The reconciliation risk in agentic advertising is not that agents will disagree. It is that they will agree, completely, and still record different things. Two agents can close a deal in full agreement, each confident it understood the terms, and produce two different structured records of what those terms were. That is not a glitch or a miscommunication. It is the predictable consequence of asking two probabilistic language models to parse the same natural language into structured deal data, independently, with no shared reference.
TL;DR. Semantic drift is what happens when the same natural-language deal terms are interpreted into different structured representations by different language models. A "premium CPM" or a "frequency cap of three" means something specific to each model's training, and those meanings may not match. Shailley Singh of IAB Tech Lab named the mechanism in August 2026 as "differences between LLMs degrading interpretation of the transactional context." Unlike fraud, semantic drift is not detectable by the party experiencing it: from inside each system, the record is internally consistent and passes standard checks. The solution both major standards bodies have converged on is deterministic transaction objects rather than natural-language confirmation: canonical structures that do not require interpretation.
What is semantic drift in the context of a media deal?
Semantic drift is the divergence that occurs when two language models each parse the same natural-language input into structured data and arrive at different results. The input was the same. The models were operating in good faith. The outputs are different.
In a media deal, the terms being parsed include CPM floors, volume guarantees, frequency caps, brand-safety conditions, delivery windows, and pricing tiers. These look like unambiguous terms to a human reader. To a language model, each is a phrase with a probability distribution over possible structured meanings, and the most probable meaning varies between models based on training data, fine-tuning, and context.
A buyer's agent trained on buy-side documentation may parse "premium inventory" one way. A seller's agent trained on sell-side inventory classifications may parse it differently. Both record what they believe was agreed. The records differ. Neither agent knows this, because neither agent has access to the other's record at the point of agreement.
How does this differ from fraud or a misunderstanding?
Fraud means intent to deceive. Semantic drift has no intent. Both agents are acting in good faith, and both records are, from the perspective of each system, entirely correct. That is what makes it harder to catch than fraud.
A fraudulent record looks wrong to someone who knows what the deal was. A semantically drifted record looks internally consistent. Every individual field in each record passes validation. The payments reconcile within each system. Standard post-campaign checks, which verify delivery against a target rather than one party's record against the other's, will not surface the divergence.
The divergence only becomes visible when both sides produce their version of the deal and compare them directly. In practice, that tends to happen at invoice time, after the campaign has run, when the consequences are already locked in.
What is the scale of the problem?
Published simulation research modelled this across 90,202 agent-to-agent transactions. Under separate record-keeping, where each agent maintained its own account, the two records disagreed on at least one deal term in 95.3% of cases. Under a single shared record that both agents wrote to at the point of agreement, divergence fell to 0.19%. The methodology and codebase are available from WPP Research.
The research models the current specification architecture rather than any live production system. But the mechanism it isolates is structural: it is a property of two parties keeping separate records of the same event, not of any particular platform's implementation. Any system where each agent writes its own version of the deal is vulnerable to this regardless of how well-engineered it is.
What did the standards bodies say about it?
Shailley Singh, head of standards at IAB Tech Lab, named the failure mode explicitly in writing, describing the need for architecture that prevents "differences between LLMs degrading interpretation of the transactional context," as reported in an ADOTAT investigation published in August 2026. His proposed solution was equally specific: "structured transaction objects, explicit transaction states, and systems of record that minimise hallucination or misinterpretation of context."
The phrase "structured transaction objects" is the key term. A structured transaction object is not a natural-language description of a deal. It is a machine-readable schema with defined fields and defined valid values, where "frequency cap" is an integer with a unit, not a phrase. When both agents write to the same schema at the point of agreement, there is no subsequent parsing step for the records to diverge in.
Brian O'Kelley of the Ad Context Protocol, answering independently for the same investigation, said a shared machine-readable semantic layer is "yes" a precondition for cross-company agent buying to function. Both standards bodies converge on the same architectural answer: move the definition of deal terms upstream, before natural language touches the record, not downstream, where two models are asked to reconstruct structure from prose after the fact.
What does this mean for the current generation of agentic platforms?
Most current agentic advertising demonstrations include agents that communicate in natural language and produce natural-language confirmations of deals. Singh's position, stated on the record, is that this architecture "cannot simply rely on a model saying that a transaction occurred." A natural-language confirmation is not a transaction record. It is a claim about one.
The current state of the standards, as O'Kelley confirmed in the same ADOTAT investigation, is that the infrastructure to prevent semantic drift is "not yet active in production at scale." The specifications exist. The reference implementations are being built. The live volume where these systems will face real reconciliation pressure is, by the only operator willing to name a number, running at roughly $3,000 a day in gross media spend, as Andrew Mole of pubX stated in the same investigation.
The window between "the specifications exist" and "the specifications are running at scale under real reconciliation pressure" is the window in which the semantic drift problem is cheapest to solve. Building shared-record infrastructure while transaction volumes are low is straightforwardly easier than retrofitting it after the infrastructure for separate records has calcified across an industry.
What should buyers require now?
Ask, specifically, whether the agents on both sides write their account of a deal to a shared schema at the point of agreement, or whether each produces its own record from natural-language confirmation.
A platform that can show a structured transaction object, with named fields, defined value types, and a single authoritative instance that both agents reference, has solved the semantic drift problem by construction. A platform that produces natural-language deal summaries and uses those as the primary record is carrying the risk, whether it knows it yet or not. The question is not whether the agents communicate clearly. It is whether the record of what they agreed is deterministic.
________________
This article references statements by Shailley Singh of IAB Tech Lab and Brian O'Kelley of AdCP published in ADOTAT in August 2026, published simulation research into agentic deal reconciliation conducted by Alkimi Exchange and available from WPP Research, and statements by Andrew Mole of pubX published in the same ADOTAT investigation. The simulation models the current specification architecture and is not an assessment of any specific production platform.