About The Investigation. Every Wednesday, ADOTAT publishes one long piece built from primary sources rather than press releases: written on-record answers, contracts, filings, transcripts, and the questions the trades tend not to ask. The format is deliberately narrow. One subject, named companies, named executives, and every claim put to the people it concerns before publication, with their answer printed as they gave it or their silence noted. Nothing here is sponsored, nobody reviews it in advance, and the reporting runs whether or not the subject is a friend of the publication.

Both protocols agree a model's word isn't a receipt. One dates the fix to 2027. The other counts in releases. And the one operator willing to give me a number said two to three thousand dollars a day.

I sent Brian O'Kelley six yes-or-no questions about whether agent-to-agent ad buying is actually built or just sold as built. He answered all six in writing, on the record.

Question three was the one that mattered: does the binding, signed record of an agentic transaction exist and work in production today?

His answer, complete, unedited: "Exist yes. Work yes. Active in production at scale - not yet."

That is the man who built the Ad Context Protocol, describing the state of the thing every agentic keynote this year has described in the present tense.

Separately, and through his press team, Shailley Singh of IAB Tech Lab answered two questions about the competing Agentic Advertising Marketplace Protocol.

He was blunter than anyone expected. "An agentic transaction cannot simply rely on a model saying that a transaction occurred," he said. "The architecture needs a deterministic, auditable record of what was proposed, approved, and executed."

Two rival standards bodies.

One shared tense.

O'Kelley said not at scale. Andrew Mole told me what the scale is.

Mole is CEO of pubX, a founding member of the Agentic Advertising Organization, and one of the few people in this category running live agent-bought, agent-sold volume rather than describing it. He answered seven questions in writing on Wednesday morning, from the UK, on the record.

The number is two to three thousand dollars a day.

That is gross media spend, not revenue. pubX books revenue only as the percentage it keeps. It covers a handful of publishers, premium and scaled, running across several sites each. And it is not continuous. There was a break in activity, one campaign having completed, with the level returning again this week.

So: the daily gross of the most forthcoming operator in agentic advertising is roughly what a mid-market brand spends on search before lunch.

Mole is not embarrassed by this, and he should not be. He is the only person who answered the question with an integer. Everyone else in this category answers it with a roadmap.

He is also, on the record, furious about the roadmaps. Asked whether he stood behind his own account of companies putting fifty dollars through a test and issuing a press release before the campaign has run, then reissuing the same release six months later, he wrote: "yes I do stand by that and I think it's harmful."

And then he said the thing nobody selling this will say in public. There is no agreed-upon definition of what agentic even means. In his telling, a number of companies are laying an agentic layer over the infrastructure they already own, which is a different animal from agent-to-agent buying, and which by his account delivers no economic benefit to publishers or advertisers because it is literally the same programmatic system.

Sit with the position he is in. He is a founding member of the industry group, he runs live volume, his own engineering lead helped build the protocol, and his assessment of most of the category's announcements is that they are the old pipes with a new word painted on them.

His timeline is the one to hold against O'Kelley's and Singh's, because it is the only one built on a commercial mechanism rather than a release schedule. Q1 2027 matters, he says, because annual spend commitments locked into existing programmatic infrastructure elapse, and that creates the liquidity headroom for something else to move into. He calls the transition epochal, puts a majority of ad spend flowing agentically inside two years, and volunteers that he is an optimist.

He is describing a trillion dollars changing channels. He is currently moving three thousand a day.

Both of those are true at once, and the gap between them is the entire story.

Disclosure, because it matters to how you weigh him: pubX has served as the sales agent in Scope3 campaigns and is a fee-paying founding member of the Agentic Advertising Organization. Mole says there is no equity relationship. pubX also sells the thesis that the existing stack can be skipped, which is a thesis with a price attached.

The first question to both principals was whether a shared, machine-readable semantic layer is a precondition for cross-company agent buying, or merely nice to have.

O'Kelley's entire answer was one word: "Yes."

Singh's was longer and more revealing. "Yes. A core principle of AAMP is continuity," he said, describing a protocol built on standards the ecosystem already understands rather than a clean-sheet spec. The goal, in his words, is to make sure buyer and seller agents work from "consistent, machine-readable definitions, rather than allowing differences between LLMs to degrade interpretation of the transactional context."

Read that clause twice. The failure mode he is designing against is two AIs understanding the same words differently. Not fraud, not latency, not fill rate.

Semantic drift between models, on either side of a transaction, involving your money.

The second question was the one that decides whether any of this is advertising or theater. Does a real agentic transaction need a deterministic, auditable record, rather than a model asserting a buy happened?

O'Kelley: "Yes."

Singh went further, and this is the passage that should end the demo season. The architecture, he said, requires "structured transaction objects, explicit transaction states, and systems of record that minimize hallucination or misinterpretation of context and that both sides can reference and audit." Existing standards like OpenDirect, he added, give a foundation for "defining canonical transaction states rather than relying on natural-language confirmation."

Natural-language confirmation.

That is the polite technical term for an agent telling you in English that it bought something.

So here is the finding, from the two organizations most invested in beating each other: if an AI agent tells you it bought your media, that is not a receipt. That is a claim.

Every demo you have watched this year where an agent narrates a media buy in conversational prose is, by the explicit standard of the people writing the standards, not a transaction.

It is a very expensive chatbot describing one.

Both men, independently, in separate emails, named the same risk: hallucination. Not as a metaphor.

As the thing the plumbing has to be engineered to survive.