Agents will run the instances. Humans will own the outcomes. Customers will bring agents of their own. And the advantage will move from software to context, trust and accountability. Eight predictions, written down so that they can be scored — and so that they can embarrass me.
It is 8.02 on a Tuesday morning in March 2027, and Maya has opened her laptop.
There is no campaign dashboard waiting for her. Overnight, her system has moved 18,400 customers into weakening attention and 3,100 into lost. It has found a replenishment play running eleven per cent ahead of its control group and widened it. It has flagged that paid reacquisition of customers already in the database rose yesterday for the third day running. And it has escalated two decisions to her, because both of them breach the pricing guardrail she set in January.
She does not ask what campaigns are going out today.
She asks: where are we losing money?
To understand why her morning looks like that, here are eight things that get settled in 2027.
0
Where we stand, August 2026.
Martech today has agents in it. Almost every serious platform now ships something described as agentic, and some of it is real. Salesforce is describing collaborative AI marketing teams where a marketer sets goals, budgets, guardrails and autonomy limits. Adobe has an orchestration layer coordinating purpose-built agents across customer-experience workflows. Braze is pushing continuous one-to-one decisioning across offer, channel, timing, frequency and creative rather than executing predetermined journeys. Shopify is syndicating merchant catalogues into AI shopping surfaces and reporting that AI-driven traffic to stores has grown many times over year on year. OpenAI has published a commerce protocol.
And yet almost all marketing still runs the way it ran in 2019. It runs on campaigns, on static segments, on journeys built once and left alone, on input pricing, and on rented attention. The agents have arrived; the operating model has not moved to meet them. That gap is what 2027 closes, at the frontier.
| A caveat that buys the right to extrapolate.
This is not a prediction that every marketing department will work this way in 2027. It is a picture of the frontier — capabilities already visible in 2026, assembled into the operating model they point towards. Maya is a leading enterprise, not the median brand. |
The brake is worth applying properly, because the noise around agentic commerce is far ahead of the behaviour. Most current agentic experiences are still conversational rather than transactional, and humans still control the overwhelming majority of purchases. Consumer trust in AI remains low — recent research puts the share of consumers who completely trust it in the low teens. Anyone forecasting a wholesale transfer of purchasing to machines by next December is selling something.
So the useful question for a year this near is not what becomes possible. It is what becomes indefensible — which habits a competent CMO will find hard to justify out loud, in a budget meeting, in front of a CFO who has read the same articles.
None of these becomes illegal in 2027. Each becomes awkward to defend, which is a stronger force.
That is the frame for everything below. Not a wave of arrival. A slow withdrawal of excuses.
1
Eight things that get settled in 2027.
What follows is not a trends list. A trends list cannot be wrong, which is why nobody ever revisits one. Each of these is written as a claim that could fail, with the test that would establish it stated alongside. I will mark them in public in December 2027, on this blog, against these words.
Resolution 1 · The campaign stops being the primary unit of planning.
The claim is not that campaigns disappear. Campaigns will still exist in 2027, in enormous numbers. The claim is narrower and more consequential: at leading brands the objective becomes the thing a human specifies, and the campaign becomes a generated execution artefact — something the system assembles on the way to the goal, rather than something a person writes and approves in advance.
The same business and the same quarter, briefed twice.
In the first panel the human has chosen the audience, the trigger and the timing, and implied the message. In the second the human has chosen the outcome, the economics and the limits, and left the rest to be determined per person. Less specifying, more governing.
The analogy is manufacturing. A factory still has production runs, but the production run is not the intelligence — the intelligence sits in the system deciding what to make, in what sequence, under what constraints. The campaign becomes the production run.
This is underway rather than speculative. Salesforce’s goal-driven marketing agent already lets a marketer define a goal, a budget, guardrails and an autonomy limit and then determines audience, content, channel and timing within them. What 2027 settles is how much of the estate moves.
How it gets scored. By the end of 2027, can a CMO at a leading brand set a commercial goal plus guardrails and have the system determine a meaningful share of audiences, treatments, timing and channels — with the campaign artefacts generated rather than authored? If the answer is still no anywhere outside a pilot, I was wrong.
Resolution 2 · The segment loses to the customer decision.
Segmentation was never a philosophy. It was a compression algorithm — the cheapest workable approximation to a problem nobody could afford to solve properly. One team could not make five million separate decisions, so it grouped people who were not alike, sent each group the average, and called the compromise a strategy.
That constraint is what has changed, so the unit of execution moves down: from segment, to customer, to individual decision. Segments survive as a lens — for reporting, governance, strategy and explaining a business to a board. They stop being the instruction.
And the most important thing now decided per person is the least glamorous one on the list: whether to contact them at all. Old personalisation meant producing more variants. Agentic personalisation can mean producing fewer interventions, because the system can see that some customers need no push. That single capability changes the economics more than any content model.
How it gets scored. What proportion of customer interventions at a leading brand are selected dynamically by a decisioning system, rather than assigned by a human-built segment or journey? If that proportion is still in single digits at the end of 2027, this resolution failed.
Resolution 3 · The CDP does not die. It disappears below the waterline.
The tempting version of this prediction is that the customer data platform category collapses. That is too strong for a single year and probably wrong in any timeframe. None of that work becomes less important. It becomes less visible. The CDP stops being the interface and becomes the substrate — indispensable, expensive to get right, and mostly invisible to the user, like the database under a modern application.
The deeper change is from profile to context. A profile tells you what is known; context tells you what matters now. Not only that she bought three months ago, but whether the item is replenishable. Not only that an offer was sent, but why it was sent, what alternative a human rejected, and what happened next.
So the buying question changes. Not which model do you use — the models will be the same models, available to everyone, at falling prices. It becomes: what does the system remember, and can the agent use it? That is a hypothesis rather than a moat anyone can declare. The same model is not the same marketer if it remembers a different history. Salesforce’s 2026 research on Indian marketers already found disconnected data limiting how far marketers trust AI to act; the bottleneck has moved from model capability to what the model is allowed to know.
How it gets scored. Do enterprise martech RFPs in late 2027 contain material sections on memory, decision traces and agent data access — as distinct from data ingestion, identity resolution and segmentation? If the RFP still reads like 2024, this was wrong.
Resolution 4 · AI-mediated demand becomes a measured channel.
This is the most important of the eight, and the one that cuts hardest against my own writing.
Every model of marketing any of us has ever drawn has two ends: the brand at one, a human at the other. The whole apparatus — persuasion, creative, subject lines, timing, relationship — assumes a person is at the far end, capable of being interested, flattered, reminded or moved.
The second diagram is the one nobody’s marketing plan is written for.
A consumer tells her assistant what she wants, in her own terms — running shoes under a certain price, comfort mattering more than weight, nothing from brands with poor returns policies. The agent discovers, compares, filters and shortlists. The brand may never get the chance to target her in the old sense at all.
Be careful about how far to push this for 2027. Direct agent-to-agent negotiation is still largely developmental; Shopify says as much itself. The defensible frontier prediction is one step earlier in the chain: that leading commerce CMOs will explicitly measure AI-mediated discovery, recommendation and transaction as a separate source of demand, with its own reporting line, rather than letting it hide inside referral or direct traffic.
The strategic consequence, though, arrives well before the volume does. If part of your audience is a machine, then part of marketing becomes making the brand legible and trustworthy to machines. Product facts. Availability. Price. Reviews. Returns policy. Reputation. Machine-readable context. None of that is advertising, and most of it is not owned by the marketing department at all.
For twenty years marketers competed for human attention. In 2027 they also compete for machine selection.
Now the part that runs against my own doctrine, and specifically against the piece of it I like most.
I have argued for a year that the way to fix email is to earn the open on the day you are not selling — the daily digest, the useful note, the message carrying no offer whose only job is to rebuild the habit of opening. That argument rests on human attachment. It assumes a person who can be pleased, who forms a habit, who comes to expect something.
My first instinct was that a machine reader kills it. The agent forms no habits and feels no warmth; it compares your product facts against five competitors in a millisecond and moves on. That instinct is too binary, and it is worth correcting carefully, because it is the difference between a real argument and a scare.
The machine has no affection. Its principal does.
An agent representing Priya knows that she prefers certain brands, trusts a particular bank, will not buy fast fashion, reads one publisher and has had excellent service from a retailer twice. Brand attachment does not evaporate because an agent mediates the decision. It becomes an input into the utility function the agent is optimising — arguably a more durable input than it was, because the agent will apply it consistently where a distracted human might not.
So the tension is not human relationship against machine facts. It is persuading the human against becoming legible to the machine that represents the human’s preferences. Those are two jobs, not a replacement of one by the other.
Which produces a more useful conclusion than my first one. Relate still matters, because it shapes the preference the agent will eventually represent. But when the purchase moment arrives, sentiment will not rescue bad price, bad stock data, bad service history or an unverifiable returns promise. Affection gets you into the consideration set. Operational truth gets you selected out of it.
That also makes marketing to agents a great deal more interesting than search engine optimisation for machines, which is how most of the industry is currently reading it.
The second gate.
There is a mechanical consequence of all this that I underplayed when I first wrote it down, and it is more immediate than the volume of agent-mediated purchasing. Today a message passes through one gate. The mailbox decides whether to accept it, on sender reputation, authentication and complaint rate — a gate every competent brand already manages, with known levers. A triaging agent adds a second gate on top, and the rule is different in kind.
The second gate is decided by the reader’s own history with that sender. Did she open the last twelve? Did she act on any of them? Has she ever replied? A brand whose base has quietly decayed does not merely get ignored more often. It stops being surfaced, which is a different and worse condition, because ignoring leaves the message on the screen and surfacing does not.
Which changes what the retention metrics are for. Click retention rate and Real Reach have been diagnostics — numbers a careful team watches to understand whether its base is eroding. In an agent-triaged inbox they become an access right. The engagement history is the thing that decides whether the next message is shown at all, and it cannot be bought at the point of need, because by then the history either exists or it does not.
Two gates, and the brand only sets the rule on the first one.
The same shift removes something the industry has never had to name, because it was free. Every programme has quietly collected attention nobody earned — the glance while deleting, the subject line read on the lock screen, the half-second before the swipe. An agent removes that residue entirely. What remains is voluntary, all of it, which raises the value of voluntary attention and takes away the floor underneath everything else.
Agents do not reduce attention. They remove the attention nobody earned.
That is also the structural reason the relationship email survives, and it is a better reason than the one I have been giving. A Sell message contains a completable task, and so does a Notify: compare, decide, confirm, track. Those are precisely what an agent exists to finish, and once it finishes them the human never needs the message. A relationship message contains no completable task. There is nothing in it for an agent to resolve on somebody’s behalf, so it either passes through to the person or it does not exist at all. Being undelegable turns out to be the property that matters, and very little brand email has it.
What follows from both gates is a design requirement rather than a positioning one. A message now has two readers and needs two payloads: something machine-actionable, so the agent can verify the price, the stock, the returns window and the provenance without guessing; and something human-experienceable, so that when it is passed through there is a reason for a person to be glad it was. Most brand email today has neither in any structured form. It has prose that assumes a human and metadata that assumes nobody.
The blur nobody has priced yet.
There is one implication hiding inside this that deserves stating, because it touches the map directly.
If customers carry agents, acquisition and retention start to blur. A customer’s agent may hold the memory of a prior relationship even when the customer does not actively recall it — that she bought from this brand before, that returns were straightforward, that size eight fitted, that delivery took two days. Your CRM has her in the lost column and has written her off. Her agent has not.
Which is a new state, and the grid has no cell for it: the brand may have lost the attention without losing the consideration. What that is worth, and whether it decays, nobody knows yet. It is the first thing I would want measured once AI-mediated demand is reported as a channel at all.
How it gets scored. By the end of 2027, do leading commerce brands report AI-mediated demand as a named channel in their own reporting — with a share of sessions or orders attached? If it is still lumped in with direct traffic, this was wrong.
A second test on the same resolution. Do any leading brands begin treating their own engagement history as a placement input rather than a retention report — suppressing sends to protect surfacing rights, or reporting Real Reach alongside deliverability? If nobody connects the two by the end of 2027, the second gate arrived later than I thought.
Resolution 5 · Outcome pricing splits into the real thing and a counterfeit.
The counterfeit will arrive first, and it will arrive in volume.
It will be marketed with exactly the vocabulary I have been using for a year — outcomes, accountability, skin in the game, pay for performance. And underneath the vocabulary the vendor will define the outcome, own the attribution model, and invoice against its own scoreboard. There will be no control group anywhere in the arrangement, because a control group is the one thing that would make the invoice smaller.
The vocabulary will be identical. The mechanism will not be.
Naming the counterfeit early is the only way to protect the real thing. Once a category has been sold a fake version at scale, the real version has to spend years arguing that it is different, and it usually loses that argument to whoever got there first with a bigger sales team.
If the vendor controls both the treatment and the counterfactual, it is not outcome pricing. It is performance-labelled attribution.
The test is that simple, and a CMO can apply it in a single meeting. Ask who defines the outcome. Ask who holds the control group. Ask whether the vendor can see and influence the counterfactual. Three questions, and every honest vendor will have prepared for them.
How it gets scored. By the end of 2027, is there a visible split in the market — with some vendors selling outcome-linked contracts backed by concurrent randomised controls and others selling attribution-model pricing under the same language? And has at least one credible buyer publicly rejected the second on methodology grounds?
Resolution 6 · Governance becomes runtime infrastructure, not a committee.
The theatrical version of this prediction is that some brand suffers a public agentic failure and explainability becomes a board-level topic overnight. That may happen. It is also unnecessary to the argument, and predicting disasters is a cheap way to sound serious.
The stronger claim is structural. By 2027, any serious autonomous marketing system needs identity, permissions, decision traces, escalation rules and a human veto as things the system executes at runtime — not as a document in a compliance folder. Adobe is already positioning its agentic architecture explicitly around governance and auditable workflows, which is a reasonable indicator of where the buying pressure is coming from.
The question a governance review asks changes accordingly. Not: what is our AI policy? But: which agent is acting and for which objective; what data, customers, channels and actions may it touch; what budget, discount, frequency and margin limits apply; which decisions need a person; where is the veto; and how do we reconstruct afterwards what it knew, what it chose, what it rejected and what followed?
Every one of those is a runtime question. None of them can be answered by a document.
That last clause is the load-bearing one. Explainability is not a compliance nicety and it is not about trust in the abstract. It is the mechanism by which autonomy gets earned. A system whose reasoning cannot be inspected does not get given more scope, and should not be. The progression — recommend, then propose, then execute with approval, then operate inside broader guardrails — is the same one every technology we have learnt to trust has walked.
How it gets scored. By the end of 2027, do enterprise agentic marketing deployments ship with runtime permissioning, decision traces and escalation as standard product surface — or is governance still a slide in the security review?
Resolution 7 · Opens lose their last claim to being an economic currency.
The temptation here is to predict that some universal verified-human-engagement standard emerges. That is too specific for a single year and I am not going to claim it.
The narrower claim, and it needs stating precisely because the sloppy version is wrong: machine-contaminated opens lose economic currency. Verified human attention remains diagnostically useful. Verified human action becomes the commercial currency.
Those are three different things and the industry keeps collapsing them into one. Privacy proxies, image pre-fetching and security scanners have made the raw open unreliable as evidence a person saw anything. But an open confirmed as human is still worth knowing — deliverability teams need it, publishers need it, and attention has value before conversion. What it cannot do is carry an economic claim.
Which produces a hierarchy rather than a single metric.
Six rungs. Real Reach and CRR live on the second; invoices belong on the sixth.
The rule that falls out of it is simple: the higher the economic claim, the higher the proof standard it has to clear. A deliverability report can rest on renders. A retention diagnostic can rest on verified human action. A supplier payout cannot rest on anything below incremental value.
There is a second reason to care about the rungs, and it did not exist when the ladder was first drawn. If a triaging agent decides what gets surfaced on the basis of a reader’s history with a sender, then the second rung stops being purely diagnostic. Verified human action is what the access right is made of. A brand that has spent a decade generating renders and calling them engagement will discover it has no record of the only thing the gate reads.
This also matters for what a decisioning system learns. Reward the machine for opens and it will get extremely good at generating opens — a failure mode with no human equivalent, because no human team was ever fast enough to fully exploit a bad metric. Reward it for a useful action, a declared preference, a state movement or an incremental transaction, and it optimises something worth having.
How it gets scored. By the end of 2027, have leading brands moved their headline engagement reporting off opens and onto verified actions? A partial test: does the open rate still appear in the board pack?
Resolution 8 · The vendor begins carrying the outcome.
The frontier martech contract moves from pay me to operate towards pay me partly for what moves.
Not all software becomes outcome-priced. That will not happen in 2027 or in 2037, and predicting it would be silly — plenty of software should be sold on access, because plenty of software delivers value that cannot be cleanly isolated. The claim is about the frontier: buyers increasingly distinguish between software that claims value and partners willing to expose some of their own economics to proof.
A fixed baseline will often remain, because delivery and infrastructure cost something before any lift exists. What enters the relationship alongside it is a share of the measured difference. The strategic change is not the percentage. It is the transfer of risk: traditional licensing hands operating responsibility to the buyer and pays the vendor whether the system was used brilliantly, badly or barely at all. An outcome-linked contract hands some execution and measurement risk back.
How it gets scored. How many major martech deals contain measurable outcome-linked components, shared baselines or incrementality clauses by the end of 2027? If the answer is a handful of pilots and nothing structural, this resolution was premature rather than wrong — but it was still wrong for 2027.
The scorecard. Eight claims, eight tests. Marked in public in December 2027.
| Resolution | The test in December 2027 |
| 1 · Campaigns | Can a CMO set a goal plus guardrails and have the system determine a meaningful share of audiences, treatments, timing and channels? |
| 2 · Segments | What share of interventions are selected dynamically rather than assigned by a human-built segment or journey? |
| 3 · The CDP | Do enterprise RFPs contain material sections on memory, decision traces and agent data access? |
| 4 · Machine demand | Do leading commerce brands report AI-mediated demand as a named channel, with a share attached? |
| 5 · Outcome pricing | Is there a visible split between control-backed contracts and attribution-model pricing under the same language? |
| 6 · Governance | Do agentic deployments ship with runtime permissioning, traces and escalation as product — not policy? |
| 7 · Opens | Has headline engagement reporting moved off opens and onto verified actions? Is the open rate still in the board pack? |
| 8 · The vendor | How many major deals contain outcome-linked components, shared baselines or incrementality clauses? |
2
Maya’s Tuesday.
The best way to see what changed is not to walk through a CMO’s day hour by hour. It is to look at the meetings that vanished from her calendar, and what replaced each one.
| The meeting Maya had in 2026 | The meeting Maya has in 2027 |
| Campaign calendar review | Outcome portfolio review |
| Segment approval | Guardrail approval |
| Creative variants meeting | Brand judgement meeting |
| Channel performance review | Route-tax review |
| Attribution argument | Holdout readout |
Each swap is small on its own. Together they change what a chief marketing officer is.
8.30 a.m. — The campaign calendar review that vanished.
The meeting used to take forty-five minutes every Tuesday. Merchandising brought the priorities, CRM brought the audience, creative brought the assets, and the channel owners argued about cadence. Someone worried that email and WhatsApp were hitting the same people. Someone asked whether the sale needed another reminder. The team left with a calendar.
That meeting no longer exists. What replaced it is an outcome portfolio review, and the difference is visible in what is on screen. Five numbers across the top: second-purchase rate, repeat margin, Real Reach, reacquisition share — a count of customers her paid channels reported as new who were already in the database, not a split of spend — and route tax. Underneath, the grid — where customers are moving between strong, weakening and lost attention, and which pools are leaking the most money.
Maya does not approve a send plan. She changes priorities. Protect the high-value customers whose attention is weakening. Grow the one-time buyers most likely to reach a second purchase. Reduce paid reacquisition of customers already known to the brand. Hold total discount cost flat. Suppress anyone likely to transact without help.
Those are management decisions. The system turns them into instances.
10.00 a.m. — Segment approval becomes guardrail approval.
Maya used to approve segments, because segmentation was where risk entered the system. Was the cohort too large? Did it include people it should not? Was it fair to give one group a benefit another would not get?
Now she approves the boundaries instead. The agent may choose from a very large action space, but it cannot invent its own economics: the margin floor, the frequency caps, the contact exclusions, the product eligibility, the protected cohorts and the actions requiring human review are all set by her team.
One decision is waiting. The system has found a group with high purchase intent and proposes a twelve per cent incentive. The model expects it to lift conversion, and the model is probably right. The expected incremental margin after the discount sits below Maya’s threshold.
She rejects it. This is the cleanest illustration of the new division of labour I can give: the model is better than she is at predicting the response. She remains accountable for deciding whether the response is worth buying.
11.00 a.m. — The brand meeting, which got longer.
This is the part that gets underplayed in every account of agentic marketing, including some of mine. Freed from operating the machine, Maya spends more time on product, pricing, story, customer experience and competitive position — the things no model can decide for her because they are not prediction problems, they are taste and strategy problems.
The pattern is familiar from every previous automation of a professional craft. When the mechanical part of a job becomes cheap, the judgement part becomes more valuable, not less. AI makes human judgement scarcer relative to everything else, and scarcity is where value goes.
1.00 p.m. — The CFO meeting, where nobody argues about attribution.
This is the meeting that has changed most, and it is the one that would most surprise a 2026 marketer watching.
There is no argument, because there is a control group. Maya and the CFO look at the baseline, the measured alpha, the carry paid on it, reacquisition share, Real Reach, customer-state movements and incremental margin. Nobody defends a model. Nobody claims a conversion that a platform also claimed. The conversation is short, because the disagreements that used to fill it were disagreements about measurement, and the measurement is no longer in dispute.
A second programme on the same screen is more interesting, because it produced no measurable alpha at all. In 2026 the team would have presented its attributed revenue and called it a win. The holdout shows that most of those customers would have come back anyway.
The CFO’s response is not disappointment. It is relief — because they have just found spend they can stop. That may be the deepest cultural change in Maya’s organisation, and it takes a year of holdouts to earn: doing nothing can now be reported as a successful decision.
Marketing finally has a language the finance function already speaks. That is worth more to the marketing department than any agent in the stack.
3.00 p.m. — The agent problem, which is not a marketing problem.
Traffic from AI shopping assistants is rising, but the brand is being selected less often than two competitors on a popular line. The product is well reviewed. Pricing is competitive. Stock is good.
The problem turns out to be returns. One channel says fourteen days, another says thirty, the marketplace copy is ambiguous, and the product feed exposes no machine-readable exception for sale items. A customer’s agent cannot establish a reliable returns promise, so it prefers a competitor whose policy it can verify.
In 2026 Maya would have asked for better ads. In 2027 the marketing fix is an operations fix: standardise the returns promise, structure it, and push it into every feed. Her job that afternoon is explaining to two departments that owe her nothing why a data inconsistency has become a demand problem.
This scene is the one to hold on to, because it is where marketing visibly stops being communications. When part of your audience is a machine that reads facts rather than claims, the truth of the operation becomes the marketing. There is no gap left between what you say and what you do, because the machine checks.
5.30 p.m. — She teaches the system.
Maya reviews five consequential decisions the system made today. A customer was suppressed because predicted organic conversion was high — approved. A recovery message used an urgency phrase that passed the brand rules and still felt manipulative — rejected, with two lines explaining why. A high-value customer was routed to a human rather than another automated offer — approved. A promotion was paused because inventory risk had changed — approved.
The fifth is the one worth watching. An agent proposed raising contact frequency for a cohort whose short-term conversion was climbing while its click retention rate was falling. Maya tells the system to protect the attention rather than maximise this week’s response — which is a judgement about the future that no reward function was going to make on its own.
Each decision and each explanation becomes part of the system’s memory.
This is her new managerial act, and it is worth naming precisely, because it looks like nothing. She is not building tomorrow’s campaign.
She is teaching tomorrow’s marketer.
3
Priya, Rahul and Ananya.
A day in a CMO’s life is only half the picture, and it is the flattering half. The other half is what all of this feels like to the person on the receiving end — which, in most accounts of the agentic future, goes conspicuously unexamined.
So: same brand, same week, three customers, three different correct answers. Two of them end with the brand doing less.
If the only illustration a doctrine can offer is a delighted customer, it is a brochure.
Priya — best, strong, and increasingly left alone.
Priya buys regularly and pays attention. In 2026 martech, that reads as high propensity, and high propensity reads as send more. She is the most contactable person in the database and therefore the most contacted.
In 2027 the system reads the same signal and reaches the opposite conclusion. She is not short of reasons to buy. Every additional message is a withdrawal from an attention account that is already full, and the model can see the withdrawal in her declining response curve long before she unsubscribes.
So her experience gets quieter. Service messages when something needs her. A composed-at-open availability alert she explicitly asked for, which is accurate at the second she reads it because it was assembled then. One useful note a week. Nothing else.
Her lifetime value rises while her message volume falls, and she would not describe any of it as marketing. She would describe it as the brand being unusually good at its job. That inversion — better marketing looking like less marketing — is the single hardest thing to sell internally and the easiest thing to prove.
Rahul — lapsed, recoverable, and the reason the holdout exists.
Rahul bought once, fourteen months ago, and has gone quiet. In today’s operating model the sequence is predictable: CRM tries, CRM gives up, paid media eventually finds him, he comes back through a rented route, and the dashboard books it as a win.
In 2027 the brand’s system knows exactly where he sits — one transaction, attention lost — and the recovery capability gets the mandate before any money goes to a platform. That capability is the one I call Team 6: the team that owns the lost column and nothing else. A brand can run it in-house, or use the outsourced version, Progency, operated by MarTech Growth Engineers working with agents. Either way it is paid for measurable improvement rather than activity, which is the whole reason the next paragraph happens.
The model believes Rahul is recoverable. His category has a reasonable repeat pattern and similar customers have come back after a year. The economics look attractive.
And because a belief is not a measurement, the cohort is randomised. Rahul lands in the control group. He gets nothing at all beyond the brand’s ordinary behaviour.
In April he comes back on his own. He needed the product again, remembered where he bought it, and returned without any prompting whatsoever.
Ninety days after randomisation. The treatment arm did not beat the control.
The treatment arm performed no better than the control. The measured lift sat inside the confidence interval. Team 6 earned no alpha and raised no invoice, and a quarter of work produced no revenue that would not have arrived anyway.
On any conventional reading, that quarter failed. Read properly, it is the most valuable thing that happened all year, because the brand has just discovered that a category of spend it was about to industrialise does not work. Every rupee it would have poured into that cohort for the next three years has been saved by one properly constructed experiment.
The system failed commercially and succeeded epistemically. In marketing, the second is rarer and worth more.
This is why prediction and incrementality must never be allowed to share a currency. The model was not wrong about Rahul — he did come back. It was wrong about its own contribution, which is a different thing, and the only instrument that can tell them apart is a control group.
Ananya — who says no.
| A short note on the customer nobody puts in the deck.
Ananya opens a message that is a little too knowing. Nothing in it is inaccurate; that is the problem. She narrows what the brand may use — no location, no cross-device history, no inference from what she browsed but did not buy. The system complies immediately, does not negotiate, does not offer an incentive to reconsider, and continues working with less. |
Her results get slightly worse. The brand accepts that, because the alternative — treating accumulated context as a licence rather than a loan — is how a company ends up on the wrong side of both a regulator and its own customers.
Knowing more about someone grants no permission to do more. Context is not entitlement. An agentic system that cannot be told to use less of what it knows is not sophisticated; it is simply not under control.
4
The supply side’s P&L.
Everything above describes what changes for marketers and for customers. An essay that stopped there would be dodging its own hardest question, because the entire argument is a claim about how suppliers must change — and I run one.
So here is the uncomfortable version, in structural terms.
Input pricing comes under pressure, because agents reduce the human usage that seat-based and screen-based pricing was implicitly measuring. If nobody logs in, what exactly is the seat for? Value migrates from seats, messages and records towards decisions and outcomes — and the vendors who resist that migration will find their pricing model quietly detaching from the value it used to proxy.
Real outcome pricing creates working-capital exposure. The supplier does the work before the alpha is known, funds delivery upfront and collects in arrears. That caps how many engagements can run at once, for the vendor and for the client’s patience alike. Any model that does not name this constraint is not being straight with you.
The vendor also starts carrying measurement risk on top of execution risk, and these are different things. Execution risk is performing badly and earning less, which is fair and easy to explain to a board. Measurement risk is performing well and discovering the intervention was not incremental — the audience would have converted anyway, the brand’s existing programme was already strong, the customer specified a stricter control than expected. A supplier has to be able to survive quarters where the honest answer is that the work was competent and the lift was zero.
Which changes which deals a rational supplier should accept. We can send this is not enough. We can probably improve this is not enough. What is needed is a declared leakage pool, an agreed current-best-effort baseline, a credible counterfactual and enough expected spread to fund the work before any of it is proved. Outcome businesses have to learn underwriting, and most software companies have never employed anyone who knows how.
Which produces a competitive dynamic worth predicting on its own: a vendor who refuses a concurrent holdout increasingly looks like a vendor unwilling to test its own claim. Not dishonest, necessarily. Just unwilling. And in a room where one competitor has offered a control group, unwilling is not a survivable position.
Two models, optimised for two different kinds of certainty.
SaaS was designed to maximise revenue certainty for the vendor. Outcome software is designed to maximise value certainty for the customer. You cannot fully optimise both.
That is the real trade-off of 2027, and it deserves its own name: ARR quality against alpha quality. A pure licence business has beautiful revenue characteristics — recurring, predictable, recognised on access, scaling at the pace of signatures. An outcome business cannot always scale at that pace, because proof has a clock: data integration, a baseline period, a holdout, an intervention window, outcome maturation. And a disciplined one will sometimes refuse revenue outright, because the leakage pool is too thin or the measurement design is too weak to support a claim.
So the honest formulation is not that outcome businesses grow slowly. An excellent one could grow very fast. What it accepts is less predictable growth and greater working-capital intensity, in exchange for stronger proof of value. That is a real trade, made deliberately, and every supplier heading in that direction should be able to say out loud that they have made it.
The likely resolution is hybrid by design: a predictable base priced for the substrate, and a variable layer wherever the causal chain can be measured credibly. Which means the interesting question about any 2027 vendor is not whether it offers outcome pricing. It is which parts of its business it is willing to expose, and why those parts and not others.
And then the part that neither I nor anyone else selling this future writes down often enough.
What happens to the people who currently operate the campaigns.
Campaign operations is the largest pool of human work in marketing. Somebody builds the segment, briefs the content, configures the journey, checks the render, schedules the send, pulls the report. Multiply that by every brand and every agency retainer and it is an enormous amount of employment.
When the machine makes the instances, that is the first line a CFO questions. Not the strategy retainer. Not the brand work. The execution hours — because those are the ones with a visible unit cost and a visible substitute. To claim otherwise, on the grounds that marketers’ jobs move up rather than out, is true at the level of the individual senior marketer and evasive at the level of the industry.
The danger for an agency is not that creativity disappears. It is that deliverables become cheap. Ten banners stop being a defensible unit of value when a system can produce a hundred. A monthly campaign calendar is worth less when the client can specify a goal and have the plan generated. Reporting retainers shrink when the analysis arrives continuously and nobody has to build the deck.
None of this is free, and it is worth being accurate about that, because the loose version of the argument is easy to attack. Models cost money. Inference costs money. Data, orchestration, quality control and governance all cost money. What collapses is not cost — it is the marginal cost of human execution, and that is the specific thing a retainer priced per campaign built is selling.
So my expectation is that the agencies which survive are the ones that stop selling deliverables and start owning an outcome — the same transition being asked of software vendors, arriving at the same moment, at organisations with thinner balance sheets and far less appetite for measurement risk. The shops that matter more will be the ones willing to say: give us the number, the constraints and the right to be measured. That is a better business if it works and a less forgiving one if it does not.
It is not comfortable to write that from inside a company selling the thing causing it. It would be less honest not to.
5 What does not change.
The failure mode of every essay like this one is that it becomes an advertisement for the future. So here is the floor — six things that will look exactly the same in 2027, and will still decide whether any of the rest works.
None of these is affected by model capability. All of them constrain it.
Economics. A thin-margin transaction cannot fund an expensive intervention, however confident the prediction behind it. The break-even on a recovery is still one divided by your gross margin, and no amount of intelligence moves that number. A system that ignores it will lose money at extraordinary speed and with excellent attribution.
Consent. Knowing more about a customer confers no additional permission to act. The two things are unrelated, and the temptation to conflate them grows exactly as fast as the context layer does. This is the single most likely place for the industry to embarrass itself in 2027.
Deliverability. A message the mailbox rejects cannot be made intelligent enough to matter. It is the least glamorous item on this list and the largest multiplier on every number above it. An agentic marketing programme sitting on poor sender reputation is a very sophisticated way of not arriving.
Brand. Ten million perfectly personalised bad offers are still ten million bad offers. Personalisation improves the fit between an offer and a person; it does nothing whatever about whether the offer was worth making. As content becomes free, distinctiveness becomes the scarce thing — and distinctiveness is a human judgement, made by people like Maya in the meeting that got longer.
Causality. A prediction is not an incremental outcome. This is the error the whole apparatus is built to make, and it will make it constantly, in good faith, at scale. A model that says a customer will return is describing the world. A control group is the only thing that can tell you whether you changed it.
Human nature. People still want relevance, usefulness, status, novelty, trust and convenience — and sometimes simply to be left alone. None of that has moved in fifty years of marketing technology, and none of it moves in 2027.
Which leads to the sentence this whole essay exists to support.
The holdout does not become obsolete as the machines get smarter. It becomes more necessary.
The logic is unavoidable. The more decisions a system makes, the more results it produces, and the easier it becomes to mistake a prediction for a cause. Ten campaigns a quarter could be assessed by argument. Ten million decisions a day cannot be assessed by anything except a control group, because there is no other instrument that can separate what the system did from what would have happened without it. Scale does not weaken the case for randomised measurement. It is the case.
And the three commitments I have been arguing for since this began do not change either, because they are not predictions. They are the floor the predictions stand on. Never lose customers. Never pay twice. Never pay fixed.
What fades, what matters more.
| Fades | Matters more |
| Campaign calendars | Business goals and guardrails |
| Static segments | Customer context and decision traces |
| Journey spaghetti | Continuous decisioning |
| More screens for humans to operate | Capabilities for agents to invoke |
| Manual campaign operations | Human judgement, taste and restraint |
| Content scarcity | Brand distinctiveness |
| Channel silos | Lowest-total-tax routing |
| Last-click attribution | Incrementality against a control |
| Vanity engagement | Movement between customer states |
| Pure fixed SaaS | Outcome accountability, and the risk that comes with it |
| Marketing to humans | Marketing to humans and to their agents |
| Renting every interaction | Owned identity, attention and memory |
Read that table in one direction and it looks like a technology story. It is not. Almost every item in the right-hand column is a governance, judgement or accountability item — things that cannot be bought, installed or prompted into existence.
Which is the surprising shape of the year ahead. Martech 2027 will contain more technology than it ever has, and marketers will spend less time using technology than they ever have. The systems get more complex underneath and simpler above. The marketer specifies the outcome. The agents run the instances. The customer, increasingly, delegates too. And the scarce things become the ones machines cannot cheaply manufacture: judgement, trust, permission, context, brand and accountability.
MarTech 2027 will be defined not by how much AI marketing uses, but by what humans no longer need to operate, what customers no longer need to endure, and what vendors are finally willing to be accountable for.
