Thinks 2150

FT: “For years, Netflix was the envy of Hollywood, a streaming juggernaut whose seemingly unstoppable growth forced the world’s biggest entertainment companies to rethink their businesses. But nearly two decades after launching its streaming service, the company is confronting a familiar Hollywood problem: a lack of hits.” NYTimes: “At the moment, despite being the most-watched paid streaming service in the world, with 325 million subscribers, the company is not growing fast enough for either its top executives or its shareholders.”

Chris Marquis: “I’ve come to realize that business schools don’t just reflect our economic system — they actively manufacture its priorities and normalize its harms, training the future leaders who will design and defend the structures driving our overlapping global crises. Tinkering at the margins with a class on corporate responsibility or ethics is not going to help. We don’t need to “fix” business schools. We need to abolish them.”

: “If we’re living in the AI revolution, why does nothing seem revolutionary? If this is the most consequential technology of our lifetimes, and maybe in human history, why aren’t average people experiencing more consequences? Why has day-to-day life barely changed for most people? Why is so much of human life in 2026 indistinguishable from 2016, or indeed from 2006 or 1996? Another way to frame this question is to ask, What have LLMs enabled human beings to do that they simply couldn’t do before? I periodically ask this question on Notes, and the first wave of responses is always filled with “Are you joking???” and “Uh, have you been under a rock for four years???” and similar – arguments to incredulity. Then I challenge people to actually name changes that are required to meet any reasonable definition of revolutionary change: they have to be significant, material (as in, not happening purely in software), and relevant to ordinary people. And they struggle. Some examples I’ve gotten have included automating note-taking, avoiding phone calls with the cable company or other corporate entities, streamlining laborious spreadsheet work…. This is all quite convenient, I’m sure. But that’s not exactly species-altering, history-shaking epochal change – and it’s not going to pay for the trillions of dollars that LLMs have to generate in order to earn back their truly incredible costs.”

: “The average citizen is used to paying ₹10 for a bus ride and a similar amount to clear solid waste for a small residential unit in Bengaluru. They pay hundreds of rupees in road tolls, airport charges, electricity and water prices. The UPI system is no different, but when you delay imposing a fee and you are confused about who it should apply to, everyone demands it for free.”

Martech: The Loop That Improves Itself

Getting Better at Getting Better: What Recursive Self-Improvement Would Mean for Martech

AI labs are building agents that improve the way they improve. What would that look like in marketing, where the test is not a benchmark that runs in seconds but a customer who takes weeks to respond?

The argument in one line. Recursive self-improvement in martech means a system that improves not only its marketing decisions but also the agents that make them, and eventually the way those agents are improved. Every competitor will have access to the same models, so the ability to propose improvements will be plentiful. The ability to verify them will be scarce: a fast and honest evaluation clock, an objective the system cannot game, and the rights and evidence to test inside real brands. The moat is improvement velocity: the capacity to generate trustworthy evidence faster and more cheaply than competitors, and to turn it into better agents. It rests on verification, not on intelligence.

1   Getting Better at Getting Better

In May 2025, Sakana AI and researchers at the University of British Columbia described a coding agent that rewrites its own code. The Darwin Gödel Machine proposes changes to itself, tests each variant on programming benchmarks, keeps the versions that score better and builds on them. Over the course of its run, its success rate on a widely used benchmark of real software bugs rose from 20% to 50%. The same month, Google DeepMind unveiled AlphaEvolve, which pairs language models with automated evaluators to discover better algorithms. A heuristic it found for scheduling work across Google’s data centres had by then been running in production for more than a year, and recovers about 0.7% of the company’s worldwide computing capacity. The technique has already reached marketing. In May 2026, DeepMind reported that WPP had used AlphaEvolve to refine the models it uses to predict campaign performance, achieving 10% accuracy gains over its own manual optimisation.

In March 2026, researchers at Meta went a step further. Their Hyperagents make the improvement process itself editable, so that the system improves not only how it performs a task but how it searches for improvements. This is what the field calls recursive self-improvement: a system that gets better at getting better.

These systems share an advantage that is easy to overlook. Their test is fast. Code runs or fails, a benchmark scores in minutes, and the loop can turn thousands of times. The Hyperagents authors make a further, subtler point. The Darwin Gödel Machine worked partly because evaluation and self-modification were the same kind of task: becoming better at coding made the agent better at rewriting its own code. Outside coding, that alignment breaks. Becoming better at reviewing research papers does not make an agent better at editing the software that does the reviewing.

Marketing is an extreme case of that break. A marketing agent that changes how it approaches dormant customers may wait sixty days to learn whether the change worked, and the answer comes from customers, not from a benchmark. The question for martech is not whether self-improving agents are coming. They are. The question is what self-improvement looks like when the test is slow, noisy and human.

This matters now because agents have become the centre of every martech roadmap. Nearly every vendor now offers agents that build journeys, write content or recommend the next action, and the features are converging. In Martech: The Moat Is the Loop, I argued that the defensible advantage lies not in any feature but in a loop that compounds with use. Self-improving agents raise the stakes of that argument. If the agents themselves can be improved by machines, and the machines that improve them can be licensed by anyone, the question becomes which company’s improvements can be trusted fastest, and why. Note what WPP’s gain measured: the accuracy of a prediction model on campaign data, a fast and automatic score. That is a real advance. It is not yet evidence that customers bought more because of it, and the gap between those two things is the subject of this essay.

My answer, which this essay develops, is that self-improving marketing is possible, but it will be bounded, governed and paced by verification. It will not be an explosion. It will be a steady acceleration, and the companies that achieve it will be the ones that can verify improvements fastest without fooling themselves.

In software, the test runs in seconds. In marketing, the test is the customer, and the customer takes weeks.

2   Four Levels of Improvement

“AI marketing” now covers so many things that it helps to separate four levels of improvement.

The first is automation. Agents do existing work faster: they assemble campaigns, write variants, build segments and schedule sends. The decisions are no better than before; they simply take less time and fewer people.

The second is optimisation. The system makes better decisions as evidence accumulates: which customers to contact, which offer to make, when to send. This is the familiar promise of machine learning in marketing, and when it is done properly, with holdouts, it is valuable.

The third is self-improvement. The system changes its own machinery: the decision rules, prompts, tools and workflows of its agents. It proposes a new version of an agent, tests it against the current one, and keeps whichever wins. The object of improvement is no longer the campaign but the thing that makes campaigns.

The fourth is recursive improvement. The system improves the process by which it discovers, builds and evaluates changes to its agents. It learns which kinds of change tend to work, which tests are worth running and how to propose the next change better than the last.

An example makes the ladder concrete. A system that finds a better subject line is optimising. A system that finds a better way of choosing subject lines, and changes its own choosing process, is improving itself. A system that finds a better method for generating and testing new choosing methods is improving its ability to improve. The recursion lives in that last step.

Level What improves Example How it is verified
1. Automation Speed of execution An agent assembles and schedules a campaign Time and cost saved
2. Optimisation Marketing decisions Choosing which offer each customer receives Lift against a holdout
3. Self-improvement The agents themselves Rewriting the rules for choosing offers, and testing the new version Champion against challenger on live, randomised groups
4. Recursive improvement The process of improvement Learning which kinds of agent change tend to work, and proposing better ones Verified gains arriving faster across successive cycles

Figure 1. Four levels of improvement. Most of what martech sells as AI today sits at the first two.

Most of what martech sells as AI today sits at the first two levels, and a good deal of it at the first. That is not a criticism: automation and optimisation are useful, and the higher levels are impossible without them. But they do not compound in the way the higher levels could.

A system that finds a better subject line is optimising. One that finds a better way to find subject lines is improving itself.

3   Three Nested Loops

In The Moat Is the Loop, I described two loops that could give a martech company a durable advantage. The evidence loop makes each deployment more likely to work, as verified results across brands improve the next prediction. The replication loop makes each deployment cheaper, as successful field work becomes reusable product. Both are about improving the marketing and the way it is installed. Self-improvement adds structure around them.

Picture three loops, nested. The innermost is the marketing loop, at the level of the customer: decide, act, measure against a holdout, learn. It improves the intervention. Around it sits the agent loop, at the level of the system: diagnose where an agent falls short, redesign it, test the new version and promote it if it wins. It improves the machinery that makes decisions. The outermost is the improvement loop: study which agent changes worked and which failed, and improve the way future changes are proposed and tested. It improves the process of improvement.

Each outer loop feeds the one inside it. Better methods of improvement produce better agents sooner; better agents produce better marketing decisions; and every marketing decision, properly measured, produces evidence for all three loops at once.

Figure 2. Three nested loops. Each outer loop improves the one inside it, and every measured decision feeds all three.

A hypothetical example shows how this could unfold over five generations of a recovery programme for a retailer.

In the first generation, an agent identifies customers who have been dormant for between 90 and 180 days and makes them offers based on their purchase history. Holdouts show which offers create incremental contribution.

In the second, the agent learns that its discounts are being spent on customers who would have returned anyway. It redesigns its policy: useful content for one group, a reminder to restock for another, and discounts only where the extra incentive pays for itself.

In the third, reviewing several experiments, the system notices that most of the improvement came not from better messages but from choosing different customers, including deciding whom to leave alone. It changes the agent that selects whom to contact.

In the fourth, the system observes that experiments on customer selection have produced more reliable gains than experiments on content, and shifts its testing effort accordingly.

In the fifth, working across many brands, the improvement loop learns which kinds of agent change carry over from one brand to another and which need local calibration. It begins to propose better starting agents for each new brand, and to test them more efficiently.

By the fifth generation something qualitatively different has happened. The system is no longer applying lessons from earlier campaigns. It is applying lessons about how to build marketing systems. That is the recursive step. It is also, as section 7 will argue, where the first danger appears, because the fourth generation is exactly how a self-improving system can talk itself out of its best ideas.

The first loop improves the marketing. The second improves the agents. The third improves the way the agents are improved.

4   The Evaluation Clock

Every self-improving system is limited by its evaluator. AlphaEvolve’s evaluators score a candidate algorithm automatically. The Darwin Gödel Machine runs each new version of itself against a suite of programming tasks. In both cases the verdict is cheap, quick and unambiguous, which is why the loops can turn so fast.

In marketing, the evaluator is customer behaviour, and customer behaviour is slow, noisy and costly to observe. A system can propose a thousand changes to its agents in an afternoon. It cannot verify them faster than customers respond, and it cannot run more experiments than it has customers and attention to spend on them.

It helps to think of the speed of improvement as the product of two things: how quickly the system can learn whether a change worked, and what share of its proposed changes turn out to be improvements. The second depends on the quality of the proposals, which is where better models help. The first depends on the evaluation clock, and in marketing the clock dominates.

The clock does not limit how many experiments a company can run. It can run a hundred in parallel, on different cohorts, with their results arriving in overlapping windows. What the clock limits is learning that builds on learning. An improvement that depends on the result of an earlier experiment must wait for that result, and the next improvement must wait for that one. Recursive improvement is exactly such a chain: each better agent is designed using what the previous one revealed. It is the length of each link in this chain of dependent learning, not the number of experiments launched, that sets the pace.

The arithmetic follows. If each link takes sixty days to read, the chain advances about six links a year. If each link takes two weeks, it advances twenty-six. Over three years that is eighteen links against seventy-eight. Running more experiments in parallel widens each link, testing more ideas at every step, but it does not shorten the chain. And parallelism has costs of its own: more simultaneous tests mean smaller groups and noisier results, experiments can interfere with one another, and a customer contacted by several of them is not a fresh subject for each.

There is also more than one clock. A completed form can be read at once; a repeat order takes a week; a renewal may come round once a year. And a retention programme can look excellent after thirty days and turn negative after six months, because it has trained customers to wait for discounts. A self-improving system needs to know which clock belongs to which promise, and must not promote a change on the strength of a clock that has not yet finished ticking.

So the strategic question is how to speed up the clock without making it dishonest. There are four ways, and each comes with a condition.

The first is dense daily signals. A brand that sends one promotion a week to an inattentive list receives one weak signal a week, most of it silence. A daily email that gives the reader something to do, whether a quiz, a poll, a prediction or a card to collect, produces engagement signals every day from every reader who opens it. I have argued in earlier essays that such emails rebuild the attention brands have lost. Here they play a second role: they turn the inbox into a laboratory with a fast clock. Changes to selection, timing and content can be read in days rather than months. I have written about earning that repeated attention in The Next Open. The condition is twofold. What matters is useful evidence density, not event volume: automated opens by privacy features and repetitive taps can be plentiful and tell the system almost nothing. And engagement is not contribution, so these readings must be checked regularly against holdouts that measure money.

The second is validated leading indicators: early measures, such as engagement in the first week, a return visit or a completed interaction, that have been shown against holdouts to predict contribution ninety days later. The system can iterate quickly on the leading indicator while the slow truth checks, at intervals, whether the indicator still predicts. The condition is that when an indicator stops predicting, the system must slow down until a new one is validated.

The third is calibrated simulation. Synthetic customers built from language models can screen ideas before any real customer sees them, and the evidence on how far to trust them is now accumulating. Amazon researchers tested simulated A/B tests against 67 historical marketing experiments. Uncalibrated, the simulations systematically overstated the size of effects; calibrating them against data from before each test cut the magnitude error sharply. The simulations matched the direction of the real result in about 70% of tests, but the authors caution that, judged against simple uninformed baselines on such noisy experiments, this is not clear evidence that the simulations had picked up a real directional signal. A separate team of statisticians, working with news-headline experiments rather than commercial campaigns, found that uncalibrated model responses recovered only 39% of the effect measured on people, and drew a sharp distinction: tests on simulated customers are correct only by assumption, while tests on people are correct by design. Simulation can be calibrated in promising ways, but its usefulness for choosing between interventions is not yet established. The condition is that it serves as a filter for rejecting weak ideas cheaply, never as a substitute for verification.

The fourth is replaying past decisions, and it needs the most careful handling. A record of decisions and outcomes allows a new agent to be run against past customer situations before it goes live. That is excellent for catching errors: a new agent that would have breached a permission, overspent a budget, contradicted a rule or behaved erratically can be rejected without touching a customer. AlphaEvolve’s scheduling heuristic was itself first tested in a simulator built from historical workloads. But replay is much weaker at estimating profit. If the old agent offered a customer a discount, we know how that customer responded to the discount; we do not know how they would have responded to something else. Replay can estimate the value of a new policy only under particular conditions: the old system must have assigned its actions with known, logged probabilities, it must have tried the new policy’s actions often enough on similar customers, and the statistics must be handled with care. Where the old system never tried an action on a group of customers, the history cannot say what that action would have done. The condition is that replay is used to reject invalid policies and test behaviour, and to estimate outcomes only where past experiments cover the ground.

Method Speed Good for Limits
Randomised holdout Days to months Establishing what was caused, in contribution Slow, and costs attention; but the final authority
Daily engagement signals Days Quick reads on selection, timing and content Engagement is not money; volume is not evidence; checked against holdouts
Validated leading indicators Days to weeks Fast iteration between holdout reads Can stop predicting; re-validated regularly
Calibrated simulation Minutes to hours Screening out weak ideas cheaply Overstates effects uncalibrated; direction not yet clearly better than simple baselines
Replay of past decisions Hours Catching errors and rule breaches; estimating value where past tests randomised the same actions Cannot say what untried actions would have done

Figure 3. Ways to speed up the evaluation clock. Fast methods propose and filter; the randomised holdout decides.

Each accelerator is a proxy, and proxies drift. The design principle that follows is simple to state and hard to keep: fast evaluators propose and filter; the randomised holdout decides. A self-improving system may use every shortcut available to discard bad ideas quickly, but nothing is promoted into the product on the strength of a shortcut alone. And the clock cannot be shortened indefinitely. A mature system will sometimes decide not to test another small variation, because the information it would yield is worth less than the customer attention it would consume. That is not a failure of ambition. It is discipline.

Self-improvement in marketing moves at the speed of honest verification.

5   The Evaluator Is the Moat

Recursion has a property that ordinary optimisation lacks: it amplifies the objective. A system that improves itself becomes steadily better at whatever it is rewarded for, and better at finding ways to be rewarded. If the reward is opens and clicks, it will learn to exhaust customers. If the reward is attributed revenue, it will learn to give discounts to people who would have bought anyway, the trap I illustrated in Martech: The Moat Is the Loop. Consider an agent rewarded on recovered customers alone. It will discover that larger discounts recover more customers, and each generation will offer a little more. Every report will show rising recoveries. The contribution, which nobody asked it to protect, will fall. A human team might drift into these habits slowly and occasionally. A self-improving system will pursue them systematically, quickly and with great ingenuity.

Recent marketing research offers a small but telling example. A team from Penn State, Rochester and Tennessee, publishing in the Journal of Marketing Research, built a framework that uses a company’s own campaign history to screen AI-generated email content. When they asked a language model to choose its own five best emails, the company’s data predicted those choices would perform poorly. Their answer was to divide the work: the model writes, and the company’s evidence judges. That division of labour is the principle on which self-improving marketing has to be built.

The right objective is verified incremental contribution: the extra profit an intervention causes, measured against a pre-agreed baseline and a randomised holdout, after the cost of incentives and fulfilment. Two kinds of protection sit alongside it, and they should not be confused. Some harms belong in the measurement: attention decay, the shrinking of the engaged base, unsubscribes and complaints. A change that lifts contribution this quarter while eroding the base of customers who still pay attention is not an improvement, and the evaluator should count the erosion against it. Other limits are not costs to be traded at all. Customer permissions, consent rules and contractual commitments are hard constraints: an action that breaches them is never taken, however much contribution it promises, and a change that would breach them is never promoted. A reward function that merely penalised such breaches would invite a self-improving system to find the ones worth paying for.

This gives outcome pricing a role larger than commerce. In No Lift, No Bill, I described a contract in which the vendor earns only on verified lift above an agreed baseline. In a self-improving system, that contract becomes the fixed point the loop is not allowed to change. The baseline, the holdout design, the formula for contribution and the measurement windows are set by people and written into agreements. The agents can propose changes to almost everything else.

That is the independence principle. The evaluator must sit outside what it evaluates. The research systems already work this way: both the Darwin Gödel Machine and Hyperagents were run with sandboxing and human oversight, and their scoring sat beyond the reach of the code being changed. In marketing the stakes are higher, because the system is acting on real customers and real money. An agent that could edit the rules that judge it would soon find that the easiest way to improve is to change the rules.

Why does this make the evaluator the moat? Not because of the scoring formula. Incremental contribution, holdouts and guardrails are not secrets, and any competitor can write them down. What is hard to copy is everything needed to run that formula well, at scale and at low cost: the rights to run experiments inside many brands, holdouts brands have agreed to, a permissioned history of decisions and outcomes over long periods, calibration to each brand’s circumstances, operations reliable enough to act on the results, and the funding to keep testing. Every competitor will have access to the same frontier models, and WPP’s experience shows that the tools for improving models are already commercially available. So the ability to propose changes will be widely shared. What will not be shared is the capacity to generate more trustworthy evidence, at lower cost, than anyone else. A rival with the same models and without that capacity can generate ideas as quickly. It cannot learn which ones work as quickly.

In a self-improving system, the evaluator is the moat: not the formula, but the capacity to produce evidence others cannot.

6   The Agent Factory

Software companies learned long ago that shipping code safely needs a pipeline: build, test, stage, release, monitor, roll back. A self-improving marketing system needs the equivalent for agents. Call it an agent factory.

Candidates enter the factory from diagnosis. An improvement agent studies the performance of the working agents and finds their weaknesses: a strategy agent that consistently overestimates the value of discounts, a creative agent that produces too many near-identical variants, an experimentation agent that spends test traffic on questions whose answers are already settled. It proposes changes to each. The candidates are tested first in a sandbox and against replays of past decisions. Survivors run in shadow mode, making decisions alongside the current agent that are recorded but not executed. Only then do they face the current version in a live, randomised champion-and-challenger test. Winners are promoted, versioned and monitored; losers are archived, not discarded. Every agent has a lineage, and any version can be rolled back.

Not every agent in the factory needs to be a language model. Most of the machinery should remain conventional computation: eligibility rules, uplift models, assignment systems, cost accounting, statistical tests and release controls. Language models earn their place where the system must interpret ambiguous evidence, propose alternative designs, write tools or reason about unfamiliar situations. They are not a substitute for a reliable experiment or an auditable record of what happened.

Proposals to change the improvement process itself travel through the same gates. An improvement agent may suggest searching a different class of problems, combining specialist agents differently or ranking ideas for testing in a new way. Each such change must earn promotion like any other. In particular, no agent can bypass the evaluator by proposing that a new evaluator would be smarter. The recursion operates inside a fenced factory, not outside it.

Promotion should be staged in the same way. A change that wins in one brand goes next to a handful of similar brands, and only then to everyone, with the guardrail measures watched at each step and an automatic stop if any of them moves the wrong way. The rate at which changes can be promoted should itself be limited, so that the system never changes faster than people can understand what it has done.

The archive matters more than it seems. The Darwin Gödel Machine keeps a library of its earlier agents, not only the best one, because a version that loses today may be the stepping stone to something better tomorrow. A marketing factory should do the same.

The factory also needs a clear boundary between what the agents may change and what only people may change.

What agents may change What only people may change
Decision policies and selection rules Brand objectives and priorities
Prompts, content generation and tool use Consent records and customer permissions
Workflow sequences and timing Financial limits and incentive budgets
How test traffic is allocated Contract terms, baselines and the vendor’s share
Proposals for new specialist agents Measurement rules, holdout design and windows
Retiring weaker agent versions The exploration budget and the evaluator itself

Figure 4. The boundary. Agents may change how the work is done; only people may change what counts as success.

This changes the role of the people involved. In the earlier essay, Forward-Deployed Marketers built gravel roads inside each brand, and a product team paved them into highways for the next brand. In the factory, many of the paving proposals come from the machine. The Forward-Deployed Marketer’s job moves up a level: judging which proposals deserve trust, defining what better means for a particular brand, handling the exceptions the agents cannot, and keeping the brand’s judgement inside the loop. People set the objectives, the permissions, the economic limits and the standards of proof. The machine searches within them. The Software Foundry series described agents that write and maintain software; the agent factory applies the same idea to the agents themselves. And the replication test from Martech: The Moat Is the Loop applies with extra force: if the factory’s tenth improvement costs as much as its first, there is no factory.

The marketer stops designing every agent and starts deciding what counts as a better one.

7   The Exploration Trap

Return to the fourth generation of the recovery example. The system noticed that experiments on customer selection produced more reliable gains than experiments on content, and shifted its effort towards selection. Each step of that reasoning is sound. Its cumulative effect can be corrosive.

A self-improving system learns not only what works but what is easy to verify. Changes to selection read quickly and cleanly. Changes to content, new mechanics and unfamiliar ideas read slowly and noisily. Over time the system drifts towards what it can measure and away from what it cannot. Every individual decision is rational. The result is a system that converges on a narrow, safe and slowly declining set of tactics, and stops discovering anything new.

The same Journal of Marketing Research study shows why this matters. The framework flags messages that are too unlike past campaigns to predict reliably, and directs them to live testing instead. In its data, an unusual birthday greeting, which the framework could not predict, turned out to have the strongest effect on spending. A system that promoted only what it could predict would never have sent it.

The research systems point to the remedy. The Darwin Gödel Machine’s archive of diverse agents exists precisely to avoid settling too early on a good-enough answer. In marketing, the equivalent is a protected budget for exploration: a share of test traffic and experiment slots reserved for novel ideas, set by people and outside the system’s control, so that the system cannot quietly reallocate it to safer bets. Novel ideas should be given longer measurement windows, because their effects are slower to read. Failed and untested ideas should be kept as stepping stones. And people should keep feeding ideas in, from Forward-Deployed Marketers, from brand teams and from customers themselves, because the most valuable ideas are often the ones no model trained on the past would propose. The protection should be disciplined, not romantic. An untested idea has no right to reach a million customers. It has a right to a small, well-designed experiment with a bounded downside. And the system’s own scorecard must not punish boldness: if the factory is rewarded only for the share of its proposals that survive the holdout, it will learn to propose safe, incremental changes and achieve excellent statistics by refusing to ask difficult questions. Survival rates should always be read alongside the novelty of what was attempted.

A system that trusts only what it can predict will optimise away its best ideas.

8   The Financing Loop

Self-improvement costs money. Experiments consume test traffic, holdouts, computing, and the time of the people who judge the results. Who pays for the search?

Under the billing meters martech has used for twenty-five years, nobody has much reason to. A vendor paid per message sent or per user tracked earns the same whether its agents improve or not, and many improvements, such as better targeting or fewer wasted sends, reduce the volume it bills. I described those meters in Martech Beyond Messages and MAUs. They reward activity, and a self-improving system that made activity more efficient would shrink its owner’s revenue.

Outcome pricing reverses the incentive. When a vendor earns a share of verified lift, every improvement its agents discover raises its earnings. Verified lift earns Carry; Carry funds experimentation capacity; experimentation produces better agents; better agents produce more verified lift. Profit from past outcomes pays for the search for future ones, and the search becomes part of the business model rather than a cost centre. The brand gains too, because the improvements are found on its behalf and it keeps most of the lift they produce. The arrangement works only if that remains true: a vendor whose share of the gains rose faster than the brand’s would soon find brands unwilling to keep the experiments running.

There are limits. More experimentation is not automatically better. Every test consumes customer attention, and every holdout means some customers do not receive a treatment that might have helped them, which is a real cost to the brand. The financing loop therefore needs rules of its own: an experiment budget for each brand, a ceiling on the share of customers in tests at any time, priority for the experiments whose answers are worth most, and the capital discipline that any outcome business needs.

Profits from verified outcomes fund the search for the next ones.

9   Where to Start

If the evaluation clock governs the speed of self-improvement, the place to begin is where the clock is shortest.

That means workflows whose outcomes arrive in days: unfinished transactions, such as abandoned checkouts and incomplete applications; renewals due this month; onboarding and verification journeys; and the second purchase after a first. Results arrive quickly, individual decisions are small and reversible, and the stakes are bounded. Recovering deeply dormant customers is the larger prize, but its outcomes are sparse and slow. It is a place to arrive at, not to start.

A first pilot could run for ninety days on randomised, comparable groups of customers, with the same measurement rules and budgets for each. One group would be served by a fixed agent, maintained by people in the usual way. A second would be served by the same starting agent together with an improvement process allowed to propose, test and promote changes within the boundaries described in section 6, on a logged release schedule. A third would receive no intervention, as a control for both.

Figure 5. Two experiments. The first, a ninety-day pilot, tests self-improvement; the second, longer and across several cycles, tests recursion.

Before the pilot begins, the brand and the vendor need to agree a short list of things: the permissions the agents may use, the size and design of the holdouts, the boundary between what agents may and may not change, the guardrail measures and their thresholds, and an off-switch that the brand controls. None of this is exotic. It is the same discipline any outcome contract needs, applied to a system that changes itself.

The first question is whether the self-improving group delivers more incremental contribution per unit of operating cost than the fixed group, without more incidents, reversals or complaints. Analysts must resist presenting a sequence of selectively chosen winning tests as one clean result; the comparison is between the two groups as wholes, under an analysis plan fixed in advance. A yes shows that self-improvement adds value over a competent agent.

It does not show recursion, and the distinction matters because recursion is this essay’s distinctive claim. A single successful upgrade is not evidence of recursion, and nor is a faster improvement cycle, which might have come from easier problems or better luck. Recursion needs a second experiment: an improvement procedure that is held fixed against one that is allowed to improve itself, compared on their ability to find verified upgrades for problems neither has seen, across several cycles or several brands. Only if the self-modifying procedure finds better upgrades, faster and more cheaply, has the system shown that it is getting better at getting better. That experiment needs enough successive improvement opportunities to be meaningful, and will usually take longer than ninety days.

Start where the clock is short, and earn the right to work where it is long.

10   How This Could Fail

If the moat is improvement velocity, it needs to be measured, and three measures capture it. The first is the time from hypothesis to verified result, reported separately for short-clock and long-clock problems so that they are not compared unfairly. The second is the survival rate: the share of changes proposed by agents that beat the current version against a holdout, read alongside how novel those changes were. The third is the full cost per verified dollar of incremental contribution, including failed experiments, incentives, human review, computing and make-goods. Each needs a minimum threshold for what counts as a worthwhile improvement; otherwise a system can look fast simply by promoting trivial changes. These measures extend those in The Moat Is the Loop, which were evidence density, prediction advantage and replication efficiency, by one derivative. They ask not how good the system is, but how fast it is getting better.

They also make the thesis falsifiable. These are the conditions that would refute it.

Test Kill condition
Self-improvement The self-improving agent delivers no more contribution per unit of cost than the fixed agent within the pilot
Recursion An improvement procedure allowed to change itself finds no better upgrades, on unseen problems, than one held fixed
Proxies Leading indicators stop predicting contribution measured against holdouts
Exploration The share of novel tests falls, survival rates rise only because proposals grow timid, and no new high-performing ideas emerge
Safety Incidents or reversals rise as autonomy increases, or any hard constraint on permissions is breached
Durability A competitor with the same models but no evidence base or agreed holdouts matches improvement velocity within a year

Figure 6. Kill tests for self-improving marketing. Each should be agreed before the system is given more autonomy.

The last test deserves emphasis. The code for systems like the Darwin Gödel Machine and Hyperagents is public, and general-purpose tools for improving agents will become widely available. If a competitor using such tools, with no evidence base and no agreed holdouts, can match a system’s improvement velocity, then the advantage was never in the loop. The claim of this essay is that it cannot, because the scarce input is not the method of improvement but the means of verifying it.

Even if every test is passed, there is a ceiling, and it should be stated plainly. Customers adapt to tactics. Responses have a noise floor below which no system can see. Lift has limits, and some brands have too few customers to run many experiments. Improvement will slow as the obvious gains are taken. That does not defeat the thesis. A system does not need to accelerate forever to win; it needs to keep compounding, and to learn faster or more cheaply than the alternatives. This is bounded self-improvement: not a runaway intelligence, but a business that becomes measurably better at becoming better, for as long as its evaluator stays honest and its exploration stays protected.

  • • •

The old martech moat was the cost of leaving. The new one is the value of staying. With self-improving agents, that value should keep rising, and rise faster than anywhere else, because the scarce asset is not the agent that proposes the next improvement. It is the system that can prove it.

Thinks 2149

David Friedman: “The fundamental problem is that there are no political institutions that make it in the interest of governments to act in the interest of their citizens: Market failure, the exception on the private market, is the norm on the political market. The reasons have been explored in public choice theory, the project of analyzing the political marketplace with the same assumption we use for the economic marketplace, that individuals act in their own interest.” [via Arnold Kling]

G. Padmanabhan: “Since 1991, India has sought to integrate more closely with the world economy. The scale of this engagement is now substantial. In 2025–26, exports of goods and services were over USD 863 billion, including USD 421 billion of services exports. Remittances are estimated at USD 135 billion. Yet these flows have not translated into commensurate investment. India is not among the ten largest recipients of FDI, and net fixed FDI has declined in nominal terms since 2020. This is striking at a time when global firms are reorganising supply chains and looking for alternatives to China. This paper argues that the lag in India’s growth, and the gap in its deep engagement with the world, is a problem of friction in the flow of data, capital and trade. Some of the problem lies in implementation: rules that are reasonable on paper are not administered efficiently.”

Rich Lesser (BCG newsletter): “Quality is the starting point in every category we tested, and it’s the biggest single driver of whether consumers think they’re getting good value. But the difference comes from the second and third drivers, and those vary by category. In skin care and pet care, it’s personal connection: trust in the brand and how the product makes people feel. In packaged snacks and prepared food, it’s quantity, meaning whether the portion feels right for the price. In staples and mobile, it’s promotions and bundled offers. And in autos, it’s service and convenience.”

WSJ: “AI can already write code and rewrite emails and summarize meetings and prove abstract math conjectures and figure out why your washing machine keeps making that strange beeping noise. But in [Demis] Hassabis’s mind, the whole point of intelligence is to do something with it. What he wants to do with it is make the kind of discoveries that actually change our lives…“The ultimate goal was to build AI and understand the mind, then use it to understand science,” Hassabis tells me. “Step one, solve intelligence. Step two, solve everything else.””

Publishers Have Attention, Brands Have Transactions, and Each Needs What the Other Has

Publishers Have Attention, Brands Have Transactions, and Each Needs What the Other Has

Why publishers must learn to sell, brands must learn to publish, and how the two can trade

The argument in one paragraph. For a century, publishers optimised for earning attention and brands for converting demand. Each outsourced the half it was weaker at: publishers rented distribution and left the money in their readers’ next steps to others; brands rented attention from platforms, increasingly to reach customers they already had. AI now squeezes both dependencies at once. Publishers must learn to sell, and brands must learn to earn attention, which is harder than owning a list. They can also become something new to each other: trading partners, with publishers supplying attention they have earned and brands supplying demand that pays only when a reader acts. Each has what the other lacks.

1  Two Industries, Two Halves of One Business

A newspaper and a bank look like different kinds of company. One produces stories; the other moves money. But underneath, every business that serves people at scale has to do two things. It has to earn their attention, and it has to turn some of that attention into something people pay for. Publishers and brands each became very good at one of these, and each organised itself around it.

Publishers built supply chains for attention. A newsroom produces something new every day, and the reader has a reason to come back tomorrow that has nothing to do with buying anything. That habit is the publisher’s great asset. A daily paper has fresh material almost every day of the month. The advantage is not the sending; it is that there is always something new worth sending.

Brands built supply chains for transactions. They know who their customers are, what they bought, what they might buy next, and how to take a payment, ship a product and handle a complaint. What most brands lack is a reason to be in a customer’s day when there is nothing to sell. Most have something new to say only a few times a month. Beyond that, a brand usually has to invent a reason, and an invented reason is usually a discount.

These are generalisations. Subscription publishers run serious customer operations, and the largest retailers now run media businesses. But the difference in what each side optimised for is real, and it explains much of what is happening to both.

Figure 1. Two industries optimised for opposite halves of the same business.

Publishers mastered earning attention. Brands mastered converting it. Neither mastered both.

2  Each Rented What It Lacked

What a business does not do itself, it rents. Publishers rented distribution. For two decades, search and social sent them most of their new readers, and the money in what those readers did next, such as buying, booking or applying, went to marketplaces, comparison sites and advertisers. Brands rented attention. They paid platforms to put them in front of people, and increasingly paid again to reach people who were already their customers but had stopped paying attention.

The two dependencies are mirror images, and they ran through the same small number of intermediaries. The three largest platform companies are expected to take 59.7% of global advertising spending outside China this year. I have written about owned and rented attention before, in “From Rented to Owned”; the point here is narrower. Both industries handed their weaker half to the same landlords.

Both industries outsourced their weaker half to the same intermediaries.

3  AI Squeezes Both Rentals at Once

For publishers, the squeeze is already visible. Answer engines summarise the story instead of sending the reader to it. Over two years, small publishers lost about 60% of their search referrals, mid-sized ones 47% and large ones 22%. Social and video networks have overtaken publishers’ own sites and apps as a source of news, according to the Reuters Institute. I made the case in “When Search Keeps the Answer”; it does not need repeating.

What is new is that the same shift is reaching brands. When a leading AI company let people buy inside its chatbot in September 2025, it described AI as becoming “a key interface for how people discover, decide, and buy”. Discovery was the publisher’s rental. Deciding and buying are the brand’s. An assistant that recommends the product, compares the options and completes the checkout stands between the brand and its own customer, just as it now stands between the publisher and its reader.

The intermediary that kept the publisher’s reader now wants the brand’s customer too.

4  Publishers Must Learn to Sell

A publisher that loses its free supply of strangers is left with the readers who come back on their own. Those readers are worth far more than strangers, but only if the publisher can do the things brands take for granted: recognise each reader, understand what they need, offer the right product at the right moment, take payment smoothly, and keep the customer.

Most publishers are early in this. Paying for news has stalled: across the markets the Reuters Institute surveys, only 17% pay for online news, essentially unchanged from a year earlier. Publishers are diversifying: among publishers surveyed by WAN-IFRA, revenue from sources other than advertising and reader payments, such as events and licensing, reached 25.4% of revenue in 2025. But much of it comes from events rather than from serving readers’ next steps. In “The Publisher’s Second Business” I argued that the opportunity lies in moving from telling readers what happened to helping them decide and act. That move needs a capability most newsrooms never built: conversion.

The difficulty is cultural as much as technical. A newsroom is trained to inform, not to sell, and rightly protective of the trust that makes it worth reading. Learning to sell without spending that trust is the publisher’s hardest lesson.

A publisher that knows its readers but cannot sell to them is leaving half its business to others.

5  Brands Must Learn to Publish

Brands face the opposite problem. Most have large lists of customers who once bought from them. They own the means of reaching those customers: the email addresses, the app installs, the consent. What they do not own is the customers’ attention. A brand with ten million customer emails has ten million permissions to ask for attention, not ten million engaged customers.

Owning reach is not earning attention. The gap between the two is why so many brand messages go unopened, and why brands pay platforms to reach people already on their own lists. Closing that gap means learning what publishers know: how to give people a reason to open when nothing is for sale. That might be a useful weekly digest, a utility such as a tracker or a calculator, a small daily habit, or a collection worth completing.

A brand does not need a newsroom to do this. It needs one recurring thing its customers want to receive. The hard parts are a steady supply of content, the credibility to be read rather than skimmed, and the discipline not to turn every message into a sale. Useful recurring messages have a far better chance of earning attention than a steady stream of offers, which train customers to wait for the next discount.

A list is permission to ask for attention, not attention itself.

6  The Shared Path

Seen this way, publishers and brands are on the same path, travelling it from opposite ends. Both start with owned reach: a permissioned way of reaching people. Both must turn it into earned attention: a reason to open, and to come back. And both must turn that attention into valuable action. Only the action at the end differs. For a publisher, it is advertising, a subscription, a transaction or a service. For a brand, it is a repeat purchase, a cross-sell, or a customer it never has to buy back.

Figure 2. The same three steps, with different money at the end.

Publishers are strong in the middle step and weak in the last one. Brands are strong in the last step and weak in the middle. That is why the two industries are converging. Publishers are moving into transactions, and brands into publishing, because each is trying to complete the path it was only ever half equipped to walk.

The architecture of attention is the same; only the money at the end differs.

7  Where They Collide: The Bank and the Financial Publisher

Convergence sounds tidy until two companies arrive at the same place. Consider a bank and a financial publisher, both chasing the same reader’s morning.

The financial publisher starts from journalism. It explains markets, interest rates and tax changes, and a loyal reader opens its newsletter every day. To make that relationship pay, it wants to move into decisions and actions: a portfolio tool, an investor course, a route to opening an account or buying a fund. The bank starts from transactions. It holds the reader’s salary, savings and loans, and sees them every time money moves. To stop paying platforms to sell its own customers a credit card, it wants a daily reason to be opened: a market digest, a budgeting utility, a weekly note on what the rate change means for their mortgage.

Both end up offering a daily financial briefing with a next step attached. That is convergence, but not equivalence. The bank holds balances, transaction histories and regulated obligations that the publisher does not; it can complete the transaction. The publisher can offer independent scrutiny and a range of views that a bank selling its own products cannot credibly provide; it can explain the alternatives. Their incentives differ too: the publisher earns from the reader’s attention, the bank from the reader’s money. A reader may well want both, but not two of the same thing every morning.

The same collision is coming in travel, where airlines publish destination guides and travel publishers sell trips; in education, in property and in health. If every brand becomes a publisher, the inbox floods, every message is worth a little less, and the scarce resource becomes the reader’s willingness to open anything at all.

When everyone publishes, the scarce thing is not content but a reason to open.

8  From Rivals to Trading Partners

There is another way for the two to meet. Instead of each trying to master the other’s half, they can trade. Publishers supply attention they have earned. Brands supply demand: offers, products, and a reason for a reader to act. The exchange happens where an engaged reader already is: an enquiry, a test-drive booking or an account opening started inside the publisher’s message, completed there where the email supports it, or through a short, secure hand-off where it does not.

Figure 3. Attention from one side, demand from the other, and useful, measurable actions in the middle.

This raises an obvious question. Publishers have sold advertising to brands for centuries, and sponsorships, lead generation, affiliate links and pay-per-action deals are not new. What would be different?

Today’s arrangement What could be better What must be proved
Advertisers pay for exposure or clicks They pay for verified, qualified actions A lower cost per customer, lead or sale than the alternatives
The reader goes through several steps before acting A short path from interest to action Higher completion rates
Publisher and brand measure results separately An agreed test of what was truly added Value created, not just credit reassigned

Table 1. What the trade has to improve on.

Selling actions is not new. Combining trusted attention, a short path to action and an agreed measure of what was truly added, in one commercial relationship, would be. None of the three is enough alone. A qualified action is not automatically an additional one, and a short path is worthless if the reader did not want to take it.

This changes what a brand has to do. A brand does not need to rebuild itself as a publisher to benefit from publishing. It can buy outcomes first. It pays when a reader acts, not when a reader is shown something, and it leaves the hard work of earning the attention to the people who are good at it.

The most interesting version of the trade involves the brand’s own customers. Today, a brand that wants a lapsed customer back often pays a platform to find that person again, at close to the price of a new one. Consider how the alternative would work:

  1. The overlap. A bank has customers who have stopped using its credit card. Some of them read a financial publisher’s newsletter every morning, and the bank cannot reach them economically through its own messages, which they no longer open.
  2. The permission. Matching happens through a privacy-preserving process, so neither side sees the other’s customer list. But matching protects data; it does not create permission. The reader must have agreed to relevant partner offers, and the bank must be allowed to use the overlap for this purpose.
  3. The relevant action. The newsletter offers those readers, clearly labelled, something connected to what they came to read: a calculator showing what the card’s new features would have saved them this year, with one step to reactivate it.
  4. The controlled test. Half the matched readers see the offer and half do not. The difference in reactivations between the two groups is the number of customers the publisher won back that would not otherwise have returned.
  5. The division of value. The bank pays an agreed price for each additional customer recovered, less than it would pay to buy them back. The publisher earns more per engaged reader than a display ad would bring, and keeps the reader’s goodwill because the offer was useful.

Each step can fail. The overlap may be small; permission may be missing; the action may not be wanted; the test may show the customers would have returned anyway. That is the point of setting it out this way: the trade has to prove itself one recovered customer at a time.

The first brand a publisher serves this way should be itself. Most media groups own subscription products, listings, job boards, education businesses and events. Each needs readers, and the group’s newsletters can send them, measured net of what would have happened anyway. Once that works, the same space can be offered to the publisher’s existing advertisers, and only then to brands it has never worked with.

Brands already pay for audiences close to the moment of action. Retail media, in which retailers sell advertising across their own sites and apps, is forecast to reach about $200 billion this year. That does not prove the publisher model; it shows the demand exists. Publishers hold a different kind of audience, one that comes back daily by choice, and they have barely begun to sell it this way.

Better to buy attention from those who earned it than to rent it from those who intercept it, and to pay only when someone acts.

9  The Reader Must Come Out Ahead

Publishers gain revenue from the trade, and brands gain customers. The reader risks getting more commercial messages. If that is all the reader gets, the trade will not last, because the publisher will be selling the very attention it is spending.

The difference lies in relevance. A financial newsletter that inserts an unrelated insurance promotion because the advertiser pays for leads is spending its reader’s goodwill. A newsletter that offers a reader who follows retirement planning a clearly labelled retirement calculator, with an optional enquiry, is extending what the reader came to do.

The best commercial action should feel like a useful next step, not an interruption.

Four further conditions make the trade fair:

  • A brand reaches its own customers through a publisher only where the reader has agreed, and without either side exposing its customer data to the other.
  • Brands pay for actions that would not have happened anyway, measured against a holdout, not for impressions or for customers who would have returned regardless.
  • Attention kept. Commercial actions are tested for their effect on readers coming back. If readers start to ignore the newsletter, the publisher has sold its future to pay for its present.
  • Fair division. The publisher earned the attention; the brand supplied the reason to act. Both deserve a share, and neither can be squeezed for long without the trade collapsing.

10  What Would Prove This Wrong

Each of these tests can be read within two years, against a suitable comparison group:

Hypothesis What would challenge it
Publishers can learn to sell Subscription and transaction revenue stays flat despite investment in identity and customer lifecycle
Brands can learn to publish Brand digests and utilities fail to build habits, and engagement decays to the level of promotional email
The trade beats the alternatives Advertisers cannot get a lower cost per additional recovered customer, qualified lead or sale than their next-best option, or publishers earn no more per engaged reader than from display advertising
Permission-based matching works Consent rules or customer reluctance stop brands reaching their own customers through publishers
Readers come out ahead Commercial actions reduce readers’ voluntary return, raise complaints or hurt subscription renewal compared with readers who do not see them
The inbox remains a refuge AI assistants summarise both publishers’ and brands’ messages, and human opens collapse

Table 2. Six falsification tests.

11  The Next Generation Does Both

For most of the last century, it was enough to be good at one half of the business. Publishers earned attention and let others convert it. Brands converted demand and paid others for attention. The intermediaries that made this division workable are now taking more from both sides, and the answer engines are about to take more still.

The companies that thrive from here will do both halves, or trade with someone who does the other. Publishers will learn to sell without spending the trust that makes them worth reading. Brands will learn to publish without turning every message into an offer. And the most interesting businesses may be the ones that connect them: publishers supplying attention they have earned, brands supplying demand, and readers getting something useful in the exchange.

Publishers spent a century learning to earn attention; brands spent a century learning to convert it. The next generation will need to do both, or learn to trade.

Thinks 2148

FT: “[The] protein boom that is reshaping the food industry. From dairy co-operatives to meat processors, companies are extracting more valuable ingredients from materials they once sold cheaply, while food manufacturers compete to secure supplies. The appetite for protein is driving billions in factory investments and acquisitions. “Protein is increasingly [dictating] what food companies produce, where they invest capital, and how they differentiate and grow,” says Brian Earnest, lead animal protein economist at agribusiness lender CoBank. Yet the commercial enthusiasm rests on a proposition that nutrition scientists question: that eating more protein necessarily means eating better. While some groups may benefit from increasing their intake, researchers say protein deficiency is rare in countries such as the US. They warn that the pursuit of one essential nutrient to excess can crowd others out of people’s diets.”

Azeem Azhar: “What is the demand for those chips to do the computing that, in some cases — most cases — produces tokens? And who are the customers of those GPUs? And when we look at this, we see these hyperscalers — Amazon, Microsoft and Google — and then new labs called neoclouds. These people buy millions of these GPUs. They put them together in data centers that will now start to cost billions of dollars, and they rent those GPUs out. And the thing that I spend a lot of time thinking about is: Is that capital expenditure on the actual productive asset turning into revenue with just enough return on investment to pay itself? And right now, the reason we know we’re in a compute-constrained environment is that people are making extremely good returns in 18 to 24 months on an asset because they’re able to charge essentially Uber surge pricing for their GPUs. And about 50 percent of their revenues go to the two big A.I. companies, Anthropic and OpenAI, for their spending.”

NYTimes: “A.I. is also creating a new type of work that may be sucking up much of the time the technology itself is saving. Call it a “verification tax,” “botsitting” or “rework” — it’s the often arduous task of checking that A.I. output is accurate, safe and usable, and it is putting a very human bottleneck on A.I. automation. “The binding constraint on growth is no longer intelligence,” Christian Catalini, a research associate and lecturer at M.I.T. Sloan, wrote in a February working paper. “It is human verification bandwidth.” In other words, Catalini told DealBook, “the only A.I. that’s useful is the A.I. that is verified.””

Benjamin Bratton and others: “Today, the best AI agents for many applications are not single models, but frameworks that use division of labor to harness multiple models into teams. This suggests that AGI may arrive not as a single general-purpose mind, but rather through societies of agents whose collective capacities exceed those of any one model. If so, AGI would not be achieved via any one “winning” architecture; it would emerge instead through cooperative interactions among models, tools, institutions, and human participants.” [via Arnold Kling]

Martech: The Moat Is the Loop

Learn Across Brands. Act Within Them. Underwrite the Outcomes.

If AI makes martech features cheap to build and platforms easy to leave, what makes a martech company hard to replace?

The argument in one line. Martech’s old moats were software, integrations, data custody and switching costs, and AI is weakening all four. The new moat is a loop with three parts: learning across brands, acting within each brand, and underwriting the outcomes. Each part can be copied on its own. The loop cannot, because it compounds: every brand served makes the system better at delivering results for the next.

1   The Disappearing Moat

A chief marketing officer is weighing a renewal. Her team has spent six years building journeys, segments, templates and integrations on its customer-engagement platform. The platform works. But a challenger can now reproduce most of what she sees on screen, an AI-assisted migration makes moving less frightening, and her own agents can already prepare campaigns across more than one tool. The incumbent points to its features and to the years of integration work. She asks a simpler question: what do we know now that we would not know without you?

Most martech companies would struggle to answer, because their customers have stayed for a different reason. Ask anyone who has run one why customers stay, and if they are candid the answer will not be about affection. Customers stay because leaving is hard.

For twenty-five years, four things made a martech platform hard to leave. The first was switching cost: years of templates, journeys, segments and campaign logic built inside one tool, none of it portable. The second was integration: the platform was wired into the brand’s commerce system, its app, its data warehouse, its service desk and its analytics, and every wire was a reason not to move. The third was custody of customer data: profiles, event histories, consent records and preferences lived on the vendor’s servers, and moving them felt like moving house with the furniture nailed to the floor. The fourth was depth of features: a mature platform did hundreds of things, the brand used dozens of them, and any challenger had to match those dozens before it could argue about the rest.

None of these moats was about results. A brand could be disappointed with what its platform delivered and still stay, because the cost of leaving was higher than the cost of disappointment. Vendors knew this, and so did their investors. Net revenue retention, the number the industry watched most closely, measured how well a platform kept and expanded its customers, not whether those customers were better off.

AI is weakening all four moats at once, though it is not eliminating them. Features are cheaper to build: a capability that took a product team a year can now be produced in weeks, and copied in about the same time. Migration is being automated: the templates, journeys and segment definitions that once took months to rebuild can be read, translated and recreated by machines, a shift I described in The Software Foundry series. Integration is becoming routine, as standard connectors and agents that read documentation do in hours what used to take an implementation team weeks. And agents increasingly work across tools rather than inside one. When the marketer’s assistant is an agent that can operate any platform, the platform starts to look like an interchangeable part. Some friction will remain: consent histories, suppression rules and the exceptions nobody documented are still hard to move, and no bank will shift its customer engagement over a weekend because AI can write code. But the direction is clear. Customers will have more credible alternatives, and friction will become a less acceptable reason to stay.

Data custody is weakening too, from a different direction. Brands increasingly keep their customer data in their own warehouses, and the platform reads from it rather than holding it. The vendor that was once the keeper of the customer record is becoming one of several tools with permission to use it.

Old moat What AI is doing to it The better reason to stay
Switching costs Machines read, translate and rebuild templates, journeys and segments Value that grows the longer a brand stays
Integrations Connectors and agents do in hours what took weeks Deployments that get cheaper and safer each time
Data custody Brands keep data in their own warehouses; platforms read it Permissioned learning without custody
Feature depth Features are cheaper to build and copy; agents work across tools Tested judgement about what causes results

Figure 1. The four old moats, what AI is doing to each, and the better reason to stay.

The industry’s response so far has been consolidation. On 20 October 2026, Salesforce agreed to acquire Braze for about $5.66 billion in enterprise value, a little over six times the $910 million or so of revenue Braze expects for its current financial year. The deal still needs shareholder and regulatory approval. It is a sensible combination, and it tells us something. Acquisitions buy capability, talent and customers, and Salesforce is buying a working company with strong engineers and real AI decisioning, not merely a feature list. What an acquisition cannot buy is evidence that the combined company has learned to improve one brand’s incremental profit because it has worked with another. That is not an asset sitting on a balance sheet. It is a process that has to be built, run and proved after the deal closes.

The old martech moat was the cost of leaving. The question for every martech company now is what replaces it.

2   Thousands of Brands, Little Learned

Consider what a large martech platform sees. Thousands of brands across dozens of categories send billions of messages a month through it. Every one of those messages is an experiment of a kind: a choice of customer, offer, channel, timing and words, followed by a response or its absence. No brand, agency or chief marketing officer sees a fraction of this. If experience were measured by the number of marketing decisions observed, the martech platform would be the most experienced company in marketing by a wide margin.

Yet ask a martech company what it knows about what works, and the answer is thin. It can tell you average open rates by industry, the best day of the week to send, and that personalised subject lines tend to beat generic ones. It cannot tell you, with any confidence, which intervention would bring back a lapsed customer of a mid-sized apparel brand who bought twice, last opened an email four months ago, and responded once to free shipping but never to a discount. That is the knowledge a brand needs, and almost nobody has it.

There are four reasons, and none of them is a shortage of data.

The first is the billing meter. Martech has been paid per message sent or per user tracked, and a vendor paid for volume has no commercial reason to find out which of its messages caused anything. I set out how those meters shaped the industry in Martech Beyond Messages and MAUs, and will not repeat the argument here. The short version is that the industry was paid for what it carried, not for what it caused, and so it measured carriage.

The second is that few campaigns ran with holdouts. Without a group of comparable customers who did not receive the message, there is no way to separate the customers who bought because of it from those who would have bought anyway. Attributed revenue, the number most dashboards show, credits the message with every purchase that followed it. It is a record of coincidence, and very little can be learned from it.

The third is that results stayed in each brand’s slide decks. When a campaign worked, the brand’s team presented it internally and perhaps at a conference. The evidence of why it worked, and the evidence of everything that was tried and failed, stayed in a folder.

The fourth is that what agencies learned walked out of the door with their people. Agencies see many brands, and good agency people carry hard-won pattern recognition. But that knowledge lives in heads, not in systems, and it leaves when the person does.

None of this means vendors learn nothing. Benchmarks, testing tools and AI decisioning all exist, and some are very good. The claim is narrower: no martech company has yet made causal knowledge from its whole operating history the central asset it sells. Its experience is immense; its memory of what caused what is thin.

The result is a strange paradox. Martech has been the most experienced company in marketing that never learned from its experience.

3   Learn Across Brands: Experience at Scale

Brands already pay for cross-brand learning. They simply do not call it that.

When a company hires a chief marketing officer, it looks for someone who has done the job before, ideally at several companies. A CMO who has run retention at five brands knows things a first-time CMO does not: which loyalty mechanics wear out, which win-back offers train customers to wait for discounts, how long a lapsed customer stays reachable. Nobody suggests she has taken something proprietary from her former employers. We call it experience, and we pay a premium for it. The same is true when a brand hires an agency that has seen hundreds of campaigns. The agency’s value is precisely that it has learned elsewhere.

What I am proposing is the same thing, made systematic, measured and available to every brand rather than only to those that can hire the most experienced people.

Two fields show what shared learning can do when it is organised well. The first is medicine. Every patient benefits from evidence gathered across hospitals and clinical trials they never joined. A doctor choosing a treatment draws on what happened to thousands of other patients, none of whom is named in the evidence. Nobody regards this as a loss; it is the reason modern medicine works.

The second is aviation safety. In April 1976, NASA and the US Federal Aviation Administration set up the Aviation Safety Reporting System, through which pilots, controllers, dispatchers, cabin crew and maintenance technicians report incidents and near-misses. The system marked fifty years of continuous operation in April 2026, and it still runs on the principles it was designed around: reporting is voluntary, confidential, non-punitive and independent. Reports are de-identified before anything is shared, so that no person or company can be recognised from them. Hundreds of thousands of reports later, every airline flies more safely because of incidents that happened at other airlines. The lesson travels; the name does not.

Marketing has nothing like this. What would it look like?

The unit of learning is a decision trace: a record of one decision and what followed. It captures four things about the decision itself. The customer state: where this customer was, in recency, frequency, value and engagement, at the moment of decision. The intervention: what was done. The context: category, season, channel and what else was competing for the customer’s attention that week. The cost: of incentives, of fulfilment, of the message itself.

The next two need care, because the whole argument rests on them. It is tempting to say that each trace also records the incremental outcome for that customer. It cannot. We can see what one customer did after an intervention; we can never see what the same customer would have done without it. Lift is a property of groups, not of individuals. So each trace records its assignment: the experiment it belonged to, and whether this customer was randomly placed in the treatment group or the holdout. The experiment then supplies the incremental outcome: the difference between the randomised groups, measured in contribution rather than clicks, and stated with a range of uncertainty. Individual traces explain who responded and in what circumstances. The comparison between groups says what was caused.

Figure 2. A decision trace. Each record describes one decision; lift is estimated only by comparing randomised groups.

Failures count as much as successes. Case studies celebrate wins, so the industry’s shared memory is a highlights reel. The knowledge of what did not work, the free-shipping offer that brought back nobody, the win-back sequence that only hastened unsubscribes, is the hardest to find and often the most valuable. A system that records every decision with its holdout learns from failures automatically, because a failed intervention is simply a trace with an outcome of zero or less.

This changes the kind of knowledge available. A best-practice library says what successful brands tend to do. It is useful, and it is how most of marketing’s shared knowledge is packaged today. Cross-brand learning does something different: it predicts which intervention is likely to work for this customer, in this category, at this point of dormancy, and how confident that prediction is. The library describes the past in general. The prediction addresses the decision in front of you.

That prediction is a starting point, not a verdict. What worked for a fashion retailer in India will not automatically work for a bank in Britain, and even within one category brands differ in margins, purchase cycles, price sensitivity and the trust their customers place in them. A model that performs well in groceries can be confidently wrong in insurance, and pooled evidence applied to the wrong brand can make results worse rather than better. So the goal is not to assume that what worked elsewhere will work here. It is to begin with evidence gathered elsewhere, test its relevance against the brand’s own history, and update the prediction as local results arrive. Cross-brand experience supplies the starting hypothesis. Brand-specific evidence decides whether it survives. The learning should make the first local test smarter; it should never replace the test.

The benefit is largest for the brands that can least afford to learn alone. A large brand with millions of customers can run enough experiments to learn a good deal from its own history. A mid-sized brand with fifty thousand lapsed customers cannot: by the time it has tested enough offers on enough segments to know what works, many of those customers are gone. Pooled evidence lets it begin where larger brands have already arrived, and spend its own experiments on the questions that are specific to it.

Every brand gets the experience of a thousand brands, without any brand’s customers leaving home.

4   Know-How Travels; Know-Who Stays Home

Shared learning works only if brands trust it. That trust cannot rest on a vendor’s good intentions; it has to rest on rules a brand can check. Here is the compact I would offer, written as a promise rather than a defence.

What travels is know-how: patterns, playbooks, and models trained on de-identified outcomes. A pattern might be that customers who lapse after a single purchase respond better to useful content than to discounts in the first ninety days. A playbook might be a sequence of interventions, with timings and cost limits, that has worked across many brands in a category. A model might predict which customers in a dormant base are still recoverable, and at what cost.

What never travels is know-who: customer records, any brand’s own results, and anything that could identify a brand or a person. No brand’s customer list is visible to another brand. No brand’s campaign results are shown to anyone else. No playbook says, or implies, where it came from.

One rule makes this work. A lesson is shared only when many brands have taught it. A pattern seen at one brand, or two, stays where it was learned. Only when the same lesson has been confirmed across enough brands that no single contributor can be inferred does it join the shared playbook. Shared learning does transfer generalised knowledge; that is its purpose. What it must never transfer is anything a rival could trace to a particular brand: its results, its distinctive tactics, its customers. What enters the pool is what many brands have independently found to be true, which is the same principle that lets an airline learn from another airline’s near-miss without ever knowing whose it was.

Removing names is not enough to keep that promise, and the compact should not pretend otherwise. The UK’s data regulator, in its guidance on anonymisation, treats identifiability as a spectrum and asks whether a determined person with the right resources could re-identify the source. In a small category, a distinctive pattern can point to its source without carrying a name. So the compact needs technical safeguards as well as good intentions: a minimum number of contributing brands for every lesson, suppression of small or unusual slices, checks against inference from repeated queries, controls on what any output can reveal, contractual limits on use, and independent audit. These measures reduce the risk to a level a brand can check. They do not reduce it to zero, and the promise should say so.

Participation is opt-in, and it works on give to get. A brand that contributes its de-identified decision traces receives the shared playbooks and predictions. A brand that prefers not to contribute can still use the platform, but learns only from its own history. Each brand can see what it contributed and what it received. The case for joining is not altruism. It is that every participant improves faster than it could alone, in the same way that every airline is safer for belonging to a reporting system it could never have built by itself.

Throughout, the brand remains the data controller. The vendor works under the brand’s permissions and consent rules, and uses its customers’ data only for the purposes the brand has authorised. Those permissions are separate: the right to read, the right to recommend and the right to act are granted one by one. A learning permission must never quietly become a sending permission, and a model’s confidence that an intervention would work is not consent to run it. The learning that travels is about decisions, not about people.

What travels (know-how) What stays home (know-who)
Patterns confirmed across many brands Customer records: names, contacts, profiles and histories
Playbooks: sequences, timings and cost limits Any brand’s own results and campaign data
Models trained on de-identified outcomes Anything that could identify a brand or a person
Lessons that many brands have taught Small samples and rare, traceable results
Evidence of what tends not to work Consent records, permissions and the decision to act

Figure 3. The trust compact: what travels across brands, and what stays home.

Know-how travels; know-who stays home.

5   Act Within Them: What Forward-Deployed Marketers Learn from Engineers

Knowing what is likely to work is not the same as doing it. Most martech companies have treated execution as the brand’s job: the vendor provides the tools and the brand’s team uses them. In practice the team is stretched, the tools are underused, and good ideas stall somewhere between the recommendation and the launch. It is the difference between an idea that has been demonstrated and one that has been installed.

The software industry has a model for closing this gap, and it comes from engineering.

Palantir built its business by sending engineers to sit inside customers’ organisations, often for months, to make its software work on their real data and their real problems. It has two main engineering roles: Devs, who build the core platforms, and Deltas, forward-deployed engineers who deploy them to achieve customers’ outcomes. According to The Pragmatic Engineer, until around 2016 Palantir had more Deltas than traditional software engineers. Deltas work alongside deployment strategists, known internally as Echoes, who carry much of the organisational work: understanding how the customer operates, managing senior relationships and finding the problem worth solving. The engineer owns the technical build. The strategist owns the navigation.

The model has since spread well beyond Palantir. Between 2024 and 2025, OpenAI, Anthropic and Cohere expanded their forward-deployed engineering teams to help businesses put their models to work, and the Financial Times reported that monthly job listings for the role rose more than 800% in the first nine months of 2025. The reason is the one Palantir faced two decades ago: a powerful product delivers nothing until someone makes it work in a messy, specific environment.

Marketing needs the same role. I call it the Forward-Deployed Marketer, and five lessons carry over from engineering.

Embed in the real problem, not the brief. A brief describes what a brand thinks it wants. The real problem is usually somewhere else: in a data feed that misses returns, an approval process that kills timely sends, or a discount habit that has trained customers to wait. A Forward-Deployed Marketer works inside the brand’s own systems and numbers, close enough to see the problem rather than its description.

Pair judgement with building. The Echo and Delta pairing maps closely onto marketing. A brand strategist, in the Echo role, understands the brand’s customers, politics and priorities, and decides what is worth doing. An agent builder, in the Delta role, configures the agents, data flows and interventions that do it. Behind them, as the Devs stand behind the Deltas, a product team turns what works into product.

Have agreed authority to ship. A consultant who can only recommend is easy to ignore. A Forward-Deployed Marketer has agreed authority to launch controlled interventions: within a budget, with a holdout, inside guardrails the brand has approved. The contract specifies which actions run automatically, which need approval, which are prohibited, and how a mistake is reversed. The brand sets the limits; the marketer does not need fresh permission for each move inside them.

Pave the road. Bob McGrew, an early Palantir executive, describes field work as building a gravel road that solves one customer’s problem, which the product team then turns into a paved highway the next customers can drive on. For Forward-Deployed Marketers, every successful deployment should become a playbook or an agent. The first win-back programme in a category is built by hand. The fifth should be mostly assembled from parts.

Measure productivity. If the road is being paved, the evidence shows up in one number: brands and outcomes per marketer. As agents take over routine work, each Forward-Deployed Marketer should serve more brands and deliver more verified lift. Count the trust numbers alongside it, such as incidents and reversals, so that speed is never bought with carelessness. If brands and outcomes per marketer stay flat, the company is not building a product; it is staffing a service.

That last lesson is where the model most often fails. In The Palantirization of Everything, a16z’s Marc Andrusko notes that job postings for forward-deployed engineers are up hundreds of percent as companies copy the model, and argues that most are copying the look without the platform underneath. Without reusable parts, he warns, a company that sends people to its customers ends up as “Accenture for X” with a nicer front-end. Copying the job title without paving the road produces a services business. A simple replication test applies here directly: if the tenth deployment costs as much as the first, there is no factory, only a very good team. And as I argued in The Expensive Last 10%, the hard part of any deployment is not producing the capability but making it dependable, which is exactly what a paved road has to absorb.

Figure 4. How the engineering roles map to marketing roles.

Consultants advise. Forward-Deployed Marketers ship, and then make sure the next brand needs them less.

6   Underwrite the Outcomes

The third part of the loop is the one that turns learning and deployment into a business.

It is easy to confuse underwriting with outcome pricing. Anyone can offer a performance fee: we will charge you a share of the results. Most such offers are a pricing tactic, often a discount in disguise, made in the hope that results will be large enough to cover the vendor’s costs. Underwriting is different. An insurer that underwrites a risk has to predict it, price it, hold capital against it, and make money across a portfolio even though some individual policies lose. A martech company that underwrites outcomes has to do the same: predict the lift a programme is likely to deliver, fund the work before any result arrives, design holdouts the brand accepts as credible, manage the downside when a programme underperforms, and make a profit across its portfolio of brands.

Few companies can do all of that. Prediction needs the cross-brand learning of section 3. Execution needs the Forward-Deployed Marketers of section 5. Capital and discipline need a company willing to carry risk on its own balance sheet.

The unit matters. Underwriting attributed revenue is meaningless, because attributed revenue includes purchases that would have happened anyway. The unit is incremental contribution: the extra revenue caused by the programme, measured against a concurrent randomised holdout, less the cost of the incentives used to produce it and of fulfilling the extra orders. A simple illustration shows why the distinction decides everything.

An illustration, not a result. A brand has 100,000 dormant customers. It assigns 90,000 at random to a recovery programme and holds out 10,000. Over the measurement window, 6% of the treated group buy, against 4% of the holdout. Of the 5,400 treated buyers, 3,600 would have bought anyway; 1,800 are incremental. Each order is worth $120 and earns $50 after product and fulfilment costs. A dashboard credits the programme with all 5,400 orders, or $648,000 of attributed revenue. The incremental revenue is $216,000, and the incremental margin is $90,000. Now add the incentive. A $10 discount is paid on every treated order, including the 3,600 that would have happened anyway: $54,000. Incremental contribution is $36,000. Double the discount to $20 and the same programme, with the same lift, costs $108,000 in discounts and loses $18,000, while the dashboard still shows $648,000.

That last case is the trap. A programme that brings customers back with a deep discount can show handsome incremental revenue and negative incremental contribution, because the discount also pays the customers who would have come back on their own. Underwriting on contribution keeps everyone honest about that.

I have set out the commercial structure elsewhere: in No Lift, No Bill, the baseline the brand would have earned anyway, the verified lift above it, and the vendor’s share of that lift alone; and in Martech Beyond Messages and MAUs, the 6× floor under everything the brand spends. I will not restate either here, because the point for this essay is a different one.

Alpha is also how the system learns. A contract paid only on verified lift forces honest measurement, because neither side will accept a number it cannot check. Every programme runs with a holdout, every result is incremental, and every cost is counted. That is exactly the evidence the decision traces of section 3 need. A vendor paid on volume generates activity data. A vendor paid on verified lift generates causal evidence, and causal evidence is the best training material the loop can get. The contract is not only how the vendor is paid. It is how the system finds out what is true.

The difference between promising results and underwriting them is evidence, capital and discipline.

7   The Loop

Each part so far can be copied. A competitor can hire marketers and give them a new title. It can build agents. It can offer a performance fee. What it cannot copy quickly is what happens when the three parts are joined.

The evidence loop: Learn, Act, Underwrite, Learn. Cross-brand learning produces better predictions of which programmes will work for which brands and which customers. Better predictions let the company choose better opportunities to underwrite: brands where lift is likely, programmes where the cost is justified. Forward-Deployed Marketers deliver those programmes inside each brand. The contract verifies the results against holdouts. And the verified results, de-identified and pooled across brands, improve the predictions. The next round starts from a better place than the last.

Each turn makes the next one more profitable. Predictions sharpen, so fewer programmes fail. Fewer failures mean smaller losses across the underwriting book. Smaller losses let the company price more keenly, or take on brands that others would turn away. More brands produce more decision traces, and more traces sharpen the predictions again.

The replication loop: every paved road lowers the cost of the next deployment. The first programme in a category takes a strategist, an agent builder and weeks of work. By the tenth, most of the programme is an assembled playbook, and the marketers’ time goes to the parts that are specific to the brand. Lower deployment cost changes which brands are worth serving. A programme that once needed a large brand to justify the effort becomes viable for a mid-sized one, and then for a small one. That opens a long tail of brands that competitors without paved roads cannot afford to serve, and the long tail feeds more experiments back into the evidence loop.

Picture both loops in a single category. The first time a team recovers dormant customers for a home-furnishings brand, almost everything is new: which customers are still reachable, which offers bring them back without training them to wait for sales, how long a holdout must run before the result can be trusted. The strategist and the builder work it out by hand, and the contract is priced cautiously, because the prediction behind it is weak. By the fifth brand in the category, the team knows roughly which customers will respond and which will not, the sequence exists as a playbook, and the agents run most of it. The prediction is good enough to price the contract more keenly, and the deployment is cheap enough to offer the programme to brands a tenth the size of the first. Each of those smaller brands adds its own traces, and the sixth programme is better again. Nothing in this story requires a breakthrough. It requires that every programme is measured, recorded and reused.

The two loops reinforce each other. The evidence loop makes each deployment more likely to succeed. The replication loop makes each deployment cheaper. Together they mean the system gets both better and cheaper with every brand it serves.

Figure 5. Two loops. The evidence loop makes each deployment more likely to work; the replication loop makes each one cheaper, which brings in the brands that feed the evidence loop.

One distinction matters more than any other here: data scale is not learning scale. A platform serving a thousand brands does not necessarily know more than one serving a hundred. It may simply have ten times as many poorly measured campaigns, and a specialist serving fifty similar brands may learn more than a generalist serving a thousand unlike ones. The advantage depends on how much of the activity produces comparable, causal evidence. Three measures capture it better than any count of brands or messages. Evidence density: the share of deployments that run as proper experiments with credible holdouts. Prediction advantage: how much better pooled learning predicts lift than each brand’s own history alone. Replication efficiency: how much time and cost a proven play saves at each successive brand. Put together, they answer the question investors rarely ask: how much more reliably and cheaply does the system produce a dollar of incremental contribution at the hundredth brand than it did at the tenth?

The compounding has to be earned, not assumed. Learning may fail to transfer. Cheaper deployments may erode quality. A tactic used across many brands may fail across all of them at once in a downturn, giving the portfolio correlated losses. And contracts paid on lift can tempt a provider to take only the easy recoveries and leave the hard customers behind. A serious system tests for all four, and does not claim compounding merely because its customer count is rising.

It is worth being precise about what is the moat and what is only the machinery. Agents can be built by anyone, on models available to all. Data integration is becoming routine. Good marketers can be hired, at a price. These are the machinery, and none of them, alone or together, is defensible. What cannot be copied quickly is the accumulated combination: a history of decisions and verified outcomes across many brands; the rights to act inside those brands, granted through contracts and trust built over years; repeatable processes that turn field work into product; and the capital and track record to underwrite outcomes at prices others cannot match. A competitor starting today with the same software and equally good people would start with none of these, and the only way to acquire them is to run the loop for years.

This turns the old moat inside out. In the old model, a brand stayed because leaving was costly. In the new one, a brand stays because staying keeps getting more valuable: its programmes draw on more experience every quarter, its deployments get cheaper, and its contract pays only for what works. Leaving would mean giving up all three and starting again with a vendor that knows less. The brand’s own data and workflows stay portable, and it remains free to go. That is the point: it stays because of what it gains, not because of what it would endure. The answer to the CMO’s renewal question is a record, against her own holdouts, of what the system has learned and delivered for her.

The new martech moat is the value of staying.

8   Three Objections

“Google, Meta and Salesforce already have this intelligence.” Much of this is true. They have vast data across advertisers, far more than any martech company, and excellent AI engineering. The engagement platforms are moving fast too: Braze now offers AI agents that build journeys from conversational prompts, and Salesforce will add that to its own stack. Pretending a challenger will have more data or better general-purpose AI is not a serious argument. Incumbents may well have components of the loop. What few have shown is the complete combination: permissioned cross-brand causal learning, repeatable deployment, and portfolio-level underwriting of outcomes. Much of the advertising platforms’ learning is about acquisition: which audiences to buy, which ads to show, at what bid. The learning in this essay is narrower and different in kind. It is permissioned, built on the brand’s own relationships with existing customers rather than on rented reach. It is about retention and recovery rather than acquisition. And it rests on causal evidence from holdouts rather than on attributed conversions. More data does not substitute for the right data. There is also a harder test, and it is the one that matters: cross-brand learning must beat what each brand could learn from its own history alone. A large brand with years of well-run experiments may need little help. It also has to beat a strong incumbent’s recommendations. Until that head-to-head test is won, the moat is a hypothesis, and the first kill test in section 10 is designed to find out.

“This is consulting in disguise.” It is, unless two numbers move. The cost of deploying a proven programme must fall with each brand, and brands and outcomes per marketer must rise. A consulting firm’s revenue grows in step with its headcount. A loop’s revenue should grow faster than its headcount, because each marketer is supported by a growing set of agents and paved roads. If those numbers do not move, the objection is right.

“Outcome pricing is a discount.” For a vendor that cannot predict results, it is. It will win the deals where lift is easy and lose money on the rest. Outcome pricing is a good business only for a provider whose prediction and execution are better than the market’s. Then it can underwrite programmes at margins others would refuse, because it knows which bets are good. The pricing does not create the advantage; it is how the advantage gets paid.

9   What the Company Must Change

A martech company that wants to build this loop has to change more than its pitch.

Finance. Performance fees are paid in arrears. The work is done, the holdout is read, and only then does the vendor earn. Meanwhile it has paid its marketers, run its agents and, in some programmes, funded incentives. Outcome work therefore needs working capital, and the need grows with success. Each engine that runs on outcomes needs its own capital plan, with clear limits on how much risk the company carries at any time.

Organisation. Forward-Deployed Marketers need a career track of their own, with progression based on outcomes delivered and roads paved rather than accounts managed. They work in pairs, strategist and builder, and the pairs need a product team behind them whose explicit job is to pave roads: to take what works in the field and make it reusable. Without that team, field work stays field work.

Incentives. Marketers should be paid partly on the verified outcomes they deliver and partly on what they turn into product. The second matters as much as the first. A marketer who delivers excellent results through bespoke work that nobody else can reuse has helped one brand. One who turns that work into a playbook has helped every brand that follows. Trust belongs on the same scorecard: an intervention that makes money but breaks a customer permission counts as a failure, not a trade-off.

The investor story. Markets give variable revenue a lower multiple than subscription revenue, because it is harder to forecast. A company moving to outcomes should not hide this. It should show compounding instead: rising evidence density, a widening prediction advantage over brand-only learning, the falling cost of deployment, and the growing contribution margin across the underwriting book. Those numbers, if they move, tell a better story than recurring revenue, because they describe a business that improves with use.

Who can do this. The trade is easier for a profitable, founder-controlled company than for one that reports recurring revenue every quarter. Moving from subscriptions to outcomes means accepting a period in which reported revenue looks worse while the loop is being built, and a company judged each quarter on subscription growth will find that hard to explain. A company that funds itself and answers to no quarterly scorecard can make the trade and wait for the loop to compound. That is an advantage of structure, not a guarantee: large companies can build outcome programmes too, and buy capability quickly when they choose to. The opening lasts only as long as the challenger improves its evidence, economics and trust faster than they do. The Braze deal shows where the subscription path often leads: into a larger platform’s portfolio, valued as a multiple of the revenue it brings. That is a fine outcome for shareholders. It is not the only one available.

10   How This Could Fail

Every argument should name the evidence that would refute it, and the tests should be designed before the company scales, not after. The cleanest test of cross-brand learning is to take a group of consenting brands and randomise the decision system itself: pooled learning for some programmes, an equally capable system trained only on each brand’s own history for others, with execution and measurement held constant. Compare the lift, and compare how well each system’s forecasts held up out of sample. Track the cost and time of deploying the same proven play across successive brands. Stress-test the underwriting book against bad years. And have the trust safeguards audited by people who are not selling the programme. These are the conditions that would refute the thesis.

Test Kill condition
Cross-brand learning No out-of-sample gain over an equally capable system trained on the brand’s own history alone
Trust Brands decline to opt in, or protected brand information is shown to have leaked
Forward-Deployed Marketers Time and cost to deploy a proven play don’t fall across successive brands
Underwriting Forecast accuracy and contribution margin don’t improve as the portfolio grows, after losses and make-goods
The loop A competitor with the same software but no portfolio matches results within a year

Figure 6. Kill tests. If any of these conditions is met, the loop is not a moat.

The commitments should be staged in the same order. Prove that one intervention creates incremental contribution. Prove that it can be repeated cheaply at another brand. Prove that pooled experience predicts better than local history. Prove that a diversified portfolio can be underwritten at a profit after failures and make-goods. Each stage earns the right to fund the next. All five tests can be read within two or three years, and the first four much sooner. If the loop fails them, the company that built it will have built a well-run services business with an unusual pricing model. That is not a disaster, but it is not what this essay claims. If the loop passes them, the moat will be visible in the numbers long before anyone describes it in a pitch: deployment costs falling brand by brand, predictions getting more accurate, and contribution margins widening as the portfolio grows.

  • • •

The old martech moat was the cost of leaving. The new martech moat will be the value of staying, because every customer makes the system better at delivering results for the next.

Thinks 2147

FT: “This is an era of Davids, not Goliaths…Why smaller powers are thwarting bigger ones.”

Andy Kessler: “AI architecture is evolving. Instead of a trillion parameters, efficient small language models might use 70 billion and run on an Apple Mac. These local models have no token costs beyond the cost of hardware and electricity. But they must be less capable than the large language models, right? Not really. Stanford researchers show that a locally run small language model “can successfully answer 88.7% of single-turn chat and reasoning queries.” No question we’ll always have high-end models running in data centers. IBM still sells mainframes. But growth will happen elsewhere. Where? Another Silicon Valley saying I’ve lived by: Intelligence moves out to the edge of the network. The real breakthrough will happen when robust AI models run on smartphones. It’s already starting. In September, California-based PrismML released a compressed version of the popular open-source Qwen3.8 27B AI model. The company got it down to under 6 gigabytes. That means it fits not only on PCs but also on high-end smartphones. The memory size is reduced by roughly 9 to 10 times. There’s that factor of 10.”

NYTimes: “Strength training isn’t just for muscle and bone health — it’s also surprisingly good cardio.”

Mint: “Over the past two decades, Indian households have undergone an extraordinary material transformation. Our household data shows electricity access rising from 75% in 2004-05 to an estimated 99% in 2025-26. Toilet access has increased from 52.9% to 95%, Liquefied petroleum gas (LPG) or piped natural gas (PNG) from 32.3% to 85% and tap water from just 19.4% to 80%. Refrigerator ownership has risen from 18.5% to an estimated 65% while air- conditioner ownership, almost negligible at 1.2% two decades ago, has reached 21.1%. Mobile phones, present in only 16.1% of households in 2004-05, are now almost ubiquitous. These are not cosmetic changes in consumption. They represent a profound improvement in everyday life. Development has entered the Indian household, its kitchen and bathroom, communications, mobility and comfort. Yet, beneath this transformation lies a less visible change: the face of disadvantage has changed faster than the structure of disadvantage.”

The Publisher’s Second Business

From telling readers what happened to helping them decide what to do next

The argument in one paragraph. For a century, publishers built their businesses around one job: telling people what happened. Readers have two more jobs, deciding what it means for them and doing something about it, and most of the value in those went elsewhere. Publishers once held part of it, through classifieds, and lost it to better software. AI now makes it cheaper to turn editorial knowledge into tools and services, but the same assistants that keep the answer also want the decision and the action. A publisher can build a second business only where it holds something the assistant lacks, can run it reliably, and can find someone who will pay. It need not own the screen: it can build the tool, supply the assistant, or partner with whoever completes the task. What matters is not who controls the interface, but who owns something the decision cannot do without.

1  The Business Newspapers Already Lost

For most of the twentieth century, a newspaper ran two businesses under one masthead. The first told readers what had happened. The second helped them act: find a job, rent a flat, sell a car, hire a plumber. That second business was classified advertising, and it was enormous. In 2000, American newspapers earned $19.6 billion from classifieds, about 40% of all their advertising revenue. Recruitment ads alone brought in $8.7 billion.

Within a decade most of it had gone. By 2009, classified revenue had fallen to roughly $6 billion; by 2012 it was $4.6 billion, down about 77%. Help-wanted advertising, the most valuable category, lost more than 90%. Many newspaper groups saw it coming and invested in online sites such as Cars.com, CareerBuilder and Apartments.com, which, as a Congressional Research Service report notes, drew classified advertising away from the papers’ own pages. Reach and a famous masthead were not enough. The winners had better software: search, filters, alerts and a market that grew more useful with every listing.

Figure 1. Publishing’s first second business, and how fast it went.

Publishers have had a second business before. They lost it because matching people with things went to whoever built the best software.

Some media groups did hold on to large second businesses, and those businesses grew bigger than the journalism. When Axel Springer split itself in two, its classifieds businesses, including the StepStone jobs platform and the Aviv property group, accounted for about €10 billion of a €13.5 billion valuation. Schibsted sold its newspapers to its main owner for about $570 million; its marketplaces stayed listed as a separate company. The Washington Post Company earned $2.9 billion, or 62% of its revenue, from its Kaplan education business in 2010, and when it sold the newspaper in 2013, it kept Kaplan.

These cases prove less than they seem to. Kaplan was bought, not built from the newsroom; it entered higher education by acquiring a chain of colleges. Marketplaces won on network effects and software, not on reporting. They show that media companies can own valuable businesses beyond journalism. They do not show that journalism itself can be turned into one.

Publisher Second business Built from the journalism? What happened
US newspapers Classified advertising Partly: local reach, not newsroom knowledge Lost most of it to specialist marketplaces
Axel Springer Job and property marketplaces No: separate marketplace businesses About three-quarters of group value; separated from the newsroom
Schibsted Online marketplaces No: marketplace software Newspapers sold to the founding trust
Washington Post Co. Education No: acquired Grew to 62% of revenue; newspaper sold, Kaplan kept
Dow Jones Professional information Closer: news archives, data and expertise About 40% of revenue and growing (section 4)

Table 1. Most successful second businesses were acquired or built as separate software businesses.

The distinction matters for what follows. A publisher can acquire a second business, as media groups have done for decades. The more interesting opportunity is to build one from the knowledge, relationships and expertise that publishing itself creates. History offers few examples of the second kind. Diversification of the first kind is already well under way: according to WAN-IFRA, revenue from “other” activities such as events and business services reached 25.4% of surveyed publishers’ revenue in 2025, up from 13.2% in 2021. This essay is about why knowledge-led second businesses may now become possible, and when they make sense.

2  Know, Decide, Do

Every reader has three jobs. Know: what happened? Decide: what does it mean for me? Do: what is my next step? Publishing has concentrated on the first, and the first has a public value no business model fully captures. But commercially, the middle job went to whoever stood nearest the decision: the search box, the comparison site, the broker, the adviser, the software vendor. Much of the third went to marketplaces.

Consider a local paper reporting that the city’s schools have changed their admission rules: new catchment boundaries, a lottery for oversubscribed schools, earlier deadlines. That is Know, and it is good journalism. But a parent reading it has a bigger problem. Which schools is my child now eligible for? How do they compare on the things I care about? When must I apply, and what documents do I need?

The paper already holds much of the knowledge to answer those questions. It has reported on every school for years. It knows the rules and the officials who wrote them. Yet it publishes the story and stops, and the parent takes the next step somewhere else, or alone.

Figure 2. Three jobs. Publishers kept the first; others captured the value in the next two.

The article informs the decision. The larger commercial opportunity often lies in helping the reader make it.

3  What Publishers Already Hold

A publisher setting out to help readers decide and act does not start from nothing. It holds six assets that most new entrants would have to build or buy:

  • Damaged, but still real. Trust in news stands at 37% across the markets the Reuters Institute surveys, a record low, yet a known masthead still carries more authority than an anonymous site.
  • Specialist expertise. Reporters who have covered schools, markets or one industry for years know where the traps are.
  • Verified archives. Decades of checked facts, figures and context.
  • Readers who come back on their own and have said what they care about.
  • Local knowledge. People on the ground who know which rule changed and which provider cut corners.
  • Convening power. The ability to bring people facing the same decision together.

Assets are not advantages until they are tested. An archive that any assistant can summarise is not proprietary intelligence. Trust in a headline is not permission to give regulated advice. And audience size is missing from the list on purpose: newspapers had huge audiences when they lost classifieds. An audience is a reason to try a second business, not a reason it will work.

The right question is not “what can we sell to our readers?” but “what can we help them do better than anyone else?”

4  When Knowledge Becomes Software

Until recently, turning editorial knowledge into a product meant hiring software teams and data engineers and waiting years. Only the largest publishers could afford it, and those that did built some of the best businesses in the industry. Dow Jones’s professional information arm, which sells risk, compliance and energy intelligence to companies, now accounts for around 40% of Dow Jones revenue, and management says a significantly greater share of its profit. Its Risk & Compliance unit grew 19% to $100 million in a single quarter this year. It is the clearest case of a knowledge-led second business at scale, built on news archives, data and editorial judgement.

AI changes who can try. A regional publisher can now build a working prototype of an answer service over its archive, a calculator that applies this year’s rules or a checklist that knows the local deadlines, in weeks rather than years. That is a prototype, not a product. Turning it into a service people rely on still means ingesting and verifying data, updating it when rules change, supporting users, securing their information and clearing the rights. Those costs can dwarf the cost of the first demonstration.

So AI does not remove the hard part; it moves it. Building gets cheaper; maintaining matters more. A wrong answer in a decision tool costs the reader far more than a wrong fact in an article. Accuracy, provenance and human oversight become the main operating cost of a second business, and its main source of value.

AI’s biggest opportunity for publishers may not be cheaper journalism. It may be more businesses that journalism can support.

5  The Assistant Wants the Same Job

There is an obvious objection, and it could sink the argument. The companies building AI assistants say plainly where they are heading. When OpenAI let users buy products inside ChatGPT in September 2025, it described AI as becoming “a key interface for how people discover, decide, and buy”. That is Know, Decide and Do in a single sentence. Assistants that summarise the news will also compare schools, model a loan and book an appointment. If they keep the answer, why would they not keep the decision and the action too?

Often they will. Generic decisions, such as the cheapest flight or the best-rated vacuum cleaner, will go to whoever has the widest reach and the best interface. That is the classifieds lesson again. Publishers without a differentiated asset have little reason to compete head-on with general-purpose assistants. Three assets are harder for an assistant to supply:

  • Verified, proprietary or local data. This year’s admission rules for one city’s schools, the cut-offs for a regional exam, the tradespeople who did the job properly. Information gathered and checked on the ground is often not on the open web, or not reliably.
  • A named institution that answers when it is wrong. When the stakes are money, health or a child’s education, people want someone answerable, not just an answer.
  • People who trust each other through the publisher: parents in the same district, investors in the same sector. Much of the value comes from the others in the room.

Holding one of these is necessary but not sufficient. A publisher may have trusted reporting on schools but no accurate, current admissions data. It may have a community but not the expertise to run a service reliably. It may be accountable but unable to serve thousands of users economically. A second business has to pass three tests:

  1. Do we hold data, accountability or community that an assistant cannot cheaply reproduce?
  2. Can we turn that advantage into a reliable, maintained product?
  3. Will readers or institutions pay enough, repeatedly, to make it profitable?

Figure 3. Three tests for a second business, and what to do when one fails.

Passing the first test creates an opportunity. Passing all three makes it a plausible business.

Job Assistant’s advantage Publisher’s possible edge Where the publisher stands
Know: general news summary Reach, speed, convenience Original reporting, provenance Best as a licensed source
Decide: generic product choice Breadth, comparison at scale Little Stay out
Decide: local or high-stakes choice Convenience Verified data, named experts Can compete, if all three tests pass
Do: generic purchase Checkout inside the conversation Little Stay out
Do: local service or application Convenience Verified providers, community Can compete, often through a partner

Table 2. The assistant’s default advantage, and where a publisher can still compete.

6  The Interface Isn’t the Business

Return to the parent. Suppose a regional publisher keeps the most accurate, current information about local schools: rules, catchments, deadlines, results. The parent does not visit the publisher’s site. She asks her assistant which school her child is eligible for. Three outcomes are possible.

  • The publisher runs its own eligibility checker and deadline tracker, and earns from readers who use it.
  • The assistant answers using the publisher’s verified data, under a paid licence, with the source named.
  • The publisher guides the parent, and an accountable specialist, such as a school-admissions service, completes the application, with any commercial relationship disclosed.

Only the first requires the parent to come to the publisher’s own screen. As assistants become the place where decisions start, the second and third may matter more. Supplying is different from licensing an archive for training: what is sold is live, maintained, verified knowledge, priced on use, and its value depends on being kept right. That is a business an assistant cannot easily build for itself, because it rests on people on the ground and an institution that answers for errors.

Figure 4. Three ways to turn trusted knowledge into a second business.

What matters is not who controls the interface, but who owns something the decision cannot do without.

7  Four Second Businesses

Four domains show where the tests are passed most often. Two deserve depth, because they show both the promise and the hardest question: who pays.

Professional intelligence is the proof that a knowledge-led second business can be large. Industry news tells a manager what changed. The second business tells the company what a new regulation means for it, whether a counterparty is safe and what to watch next, delivered inside the work through screening, monitoring and alerts. The advantage is proprietary, structured data plus accountability, and the customer is a company that pays for reliability because errors are expensive. Dow Jones shows the scale this can reach. It also shows the cost: the data must be maintained continuously, enterprise buyers are demanding, and liability comes with the territory. It proves the category, not that a small newsroom can copy it.

Education shows where the economics get hard. Education pages already report exam dates, results and policy changes. The second business helps a student or parent decide which exam, course or school fits, then do the work: eligibility checks, practice, assessments, applications and learning plans. The advantage is verified local data and a high-stakes need. But who pays for a school-admissions tool? A parent may use it intensely for a few weeks and never again. Schools paying for placement create an immediate conflict. The options each change the business: a membership that bundles the tool with other services, a contract with an education authority, sponsorship kept visibly separate from rankings, or supplying the verified data to the services and assistants parents already use. A reader’s need does not automatically mean a willingness to pay.

Finance follows the professional model for consumers: calculators, plain-language explainers and investor education, with next steps routed to licensed partners. Its boundary is regulation; in most markets, explaining is journalism, but personalised investment advice needs a licence. Local turns civic reporting into verified directories, bookings, events and memberships. Its advantage is local data and community; its risk is becoming one more directory, so “verified” has to mean something.

Domain Decide Do Likely model Who pays
Professional What this means for my company Screening, monitoring, alerts Build Enterprises, through licences
Education Which exam, course, school Eligibility, practice, applications Build, supply or partner Hard: members, institutions, licensees
Finance What this means for my money Next steps via licensed partners Build and partner Subscribers; disclosed partner fees
Local Which contractor, clinic, school Verified directory, bookings, events Build and partner Members, verified providers, events

Table 3. Four second businesses, each built on knowledge the publisher already holds.

8  The Trust Tax

Every second business tempts a publisher to spend the trust that makes it possible. A lender will pay for leads. A school will pay to be listed first. A retailer will pay a commission on every sale. Readers may trust a publisher’s journalism without wanting it to become their broker or their adviser.

Kaplan is the cautionary case, even though it was acquired rather than built from the newsroom. A Senate committee found that Kaplan’s higher-education business drew 85.9% of its 2010 revenue from federal student aid, and 87.9% including military education benefits. The same investigation documented aggressive recruiting and high withdrawal rates, while also crediting the company with significant later reforms. Meanwhile the parent company’s chief executive led the industry’s effort to soften new federal rules, as his newspaper covered the same policy fight. A second business had put the first in an awkward position.

The boundaries are not complicated, but they have to be built in from the start:

  • Editorial independence. The newsroom does not write for the second business’s clients.
  • Every commercial relationship and ranking method is visible at the point of decision.
  • Reader data is used only for what the reader agreed to.
  • Information, advice and advertising are labelled and kept apart.
  • Licences and oversight. Where advice is regulated, the publisher holds the licence or works through someone who does, with expert review and a visible process for complaints and corrections.

The hardest case is a service that succeeds commercially while trust in the newsroom falls. A publisher should decide in advance which trust measures would trigger a review, and be prepared to change or close the service if they move.

Trust is the asset only if the second business does not spend it.

9  Who Can Afford a Second Business?

Smaller publishers are under the most pressure. Over two years, small publishers lost about 60% of their search referrals, mid-sized ones 47% and large ones 22%. But pressure is not capability. Many smaller publishers lack product teams, capital, proprietary datasets or enough customers to support a software service. Some will do better through membership, philanthropy or specialist reporting than through any second business.

For those that do have an advantage, three structures are feasible without a large balance sheet: a focused local utility that does one job well; a shared service built with peers, so that several publishers spread the cost of the data and the technology; or a supply or partner arrangement that turns their verified knowledge into revenue without building a product at all.

Smaller publishers may need a second business most, but need is not evidence that they can build one.

10  The Economics of the Second Business

The test is contribution, not traffic. Contribution means revenue minus the costs that a second business adds: maintaining data, expert review, technology, support, liability, distribution and acquiring customers. A decision tool used by a few thousand people who renew every year can be worth more than an article read by a million, or it can lose money on every user. Only the arithmetic, against a fair baseline, will say which.

Revenue comes in several forms: subscriptions to tools, service fees, licences to businesses and assistants, disclosed partner fees, and events and memberships. Most second businesses will combine two or three.

The organisation sits beside the newsroom, not inside it. Journalists supply the knowledge; product builders turn it into tools; editorial standards govern both. Start narrow: one decision readers face repeatedly, with high stakes, where all three tests can plausibly be passed.

Governance needs deciding early. Table 1 showed what happens when a second business succeeds: when it is worth more than the journalism, investors want the two apart. A publisher should decide at the outset whether its second business exists to fund journalism, and design ownership accordingly.

11  What Would Prove This Wrong

Each of these tests can be run within 12 to 24 months, against a suitable comparison group:

Hypothesis Evidence that would challenge it
Publishers hold a real advantage Users of a publisher’s decision tool report no added value over a general assistant, and usage falls as assistants add the same features
Someone will pay Free use is high, but trials do not convert into recurring payments, licences or other revenue that covers costs
AI lowers the cost enough Verifying, updating and supporting the product consume most of its gross profit
Trust survives commerce Trust in the publisher falls among users exposed to the service, relative to those who are not
The interface need not be owned Publishers cannot earn from their knowledge either directly or by supplying assistants and partners
It is not only for giants After 24 months, profitable examples remain limited to the largest publishers

Table 4. Six falsification tests.

12  The Publisher of 2030

The publisher of 2030 will still be defined by its journalism. Reporting earns the trust, the expertise and the relationships that every second business depends on. Without the first business, there is no second one.

But journalism no longer needs to be the only product. Not every publisher will become a software company; not every archive will become a product; not every reader wants a publisher to act on their behalf. Those with a clear right to win will turn their knowledge into something readers and institutions can use: to choose a school, manage their money, run a business or find their way in their own city. Some will own the service. Others will power it from behind an assistant, or through a partner. The question for 2030 is not how many articles a publisher produces, but what people can reliably do because that publisher exists.

The first business of publishing was making information worth reading. The second may be making knowledge worth using.

Thinks 2146

FT: “The AI trade is no longer new and it is no longer “a” theme. Instead it has morphed into “the” theme — a vast, suffocating blob slathered on top of the global financial system, starving everything else of oxygen, distorting market reactions and leaving every portfolio, everywhere, alarmingly reliant on the fortunes of just one bet. Stocks, real estate, infrastructure, energy, emerging markets, some vintages of private equity . . . they’re all, to a large extent, the same trade.”

Mint: “Two reforms can help India move from a slow, expensive manufacturing destination to a fast, cheap one: averaging regular hours over a longer period and reducing the overtime premium.”

WSJ: “When we eat, the energy in the food is shared between our own cells and the trillions of microbes that inhabit our intestines. How much energy we absorb depends on the quality of the food and our gut microbiome. Controlled trials show how important this can be. On highly processed diets, the body can absorb roughly 97% of consumed calories. On minimally processed diets rich in intact foods, absorption can fall to 85% or lower. For a 2,500-calorie diet, that difference is about 300 calories a day. Yet the calorie counts appear identical. Calorie totals also don’t tell us where the absorbed energy goes. The quality of what we eat influences hormones, immune function, inflammation, insulin signaling and our gut bacteria. Together, these systems determine how nutrients are partitioned in the body.”

NYTimes: “Fourteen authors in the New Directions stable have received the Nobel Prize in Literature, including this year and last. How does this small press pack such a punch?”

Sadanand Dhume: “China still holds lessons for India—and any democracy. Julian Gewirtz, a China expert at Columbia University, calls the country’s economic takeoff “as much a story of bottom-up experimentation and exploration of international ideas as anything else.” By embracing meritocracy, China has made its universities world-class. China sees its cities as engines of growth, not merely cash cows for rural hinterlands. But these are arguments for improving Indian democracy—so it may one day offer the developing world a better example than China.”

Making AI Affordable for All

Private, Useful AI for Every Employee, at Emerging-Market Prices

**

What would it take to give every employee in an emerging-market company the AI tools that a few people in a rich company take for granted?

AI has an access problem hidden inside a pricing problem. The tools that make office work faster are priced in dollars, per seat and increasingly per task, for buyers in the richest economies. The leading office-suite AI add-on lists at $21 to $30 per user per month, on top of the base licence. A company in Mumbai, Jakarta or Lagos often pays much the same dollar price for employees whose salaries are a fraction of their counterparts’ in New York or London. So it buys a few seats for senior people, and everyone else goes without, or uses whatever free tool they can find.

This essay completes a series of three. The Compute Strait asked how much inference a job needs. The Compute You Already Own asked where the inference that remains should run. This one asks who gets it, and at what price. The goal is not cheaper AI for a few. It is useful, private AI for every employee, at a price an emerging-market employer can sustain.

My answer is not a set of cheaper substitutes for famous AI products. It is a shared utility inside the enterprise: a few everyday AI jobs, joined by one layer that remembers the work, running first on compute the company already owns, under rules the company sets. The rest of the essay explains why that combination could cost a tenth of today’s prices, what competes with it, what it would take to make the economics work for the supplier as well as the buyer, and how it could fail.

Figure 1. Three essays, three questions

1   AI is priced for the few

Run the arithmetic for a company of 500 people. At $20 per employee per month, below today’s list prices for the leading business add-ons, covering everyone costs $120,000 a year. For a services firm in India, that is the salary of several employees, spent on software for all of them. Few companies make that trade. They ration instead: seats for managers and a handful of enthusiasts, and nothing for the support agent, the operations clerk or the sales coordinator who would gain the most from help with routine writing and reading.

Figure 2. Covering every employee in a 500-person company (prices assumed)

When AI is priced per seat in dollars, an emerging-market company does not decide how to use it. It decides who may use it. Metered pricing makes this worse. A finance team that cannot predict the bill caps the usage, and the cap falls hardest on the people with the least authority to ask for an exception.

The market is not standing still, and an honest argument has to start there. The largest office-software vendors are folding AI into what companies already buy. One has bundled its AI assistant into its business plans, funded by a base-price increase of $2 to $4 per user a month rather than a separate add-on; the other has offered promotional prices on its business AI add-on. At the same time, employees without a company seat do not wait. Many paste work into free consumer tools, with no governance over what leaves the company or what comes back. The competitor is not only the $30 seat. It is AI bundled into software a company already pays for, and free tools its employees use without asking.

So the opportunity is not to introduce AI where none exists. It is to offer a coherent, governed and affordable alternative to fragmented tools and uncontrolled use. A bundled assistant serves the company that runs its whole office on one vendor’s suite. Many emerging-market firms do not: email may sit with one vendor, documents with another, accounts in a local system, customers on a messaging app. What they lack is AI that works across that mixed environment, keeps the company’s memory under its own control and costs a predictable amount. That is the gap this essay is about.

2   Four jobs, one suite

Most everyday AI use in a company reduces to four verbs. Ask: a private assistant that answers questions from approved company knowledge. Act: a mail and work agent that drafts replies and follow-ups, with nothing sent without a person’s yes. Listen: meetings transcribed and written up, ideally on the machine that recorded them. Watch: company records checked against rules, with each exception sent to its owner, with evidence.

None of these is new on its own. Self-hosted chat over company documents is already available as open-source software, and cheap tools exist for each of the other three. Four inexpensive applications would not be much of a business. What changes the picture is the work that passes between them.

Take a single meeting. On Tuesday morning, a team agrees three actions. The Listen job writes up the meeting and records who promised what, by when. The Act job drafts the three follow-ups, and each goes out only when its owner approves. On Friday, the Watch job checks the records: two actions are done, one has slipped, and its owner is told, with the evidence. The following week, a manager asks where Tuesday’s actions stand, and the Ask job answers from the same records, showing only what that manager is allowed to see. Nobody copied anything from one application to another.

Figure 3. One meeting, four jobs, one shared record

Separate subscriptions can be connected, of course, through integrations. But integration across four vendors is usually partial, expensive to maintain and fragile when any one of them changes. A shared architecture makes the connection the default rather than a project. The value is not four inexpensive AI applications. It is one inexpensive system that remembers what happened across them and helps finish the work.

What runs underneath matters less than buyers assume. The leading frontier models are proprietary; their weights cannot be installed on an office server. Open-weight models can be, and they are closer than most buyers think. Epoch AI finds that the best open-weight models lag the closed frontier by about four months on its capability index, and that open models able to run on a single top-end consumer graphics card typically match the frontier of six to twelve months earlier. Last year’s frontier now fits on a well-equipped desk. Not on every desk: that finding is about a high-end graphics card, not an ordinary office laptop. But narrower models doing narrower jobs reach far more modest hardware, and for summarising a meeting, drafting a reply or answering from a policy manual, that is often enough. It has to be proved task by task rather than assumed. The company is buying completed work, not a model name.

3   The office already owns much of the infrastructure

Every office already owns an AI computing infrastructure. Most of it has never been organised to work together. It sits on desks: well-specified laptops and desktops, some with enough memory to hold a capable open model, idle for much of the day while their owners are in meetings, at lunch or at home.

The architecture that uses it has three parts, and the order of importance matters.

  • An always-on office server provides the service guarantee. It routes every question, holds the models and their updates, and runs anything sensitive. If every employee laptop were switched off, the service would still work.
  • Idle employee machines contribute optional capacity. They take eligible, independent tasks while their owners are away and the machines are plugged in, each running a whole model so that nothing is split across machines.
  • The router uses an idle machine only when it helps. It sends work there only when the gain is measurable and the owner is unaffected; otherwise the server does the work.

In The Compute You Already Own I set out why such an office pool is plausible where a public pool of strangers’ devices is not, and the questions that decide whether it pays. The most important is timing. What matters is not total idle computing power. It is dependable compute available in the hours employees need AI. An office might have twenty idle machines at lunchtime and three at eleven in the morning, precisely when questions peak. If one dedicated server handles the load more cheaply, use the dedicated server. The business does not depend on proving that distributed office compute always wins.

It is tempting to call this an “Airbnb for compute”. As an image, it works: an underused asset put to productive use. As an operating model, it misleads. Airbnb rents strangers’ property to strangers. This is a private pool inside one company, among colleagues, managed by IT.

The economics need stating carefully too. On hardware the company already owns, the marginal compute cost of an answer is mostly electricity. The service around that compute is not free. Models have to be kept ready in memory, machines’ availability detected, work moved when a machine disappears, employee endpoints secured and failures diagnosed. Those costs belong in any honest comparison, and they are the reason the dedicated server comes first.

4   The meta-model decides what runs where

If every question goes to a large model, the system is expensive. If every question goes to a tiny one, it fails on the hard cases. The answer is a meta-model in the practical sense: a routing layer that chooses the cheapest way to complete each task correctly. It follows the ladder from The Compute Strait. Code handles dates, sums and rules. Validated memory answers what is already settled. A small local model extracts, classifies and drafts. The office server’s larger model handles harder shared work. A premium external model is called only for the cases that need it, and only where policy allows that data to leave the approved environment.

The router must itself be cheap: rules, capability checks and past evaluations first, never an expensive model asked whether an expensive model is needed. And it should record what it learns: completion, errors, corrections, response time and cost per completed task, by type of work. Over time the company comes to own an evidence-based map of what its own work requires.

Figure 4. One suite, one shared layer, compute the office owns before compute it rents

5   A tenth of the price, if the economics hold

The commercial target should be bold: roughly a tenth of developed-market per-seat prices, about $2 to $3 per employee per month for everyday workplace AI. Four things make it plausible. Open-weight models carry no per-token licence. Much of the compute is already owned. One shared layer serves four jobs instead of four vendors each building their own. And frugal routing keeps premium spending to the cases that need it.

A low price has to work for the supplier as well as the buyer, and that is where the argument is most exposed. At $2.50 per seat, a 500-person company brings in $1,250 a month and a 50-person company $125. Against that sit installation, updates, security patching, monitoring, compatibility testing and support. Those costs do not shrink in proportion to headcount. Supporting a 50-person office does not cost a tenth of supporting a 500-person one.

Figure 5. Supplier economics per office at $2.50 a seat (all costs assumed)

Office size Monthly revenue Cost to serve Monthly margin
50 seats $125 $425 –$300
200 seats $500 $500 $0
500 seats $1,250 $650 +$600
2,000 seats $5,000 $1,400 +$3,600

Assumes $400 a month fixed per office (updates, monitoring, security, support) plus $0.50 per seat. Illustrative, not measured.

Three design choices follow. A tenth-price seat needs a minimum commitment, either a seat floor or a fixed office fee, so that small offices cover the cost of serving them. It needs standardised installation and predominantly self-service operation, because a support visit can wipe out a year of margin. And it probably needs local partners who install and support it, paid from the hardware and services they sell rather than from the subscription.

The buyer’s comparison must be complete too. For the 500-person company, the full cost includes the subscription, the office server, IT time and approved premium use.

Figure 6. The buyer’s full annual cost, 500 employees (all inputs assumed)

Item Assumption Annual cost
Subscription $2.50 per employee per month $15,000
Office server $4,000, written off over three years $1,333
IT time Two hours a week at $15 an hour $1,560
Premium AI Administrator cap $3,000
Total   about $20,900

Against $120,000 for developed-market seats, that is about a sixth, not a tenth, once everything is counted. Against AI bundled into an existing suite, the money may be similar, which is why price cannot be the whole case. The case is coverage across a mixed environment, memory the company keeps and control over where data goes. The $2 to $3 seat is a design constraint to be proved in pilots, not a price the architecture has already earned. The test is cost per completed task, including hardware, IT effort and human rework, not the licence fee alone.

A flat price also cannot mean infinite work for infinite users. It means a predictable price per employee, with fair-use and concurrency limits set by the office’s own capacity: when the office is busy, answers queue or slow down; they do not generate a bill. Premium external inference is different. It should be approved by the administrator, priced separately and visible before it is spent. A flat price is a promise about the bill, not about unlimited compute.

6   Sovereignty is an operating promise

For a bank, a manufacturer, an insurer or a services firm, the attraction is not only cost. It is control. Questions, documents and answers can be processed on machines the company owns, under permissions the company sets.

The promise needs precise wording. “No data leaves the building” is too narrow: companies have remote employees, several offices, existing cloud software and, sometimes, a good reason to use a premium external model. The real promise is this: no enterprise data leaves the approved processing environment without explicit policy authorisation. That environment may span offices and a private cloud the company controls. Telemetry, software updates, connectors to other systems and premium escalation can all carry data out unless they are configured not to, which is why the promise is a policy to enforce and audit, not a slogan inherited from where the server sits. Sensitive classes of work, such as legal, HR and finance, should run only on machines IT controls, by default rather than by request.

The record-keeping needs the same precision. Storing every raw prompt indiscriminately would defeat the purpose. But a ledger that records nothing but usage is too thin for work that needs evidence. Audit records may need source references, decisions, approvals, model versions and access logs, kept under retention rules the company sets. The objective is that the enterprise keeps control of sensitive information and can prove where it went, who saw it and what was done with it.

Regulation is a reason for some buyers, not all. In India, the data-protection law does not impose blanket localisation: its cross-border framework allows transfers except to countries the government restricts, while sectoral regulators such as the RBI can impose stricter rules. Local processing can simplify governance and satisfy contractual or sectoral requirements. For most firms, it is a choice about control rather than a statutory necessity, and the essay should not claim otherwise.

Open source helps, but it is not a guarantee. Code that can be inspected makes a backdoor findable, not impossible. The stronger promise is verifiability: published code, signed releases, builds that independent parties can reproduce, documented components and external audits, practices of the kind NIST’s Secure Software Development Framework sets out. Trust comes from what can be checked, not from what is promised.

7   The layer that lasts

Underneath the four jobs sits the asset that matters most: the shared operating layer. It records what each application knows, what each employee is authorised to see and do, where each task should run and whether completed work met the agreed standard. Applications are replaceable. A company might change its meeting writer, its chat assistant or its local model next year. It should not have to replace its memory, its policies or its operating history when it does.

This is the argument of From Rented to Owned, carried into the office: own what compounds, rent the rest. Here, what compounds is the record of the company’s work and the rules around it. The models underneath can be swapped as better ones appear, and the routing adjusts to use them.

It also clarifies what kind of business this is. A suite of cheaper substitutes competes on price. A trusted, portable operating layer that helps several applications work together competes on continuity, convenience and control. The first can be undercut by any incumbent willing to cut a price. The second becomes more valuable the longer a company uses it.

8   The business model

The opportunity grows in three levels. At the first, a company saves on AI bills by running routine work on hardware it owns. At the second, it gains private AI infrastructure: local routing, permissions, memory and controlled escalation, valued for control as much as cost. At the third, several applications share one operating layer for office intelligence. The lasting advantage sits at the third level, and it comes from making many applications cheap and dependable to run, not from idle processors alone. It would be a mistake to claim all three before proving the first.

Figure 7. How the business earns, and what it refuses

Revenue line How it works Why it fits
Base subscription A flat price per employee, about a tenth of developed-market seats, with a seat minimum or office fee Predictable for the buyer; covers everyone; covers the cost to serve
Premium escalation A disclosed commission on approved external inference, capped by the administrator Pays for hard cases without hidden metering
Hardware and partners The always-on office server, sold or leased, installed and supported by local partners One machine turns idle capacity into a dependable service
Advertising Not offered A product sold on keeping data inside cannot be paid for by anyone who wants to know what is in it

Two cautions belong in the model itself. If the escalation commission becomes the main source of margin, the incentive tilts towards sending work outside; disclosure and caps exist to prevent exactly that. And the natural first market is emerging-market enterprises, where the price gap is widest, software environments are most mixed, local languages matter and data rules are tightening. Those buyers do not need a lesser version of AI. They need a different economic architecture for delivering it.

9   How the thesis could fail

Three bets should be tested separately. Product fit: does one integrated suite complete everyday work better than separate tools or the AI bundled into existing software? Economics: does a seat at around a tenth of developed-market prices, with a minimum commitment, cover models, support, updates and margin? Trust: will security and compliance teams approve a shared local service, and will employees use it without feeling their machines or their data have been commandeered?

Figure 8. Eight ways the thesis fails

Failure What would show it
Bundled AI is good enough Companies on one vendor’s suite see no gain worth switching for
Open models fall short on real office work The share of tasks escalated to premium models stays high, or human rework rises
The shared record adds little Users rarely move work between the four jobs; the suite behaves like four separate tools
The idle pool adds nothing One office server matches the pool at equal cost; then sell the server and drop the pool
The price cannot fund the service Cost to serve exceeds revenue even above the seat minimum
Buyers trust global brands more than local control Pilots convert slowly even where the product performs
Escalation becomes the real business Commission revenue outgrows subscriptions
Global vendors cut prices for emerging markets Developed-market tools arrive at local prices, removing most of the gap

A ninety-day pilot in one office of 20 to 50 consenting users would answer much of this. Run the four jobs on real work against current practice, including whatever bundled or free AI tools employees already use. Serve them three ways: one dedicated office server, the server plus an idle-machine pool, and a hosted small model. Measure completion, errors, response time, availability at peak hours, human rework, energy, the share escalated outside, the share of work that crosses between jobs, the supplier’s cost to serve and cost per completed task. Set the stop rules before starting.

If the suite needs constant human rescue, the shared record adds nothing, or incumbents close the gap on price and control, this particular route is wrong. The need it addresses does not go away.

**

The first wave of AI sold extraordinary intelligence to those who could afford the invoice. The next should make dependable intelligence ordinary, private and affordable across whole organisations, in the countries where most of the world’s office workers live. The answer is not merely to make intelligence cheaper. It is to make it a shared utility inside the enterprise: available to every employee, across applications, with the company keeping control of its data, its memory and its costs.

Affordable for one person is a subscription. Affordable for everyone is infrastructure.