Negotiating the Nouns
Season 2's fourth Friday poll closed with the tightest finish of the run: 55 operators, Success criteria 34%, Price 32%, Audience definitions 16%, Liability 16%. The essay names the mechanism — Negotiating the Nouns: automated negotiation is mostly semantic negotiation wearing a price tag. Machines have negotiated price for fifteen years (real-time bidding), and price stayed easy only because every other term was frozen by convention: the audience a list name, success handled at the QBR, liability boilerplate. Agents can't execute ambiguity, so the deferred arguments get pulled forward from the quarterly review into the deal schema itself. The author voted Success criteria — the clause the other nouns drain into — and defends the 32% price bloc: every unsettled noun eventually invoices as a price dispute. Checkable close: watch for a machine-readable success-criteria field (metric, source, window, baseline) an agent can reject on parse.
In English, please
A recurring reader poll asked: when AI programs ("agents") handle ad deals for both the buyer and the seller, which deal term will they fight over hardest? Four options: the price, the definition of the audience being bought, the criteria for calling the campaign a success, and who is liable when things go wrong. Of 55 voters, success criteria won with 34%, price came second at 32%, and the other two took 16% each — the closest finish this season.
The essay's starting observation: price is the strange one on that list, because machines have negotiated ad prices automatically for fifteen years — that's what real-time ad auctions are. Price could stay simple because everything else was held still by habit: the audience was a name on a form, success was argued about in quarterly reviews after the money had moved, liability was legal boilerplate nobody read. Human deals ran on useful vagueness — leaving terms fuzzy let deals close, and arguments were had later, between people, when they actually mattered.
Software can't work that way. A program can't buy against a vibe or sign a clause it can't read precisely; every fuzzy term has to be pinned down before any money moves. So automation doesn't remove negotiation — it moves it, out of the price and into the definitions. That's the week's term, negotiating the nouns: the fight shifts from what the deal costs to what the words in the deal mean — what counts as success, measured by whom, over what period, compared to what.
The author voted for success criteria, which won by two points. His reasoning: it's the definition the others depend on — you can't argue about who pays for a failure until you've agreed what failure means. He also defends the third of the room that voted price: every unsettled definition eventually turns into a price argument (the discount, the refund, the make-good), so price is where unresolved fights get paid off. His test for whether this is really happening: watch the technical specifications for agent-to-agent deals. The day they include a standard, machine-readable field naming the success metric, who measures it, and over what window — one an agent can reject automatically if it doesn't parse — the definitions have started being settled once, industry-wide, instead of argued deal by deal.
On this page
The cold open
Here’s the question as it ran: what will buyer and seller agents negotiate hardest? Options: Price. Audience definitions. Success criteria. Liability. Core thesis at launch: automated negotiation expands beyond price into definitions, proof, and responsibility.
Notice what’s odd about the list. One of those four terms has been negotiated by machines for fifteen years. Real-time bidding is nothing but automated price negotiation — billions of times a day, at the millisecond, with no human in the loop. If price were the hard part, the industry solved the hard part in 2010 and nobody noticed. So the poll was secretly asking a different question: which term still needs people in the room? And two-thirds of the room answered with a definition.
Last week’s poll asked what autonomous agents will ultimately buy, and 57% said outcomes. That essay ended by handing this week its assignment: a room that just voted to buy outcomes was about to reveal which clause of the outcome contract it thinks is the fight. It answered. The fight is over what counts as success.
The vote
The poll closed with 55 votes. (Shares are LinkedIn’s rounding, which is why they sum to 98.)
| Answer | Share |
|---|---|
| Success criteria | 34% |
| Price | 32% |
| Audience definitions | 16% |
| Liability | 16% |
Two points. After a season of clear verdicts — last week produced a 57% majority on 73 ballots — this one nearly split down the middle, and the split is the story. Read one way, Success criteria won and my ballot was on it. Read the more useful way, the room divided into a 32% bloc that says the money line is still the fight and a 66% bloc that says the fight has moved into the words around the money line: what counts as success, who counts as the audience, who owns the miss. I’ll defend the minority position later in this essay, because it’s stronger than the two-point loss suggests. But first, the majority deserves its name.
Negotiating the Nouns
Negotiating the Nouns: when agents transact directly, negotiation doesn’t shrink — it migrates. Price, the verb of the deal, is already cleared by machines. What remains contested is every noun the price hangs off: the audience (whose taxonomy decides who counts as in-market), success (what counts, who counts it, over what window, against what baseline), the miss (who is liable when a definition turns out to be wrong). Automated negotiation is mostly semantic negotiation wearing a price tag.
The reason this is new work is that human deals ran on ambiguity, and the ambiguity was functional. “Success” stayed vague so the deal could close by quarter-end; it got argued about later, at the QBR, relationally, after the money had moved. The audience was a list name on an insertion order — a handshake, not a definition. Liability lived in an indemnity section that was copied between contracts and read by nobody. None of this was sloppiness. It was a working system for deferring expensive arguments until they were actually needed, lubricated by the fact that both sides employed humans who could absorb a vague term and interpret it charitably.
An agent can’t do that. It cannot buy against a vibe, judge against a feeling, or sign a clause it can’t parse. Every deferred argument has to be settled before a dollar moves, in a schema, explicitly. The arguments don’t disappear — they get pulled forward from the QBR into the deal itself. That’s the mechanism under this poll: agents don’t remove negotiation from advertising, they relocate it, from the price line everyone watches into the definitions nobody used to read.
Run the four
Price. The strongest minority vote of the season, and I want to state its case properly: every noun you fail to settle eventually presents its invoice as a price dispute. The make-good, the discount on renewal, the clawback — those are semantic disagreements being paid off in the only unit both sides share. On that view, price never stops being the fight; it’s where the fight settles. What the 32% miss, I think, is the direction of causation. Price stayed the whole negotiation only because every other term was held still by convention. The auction cleared the number in a millisecond because the definitions underneath it were frozen — same impression standard, same viewability rules, same measurement, industry-wide. Agents un-freeze them.
Audience definitions. Sixteen percent for the oldest unsettled noun in the business. Two agents can’t transact on “auto-intenders” until both parse the same answer to who’s in it, whose data decided, and how stale it’s allowed to be. The definition was priced as a data fee for years precisely so nobody had to warrant it — last week’s essay called that the market quietly repricing a promise into an ingredient.
Success criteria. The winner, by two points, and clauses one and two of last week’s outcome contract wearing a job title: a definition both sides accept, a measurement both sides trust. iROAS is a negotiation, not a number made the case for a single metric; this vote generalizes it. Metric, source, attribution window, baseline, holdout design — each is a term an agent has to fix before it can act, and each is currently fixed bilaterally, deal by deal, if at all. There is no protocol field two arbitrary agents can point at that answers whose measurement settles this deal. That absence is the fight.
Liability. Sixteen percent for the clause that the Season 1 finale argued is the only one whose absence stops a deal outright, and that the underwriting essays priced: who pays when the outcome doesn’t arrive. I’d have expected more than 16%, and my read on the low number is sequencing rather than disagreement — you can’t argue about who owns the miss until you’ve agreed what a miss is. Liability is downstream of success criteria. The room ordered the fights correctly.
My vote
I voted Success criteria, and it held by two points.
The practical reason: it’s the clause the other two nouns drain into. An audience definition is a success criterion for targeting. A liability clause is what happens when the success clause fails. Settle what counts and who counts it, and the remaining nouns become tractable; leave it open, and every other term is negotiating on sand. It’s also where the work actually is right now — the agentic protocols can already carry price, formats, and flight dates in machine-readable form, and are thinnest exactly where this vote points: the field that names the metric, the measurement source, and the window in a way a counterparty’s agent can accept or reject. The nouns will get negotiated one way or another. The only question is whether they’re negotiated once, in a schema working group, or forever, deal by deal.
And the near-tie deserves respect. A 34–32 finish means this room — practitioners, mostly — looked at machine-to-machine dealmaking and split almost evenly on whether the hard part is the economics or the semantics. Both camps agree on the underlying fact that the deal terms are moving into code. They disagree about which line item bleeds first.
What the next weeks test
Week 5 is already live and asks which intermediary function remains most valuable when agents transact directly — finding counterparties, negotiating price, verifying results, or assuming risk. Two of those options are this week’s winners wearing job titles: verifying results is success criteria as a service, and assuming risk is liability as a business model. A room that just said the fight is over what counts is now being asked who gets paid to do the counting.
The reading in this essay is checkable the ordinary way. Watch the deal schemas, not the demos. The day a production agent-to-agent deal carries a machine-readable success-criteria field — named metric, named measurement source, named window, named baseline — that a buyer’s agent can reject on parse, the nouns have started being negotiated once instead of every time. If, a few years out, those fields are still free-text or bilateral, then the 32% were right after all: the definitions stayed frozen, and the fight went back to being about the number. Either way it shows up in public specs on a normal release schedule. The poll asked what agents will negotiate hardest. The schemas will answer in writing.