The Mandate Finished Last
Season 2 opened by claiming that once every advertiser has a capable buying agent, advantage moves to the quality of the mandate — and then 42 operators put that answer last, at 4%, while proprietary data took 78%, the biggest share these polls have produced since the very first one. The room is right about the market it works in, and the mechanism is the Legibility Lag: an asset commands a price when a buyer can see the difference between one holder's version and another's and can hold it, and causal weight is not one of those conditions. Run all four options through it and the ballot order falls out. Capable agents on both sides of a trade collapse execution variance toward zero, which makes a data advantage sharper today, not flatter. The mandate becomes purchasable only once it is a versioned artifact with results attached to a version — the same paperwork accountability already needs.
In English, please
A recurring reader poll asked what will give one advertiser an advantage over another once every advertiser has an equally capable AI program ("agent") doing the actual ad buying. Four options: owning customer data nobody else has, running faster, having special access to inventory or platforms others can't get, or simply giving the agent clearer instructions about what the business wants. The author had launched the season arguing the last one — clearer instructions — would become the deciding factor. Of 42 voters, 78% picked data. Clearer instructions came dead last, at 4%.
The essay's explanation is that markets don't pay for whatever matters most. They pay for whatever they can see and own. Data passes both tests: you know exactly what you hold, a competitor can verify it in a deal, and you can buy it, license it, and stop rivals from getting it. A clear set of instructions passes neither. Nobody publishes their brief, so there's no way to compare yours against a competitor's, and you can't buy a good one — someone inside the business has to write it. Something can be the real cause of a result and still be worth nothing on paper, because there's no way to trade it.
There's a second reason the voters are right for now. Giving both sides an equally capable agent doesn't make data matter less; it makes data matter more. Today a lot of the gap between two advertisers comes from one team simply being sharper than the other, which can hide a data disadvantage. Take skill out of the equation by handing both sides the same quality of machine, and the only thing still separating them is what each one feeds it.
What would have to change for the original argument to come true: instructions to an agent would have to become a written, versioned document with a record of what the agent did under each version — the same paperwork usually proposed for figuring out who's to blame when an agent goes wrong. Once results can be traced back to a specific version of the instructions, you can compare one against another, and only then is there something a buyer can price. Separately, the data advantage has to shrink as more of the same data becomes purchasable by everyone. Neither has happened, so the author voted with the room, against his own opening argument.
On this page
The cold open
Here’s the question as it ran: when every advertiser has a capable buying agent, what creates the edge? Options: Proprietary data. Faster execution. Privileged access. Clearer mandate. Core thesis at launch: intelligence becomes infrastructure; advantage moves to the quality of the mandate.
That thesis is mine. I wrote it as the opening claim of a season about where advantage goes once agents can do the work, put four options underneath it, one of which was the thesis itself, and asked a room of operators to vote. They put my answer last, by a distance.
The vote
The poll closed with 42 votes:
| Answer | Share |
|---|---|
| Proprietary data | 78% |
| Faster execution | 9% |
| Privileged access | 7% |
| Clearer mandate | 4% |
The shape of it
This isn’t a four-way ranking. It’s one answer and a remainder, and the remainder is too small to rank. Nine percent and seven percent of 42 ballots are four votes and three. Four percent is one or two people. Reading “faster execution beat privileged access” out of that is reading a single ballot as a finding. Second, third and fourth place here are one undifferentiated 22% that means something other than the winner.
So the finding is the 78%. Only one poll in this series has ever produced a bigger number: the first one, back in June, where 87% of 72 votes said standards matter more than ever in a world where AI can understand anything. That question had its answer built in. The dissents on offer were “fewer than today” and “AI replaces them,” and neither is a live position among people who build ad infrastructure for a living. This one had no such tilt. All four options are assets that real companies are spending real money to acquire right now, each with a budget line and an org chart behind it. Seventy-eight percent on a question where every option is somebody’s current strategy is the strongest signal these polls have produced.
It also sits oddly against the room’s own history. Season 1’s fourth week asked what becomes the most valuable asset in an agentic market, and that room put Trust at 52% and Data at 33% — the result What Agents Can’t Manufacture was built on, and that essay pushed further, arguing that a large private data lake looks like an unauditable black box to a counterparty’s agent. Ten weeks later a data-shaped answer takes 78%. The two results don’t fight, because they describe different jobs for the same asset. Data you present to a counterparty has to survive their verification, and scale makes that harder. Data you consume privately has to survive nothing at all. A retailer’s purchase log makes its own agent bid better whether or not anyone else can audit it. And trust, whatever else it is, functions as a license: it decides whether you transact, and it does nothing to separate two companies that both hold one. Asked what separates two advertisers who both have capable agents and both clear the trust bar, the room stopped naming the entry requirement and named the differentiator.
The Legibility Lag
Here’s the mechanism, and it accounts for the shape of the whole distribution rather than just the winner.
An asset commands a price when a buyer can do two things with it: see the difference between one holder’s version and another’s, and hold it — acquire it, contract for it, keep a competitor from having it. Causal weight is not on that list. An asset can be the thing that decides the outcome for years before it satisfies either condition, and during that interval the market pays for whatever visible asset sits closest to it.
Call the interval the Legibility Lag: the gap between when an asset starts deciding results and when the market can see it well enough to buy it. Inside the lag an asset trades near zero no matter how much of the outcome it explains, because there is no transaction available in it. Run the four options through the two conditions and the ballot order falls out.
Proprietary data clears both. You know what you hold and so does everyone else, because exclusivity is a contract fact — disclosable in diligence, enforceable in court. It is acquirable, licensable, and it sits on a balance sheet. It is the most legible asset this industry has, which is why twenty years of adtech M&A organized itself around buying it.
Privileged access clears both, on a timer. A first-look arrangement or a beta seat is a contract, so it is holdable, and it is visible because the counterparties announce it. What it fails is duration, and six months of that decay is already on the record. In the window The Intelligence Economy tracked, the entry price for advertising inside ChatGPT went from a $200–250K minimum in January to $50K in May to no minimum at all. Early access to the most-discussed new surface in the market was worth something for about two quarters.
Faster execution clears visibility and fails holding. Latency is measurable to the millisecond, so the difference is easy to see. It is also rented from the same few clouds and model providers your competitor rents from, and available to them next quarter on a purchase order.
A clearer mandate clears neither. Nobody publishes their brief. There is no diligence process that reads a competitor’s objective function and reports back on whether it is any good — no benchmark, no rate card, no comparable transaction. You also can’t buy one. A mandate has to be produced, by people who understand a business well enough to state what it wants in terms a machine can act on. An asset with no observable variance and no market to acquire it in is, from a pricing standpoint, not yet an asset.
Four options: one clears both conditions, one clears both on a timer, one clears half, one clears neither. The room voted them in that exact order. It was ranking them by which one is priceable, and on that question it was right.
The same thing already happened one layer down
This isn’t a new observation, only a new name for it at the level of strategy. Three weeks before this poll, Business Outcomes Isn’t a Number found the same shape one layer down, at the level of what an agent optimizes for. Cost and ROAS dominate as optimization targets for a reason that has nothing to do with being the right goal: they come pre-instrumented, so the system can already see them. A customer-defined outcome is an instrumentation project before it is an objective function, and it loses to a worse target that happens to already be measurable.
Swap “optimization target” for “competitive asset” and the argument is the same one. The market selects the measurable option, then rationalizes the selection as a judgment about value. The 78% is that selection running at the strategy layer, and it will keep running there until the alternative gets instrumented.
Capable agents raise the value of data
The launch thesis had a hidden step in it, and the step is wrong. It assumed that commoditizing intelligence would flatten the outcomes intelligence produces. The first-order effect runs the other way, for arithmetic reasons.
Variance between two advertisers’ results comes from the quality of their inputs and the quality of their execution against those inputs. Execution variance is currently enormous. A sharp analyst with a mediocre dataset routinely beats a mediocre analyst with a good one, and a fair amount of what gets reported as data advantage today is really the second effect wearing the first one’s clothes.
Now give both sides a capable agent. The execution term collapses toward zero: comparable reasoning, comparable speed, comparable stamina on both sides of the trade. What is left is the inputs. Commoditizing the processing layer doesn’t dilute a data advantage; it strips away the noise that was hiding it.
The arithmetic, drawn: commoditizing the processing layer shrinks the total spread and hands what remains of it to whoever holds the better inputs.
That sits fine with the Self-Supply Test, which held that data fails as a durable moat because a smarter agent can synthesize or simply buy more of it. That is a claim about what survives a long horizon. It says nothing about what decides a quarter. An asset can be the sharpest differentiator on the table today and the one that erodes fastest, and proprietary data is currently both.
What would have to change
Two conditions, and each is checkable rather than a matter of taste.
First, the mandate has to become an artifact. Today it lives in a kickoff deck, a Slack thread, and somebody’s head, which is exactly why there is nothing to compare. It becomes visible the moment it is a versioned document an agent reads and a ledger records it acting under — the machinery Who Owns the Agent’s Decision? argued for on accountability grounds: versioned operating rules plus an append-only record of what was done under which version. That was framed as infrastructure for assigning blame after something goes wrong. It is the same instrumentation that makes mandate quality visible in the first place. Once a result attaches to a mandate version, you can compare versions, compare the people who wrote them, and pay the difference. Only then is there anything for a market to price.
Second, the data advantage has to decay from the supply side. Part of that is ordinary commoditization: as more of the same data becomes purchasable through clean rooms, retail media networks, and licensing, the gap between what two advertisers can see narrows. Part of it is specific to agents, and it is the black-box problem read forward. A counterparty’s agent has no mechanism to verify a large private dataset and every reason to discount it, so the party holding less data with better provenance gets believed faster and cheaper.
Neither condition is close to met. Until both move, “advantage moves to the quality of the mandate” is a forecast about a market that hasn’t arrived, and the room was voting on the one it works in today.
My vote
I voted Proprietary data, against the thesis I opened the season with.
Some of that is a fair reading of the tense. The question was present-tense: what creates the edge today. My thesis describes a market where every advertiser actually has a capable buying agent, which is not the market of August 2026, and voting for the mandate would have meant answering a question that wasn’t on the ballot.
The less comfortable part is simpler. When I sat with the four options, I could name, for any client I have worked with, exactly which data assets they held that a competitor didn’t. I could not name a way to tell whether their brief was better than their competitor’s. I wrote a thesis about the thing I can’t measure and then voted for the thing I can. That is the Legibility Lag operating on the person who named it, which is roughly the level of evidence I would want before believing a mechanism.
Where the season starts from
Season 2 runs eight weeks on where money and power settle once agents can act, and Week 1 sets the baseline. The baseline is that the room starts from the answer that won the last cycle. Every technology transition begins there. The question is what moves it.
This room has moved before. In Season 1’s fourth week, Trust overtook Data 52% to 33% while the ballot was still open, after the question got reframed mid-week. So the 78% is not a fixed position. It is what this room believes when the question is put this way, and it will hold until a different answer becomes purchasable.
That gives a marker for the rest of the season. If mandate quality is ever going to be the answer, the first sign will be this room starting to pay for whatever makes a result checkable, because verification is what turns a brief into something a buyer can compare, and comparison is what turns an asset into a price. The rest of the season either finds that shift or it doesn’t.