Author: Rajesh Jain

Context is the moat — and you pay twice

 Here is the objection that ought to follow, and it is the right one. If agents are what changed, then everybody gets agents. The models are the same models, available to your competitor on the same terms, at the same price, next quarter. Whatever advantage you buy on Monday is commodity by the following Monday.

That objection is correct, and conceding it fully is what makes the rest of the argument work. Intelligence is becoming a commodity, and commodities do not differentiate. Anyone selling you an agent as a moat is selling you a rental agreement.

What is not rented is the context. Only you have this customer’s history — what they bought, what they returned, what they browsed and abandoned, which of your messages they opened at seven in the evening and which they have ignored for four months. It deepens every single day. It cannot be bought at any price, only accumulated. And it is the input the model needs to be worth anything on your particular business.

Which produces the sentence to take away from this section: two brands with identical agents will not get identical results. The one that knows the customer wins. Same models, same prompts, same vendor — different outcomes, because the models are answering questions about different people with different amounts of knowledge behind them.

Now the turn, and it is the reason this edition exists. Context is not bought — it is accumulated, one interaction at a time — and it accumulates only on surfaces where the interaction belongs to you.

Before the comparison, a fair word about the rented surfaces, because the argument is stronger without a caricature. Google finds demand you could not have found. Meta reaches a consumer whose attention you have already lost. Marketplaces create discovery and convenience at a scale you will never build. None of these channels is bad, and none of them should be abandoned. The claim is narrower and harder to argue with: their economics are different, and the difference is not only the fee.

The same transaction, two surfaces. Only one of them leaves you anything to compound.

A rented surface returns a transaction. The platform saw the behaviour, not you. You get the sale and none of the memory. Which means next time costs exactly what last time cost, and the time after that costs more, because the auction inflates. An owned surface returns a customer. Every open, every tap, every configuration and every silence is yours, recorded as a fact attached to a person, consented and governed. The record compounds. Next time costs less and works better.

So the sentence that has anchored this doctrine needs upgrading. Never Pay Twice was always a pricing claim: do not buy the same customer from a platform you already sold to. It is more than that. You pay twice — once in margin, and once in memory. Adtech takes roughly a third of the transaction and keeps what it learned. Meta knows what happened on Meta. You do not. So the next recovery costs you exactly what the last one did, forever, because you never got to learn anything from paying for the first.

That reframes the tax ladder. It has always measured what each route to a transaction costs you. It also measures what each route leaves you with.

Read down that third column and the strategy writes itself. The job is to move transactions down the ladder toward the cheapest proven route, and to spend adtech last — for what nothing cheaper could reach. But the second reason is the one that compounds: every transaction you move down the ladder does not just cost less this quarter, it makes the next one cheaper by leaving you something to learn from.

The two arguments, joined. Agents lift the arithmetic that kept the plays from being run. Context is what stops agents from being a commodity. And context compounds only on ground you own — which is why the owned-channel argument stops being about postage and starts being about the only asset in marketing that appreciates.

You pay twice. Once in margin. Once in memory.

Your job moves up.

 Not out. Up. This distinction matters more than any other sentence in the essay, because it is the one every marketer is privately testing while they read the rest.

The judgement stays exactly where it has always been — with the people accountable for the number. What moves is the execution of the instances.

The line between what a human decides and what an agent runs. Four of these five human items are things no model can decide for you.

Humans own the outcome: which number you are accountable for this year, what a good offer is and what it costs the business, what the brand sounds like, which trade-offs are acceptable, and — the item nobody should ever delegate — what must never be sent and when to stop. Agents own the instances: which of the six plays this particular customer needs, on the channel they still respond to, at the moment they are paying attention, with content assembled for them, ten million times over, every single day.

Underneath that line the architecture is simpler than the vocabulary suggests. There is an orchestrator — a unified co-marketer that you brief and that executes nothing itself; its job is to hold the team, call whichever specialists a task needs, sequence them and hand the result back. There is a decisioning engine, which is not an agent but the thing every agent calls: it scores the next best move for each individual, in real time, against the goal you set.

And there are the specialists. An insights agent that reads every campaign and reports what performed and why. An audience agent that builds and refreshes the segments worth targeting — and refreshes them, which is the part humans never get to. A content agent that drafts from master content, on brand, for hundreds of micro-segments rather than ten. A scheduler that delivers to each person at their moment on their channel. A shopping agent that carries a customer from discovery to checkout. And, increasingly, custom agents a brand builds itself for the jobs only its business has.

One agent decision deserves special mention because it is the one humans almost never make: suppression. Should this customer be left alone entirely, because she is already likely to transact without any intervention at all? A marketing team measured on sends will never choose that. A system measured on incremental profit will choose it constantly, and the saving is real.

Judgement enters at the top. Everything below it is execution — and everything comes back as memory.

None of this works without memory. The context layer has to remember what was sent, what was rejected, what a human overruled and what happened next — the decision traces. Those traces are what make an agent’s choice explainable after the fact: why this offer, to this person, on this channel, and what the outcome was. And explainability is not a compliance nicety. It is the precondition for autonomy. A system whose reasoning cannot be inspected does not get to be trusted with more, and should not be.

Which is why autonomy should be earned in rungs rather than granted in one decision: advisor first, then co-pilot, then semi-autonomous within tight guardrails, then autonomous only where the outcomes have been measured long enough to justify it. Start with a cohort. Keep a holdout. Widen what the agents may decide as the evidence accumulates, and not before.

There is a corollary for vendors, and it is uncomfortable for most of them. If a platform is confident enough to make these decisions on a brand’s behalf, it should be confident enough to be measured on the result. The honest version of the agentic pitch is not another agent to buy. It is a willingness to move from vendor to owner — to take on the KPI, within agreed parameters, and to link part of the compensation to the outcome. Everything else is a demonstration.

The last era of martech gave marketers more software. The next era takes work away from them.

Nothing here is hard. There is just too much of it

 The answer is not talent, and it is not effort, and it is not budget. It is arithmetic — and arithmetic is the one thing that has changed.

Take a play. Any of the six. Executed properly, it is one decision, for one person, on one channel, at one moment. Now multiply that out across a real business.

Nine cells on the grid, ten million customers, six plays, five owned channels, three hundred and sixty-five days.

Billions of individual decisions a year is what the grid demands if every customer is to get the right play on the right channel at the right moment. A very good marketing team, working hard, makes a few hundred campaign decisions a year — each one briefed, argued over, approved, shipped, and made brilliantly by people who are excellent at their jobs. The distance between those two numbers is not a performance gap. It is a category error.

This also explains something that has puzzled the industry for a decade: why retention has consistently underperformed its own business case. The answer is an asymmetry in how the two halves of marketing got automated.

The same marketer, the same budget, two completely different amounts of work per decision.

Hand a budget to Google or Meta and the system does the work. Targeting, bidding, creative rotation, frequency management — all automatic, running overnight, at any scale, without a brief. The effort per decision is close to zero. Now look at the other side. Every segment needs analysis, then content, then a campaign build. Every send needs a brief, an approval and a report. The effort per decision is a person for most of a day. So ten segments get built once and left in place for months, because eleven would cost another person.

Put those side by side and the last decade explains itself. Martech got hard, so it did not get done, and the money went back to the channel that made it easy. Not because anyone believed adtech was better. Because adtech was the only one that could absorb the volume.

There is a second consequence, and it is the one worth sitting with. The campaign itself was never a strategy. It was a workaround.

Nobody ever chose to talk to a segment. We invented the segment because talking to a person, one at a time, at scale, was impossible to afford. One team, millions of people — so we grouped people who are not alike, sent the average to all of them, and called the compromise a strategy. A segment is a compression algorithm. It was the right answer to a real limit.

The limit has lifted. Per-person is now affordable. Per-moment is now possible. Per-channel is now automatic. The constraint that created the campaign is gone, and the campaign is the only thing still standing on top of it.

What replaces it is not a better campaign calendar. It is a different operating model, and the differences are worth setting out line by line, because every row is a place where a habit has to be given up.

Read the last row twice. It is the one that changes what gets reported upward. Activity metrics tell you the engine was busy. Movement between states tells you whether the base got healthier — which is the only thing the P&L eventually responds to.

Every objection to owned-channel retention for the last ten years — it is too manual, it does not scale, we cannot resource it — was correct. That is why it deserves a straight answer rather than an argument. The constraint was real. It is the constraint that has changed, not the conviction.

The gap was never talent. It is arithmetic — and arithmetic is the one thing that has changed.

Five plays your team already runs, and a sixth it was never built for

 Read as moves on the table, the work resolves into a small set of plays. Each one is either a climb up a row or a hold against the drift rightward.

• Capture turns anonymous and intermediated buyers into known, reachable customers. It is the on-ramp onto the grid, and everything else depends on it.
• First moves a known customer from nought transactions to one.
• Second moves them from one to two, across the tripling inflection. This is the highest-value single move in consumer marketing.
• Repeat keeps the best customers buying, and keeps them in the strong column while they do it.
• Protect pulls drifters back left before their attention is lost — the play that acts on the lead indicator rather than the lagging one.

Those five are the Grow and Protect work your CRM team already does. Agents and in-channel interactivity make every one of them sharper and cheaper to run, but none of them is new, and none of them would be worth an essay on its own.

They translate cleanly outside commerce, which is worth stating because the vocabulary can sound retail-specific. In banking and insurance: Capture is KYC-complete activation, First is the first product, Second is the second product, Protect is a dormancy early-warning. In telco: First is the first recharge, Second is the second, Protect is the pre-churn signal. Same grid, same plays, different nouns.

The sixth play is different in kind. Recover takes the lost column — the customers whose attention has gone dark entirely. Nobody owns it, and that is not an accident. A CRM team is built and measured to keep the engaged engaged. The lost are always next quarter’s problem, so they are never worked. They fall through the floor, and the floor they land on is adtech.

Now the concession that sets up everything after it. None of these six plays is clever. Every marketer reading this already knows all of them, has known them for years, and could have written the list unaided. Which raises the only interesting question in the essay: if the plays are obvious and the money is enormous, why is none of this being done?

Five plays climb the rows and hold the columns. The sixth wins back the customers everyone else has given up on.

Attention is the lead indicator — and the grid that shows it

Every transaction is downstream of attention. Nobody buys from a brand they have stopped noticing. That makes attention the lead indicator and revenue the lagging one, and the gap between them is exactly where brands lose customers without realising it. By the time a sales report dips, the attention left months earlier. The dashboard simply had no column for it.

Attention is not abstract, either. It is a set of things you already log: a WhatsApp message opened, an email clicked, a push notification acted on, a site visited, an app opened, a store walked into. Every one of those is a signal you own, sitting in a table somewhere, unused as a forward-looking measure. Track it and you can act before the loss instead of after it — re-engage a fading customer while there is still a relationship to save, rather than paying to reacquire them once they are gone.

Which leads to the map. RFM ranks customers by what they have already done. The Transaction–Attention Table adds the axis that predicts what they will do next. The rows are lifetime transactions since the first order; the columns are attention right now, split at thirty and ninety days. Ninety days of silence means the relationship is broken, whatever the purchase history says — and that holds even in a long-cadence category like insurance or furniture, because the cadence governs when they buy, not whether they are still listening.

The nine cells. Percentages are illustrative of a typical consumer base, not a benchmark — your own audit finds your real distribution.

A transaction here is whatever you sell: an order, a recharge, a deposit, a policy renewal, a subscription month. The grid is category-agnostic by construction, which is why it travels from D2C to banking to telco without translation.

The three columns hand you three jobs, and this is the part worth committing to memory. Strong attention is to be grown. Weakening attention is to be protected before it slips. Lost attention has to be recovered. Grow, Protect, Recover — and most of any brand’s base sits in the weaker rows and the righter columns, drifting in a direction RFM is blind to.

Look at what the illustrative distribution does to your intuitions. Roughly fourteen per cent of buyers sit in B — best and engaged — and they carry something like thirty-eight per cent of trailing revenue. Meanwhile thirty-nine per cent of buyers sit in R2: bought once, lapsed, gone quiet. They carry fourteen per cent of revenue and almost all of the unrealised value. Nearly two in five of your customers are in a cell nobody in your organisation has a plan for.

The rows hide the single most important number in consumer marketing. For most brands, sixty to sixty-five per cent of customers buy once and never return. The second transaction roughly triples lifetime value, and each purchase after it makes the next more likely. One purchase is a trial; the second is a customer. In banking it is the second product; in telco, the second recharge. Almost everything worth doing is in service of getting from one to two — and then never letting two go dark.

Which reframes the customer journey. It is not one acquisition funnel. It is a sequence — Unknown to Known to First to Second to Repeat — running against a second force that needs no budget and no campaign to operate.

The path you drive, and the path that happens anyway.

Attention decay is the default state of every commercial relationship. Nobody has to do anything for it to happen; it happens while you are busy running the calendar. Marketing’s job, stated as plainly as it can be, is to move customers along the top path faster than entropy pulls them along the bottom one. The TAT is simply a way of seeing both movements at the same time, which no list and no RFM score can do.

One purchase is a trial. The second is a customer — and it roughly triples their value.