Yes. AI buying agents run on language models, and language models can produce plausible deal terms that nobody agreed, from a CPM filled in when the real price was unknown to a frequency cap restated wrongly in a confirmation. The reliable defence is procedural: write every term once into a structured record both parties hold, and check the agent's account of the deal against that record before any money moves.

TL;DR. In media buying, hallucination is less about agents inventing facts about the world and more about agents drifting from the brief, the seller's offer or their own earlier statements. IAB Tech Lab's AAMP 2.3 release added pricing provenance so agents cannot fill in CPM values when pricing data is missing, which shows the risk is real and already being engineered against. Six checks catch most invented terms: a shared record, price provenance, a field-by-field comparison, a portfolio check, repeat runs and a human approval point for anything the record cannot confirm.

Do AI agents hallucinate in media buying?

AI agents can hallucinate in media buying because the models underneath them sometimes guess. In the 2025 paper Why Language Models Hallucinate, Kalai and colleagues describe models "producing plausible yet incorrect statements instead of admitting uncertainty", and argue that "training and evaluation procedures reward guessing over acknowledging uncertainty". A buying agent faces the same pull: a confident number looks like progress, while "I do not know" looks like a stall.

The survey by Huang and colleagues splits hallucination into factuality hallucination, a conflict with verifiable real-world facts, and faithfulness hallucination, a divergence from the user's input or a lack of self-consistency. Faithfulness is the type that matters in a deal, and its three subtypes map onto buying: instruction inconsistency is a deal that departs from the brief, context inconsistency is a confirmation that departs from the seller's offer, and logical inconsistency is a set of contracts that does not add up to the budget the agent says it kept.

Where do invented terms get into a deal?

Invented terms usually enter at four points: gaps in data, rewording between systems, drift across repeated negotiations, and summaries written for people. Gaps in data are the best documented. IAB Tech Lab's AAMP 2.3 release post describes fixing "the problem of CPM hallucination", where fabricated CPM values could appear when pricing data was missing, which it called "a dangerous failure mode when an agent was negotiating real money".

Rewording is subtler. When each side's model turns the same natural-language terms into its own structured record, two agents acting in good faith can agree completely and still store different things, a problem examined in Semantic Drift Between Models. Repetition adds a third route: in the 2024 τ-bench study, state-of-the-art function-calling agents succeeded on fewer than half of tasks that required following domain rules, and the chance of succeeding on all eight of eight repeated trials fell below 25% in the retail domain.

Summaries are the easiest to miss. A status report saying a deal "includes a 10% make-good" (a hypothetical example) is itself model output, and nothing guarantees the signed terms match unless someone checks.

How do you catch invented terms before money moves?

Each of the six steps below targets one of those entry points, and together they replace the question "did the agent sound confident?" with "does the record agree?"

Step 1: Fix the terms in one structured record both sides hold

The most effective single check is to stop each side keeping its own version of the deal. The DealSheet in the Alkimi Marketplace is a concrete worked example: it is written at the moment of agreement, timestamped, held by both parties and retrievable by either side independently. Any term an agent later reports can be checked against it, and a term that is not in the record was not agreed.

The size of the gap shows up in simulation. In the same company's Agree. Transact. Verify. simulation, 95.3% of deals had mismatched records when each side kept its own, against 0.19% when both worked from a shared record. These are simulated deals, not live campaigns, but the direction is clear: separate books let invented or drifted terms survive unnoticed.

Step 2: Demand a source for every number

Every price, volume and percentage in a deal should carry a note of where it came from. AAMP 2.3 took this route for pricing by adding a field that shows pricing provenance, so organisations "can trace where every price came from through the negotiation chain". The same post puts the goal plainly: "An agent that says 'I don't know the price' instead of confidently making one up is an agent you can trust once it is deployed."

Configure the buying agent so that a missing input produces a stop or a question, never a default. Any number with no traceable origin, whether a CPM, an impression volume or a viewability floor, should be treated as unagreed until the counterparty or a person confirms it.

Step 3: Compare the confirmation with the record, field by field

Read the agent's confirmation as a claim to be tested, not as the deal itself. Check each field (price, volume, flight dates, geography, format, frequency cap, make-good terms and brand safety conditions) against the structured record, and flag any field that appears in one but not the other.

Natural-language summaries are where paraphrase turns into invention, so compare structured fields wherever possible. The draft OpenProposal specification, which IAB Tech Lab has published as the core object of AAMP 3.0, follows the same instinct: "Every line item is fully self-contained. A field never takes its value from the parent proposal or a sibling line item."

Step 4: Check the portfolio against the brief, not just each deal

An agent can sign individually valid contracts that together break the brief. Three ways an AI buyer fails a media brief describes this budget overrun pattern, where each contract is valid in isolation but the combined commitment exceeds the stated budget. Add up total spend, audience reach and required formats across every deal, and compare those totals with the brief rather than with the agent's own running tally.

Step 5: Run the same negotiation more than once before go-live

A single clean test proves little, because a model that guesses can guess right once. The τ-bench authors proposed the pass^k metric for exactly this reason: it is the chance that all k repeated trials succeed, rather than at least one. Rehearse the same brief several times against a counterparty agent, and treat any term that changes between runs without a matching change in the inputs as a hallucination risk.

Step 6: Keep a human approval point for anything the record cannot confirm

Autonomy is safer earned in stages than granted on day one. A five-stage ladder of earned autonomy (Observe, Recommend, Draft, Human-approved action, Bounded automatic action) gives a workable rule: an agent moves up a stage only when its outputs at the current stage consistently match the record. Until then, any term the shared record cannot confirm goes to a person before it is signed.

Are the standards bodies addressing invented terms?

Yes, from complementary directions. IAB Tech Lab's Agentic Advertising Management Protocols (AAMP) treat fabricated prices as a fixable engineering defect, through the 2.3 provenance field, and the OpenProposal draft moves towards self-contained, structured line items that leave less room for interpretation. AgenticAdvertising.org, which publishes the Ad Context Protocol (AdCP), makes the governance case on its homepage: "speed without judgment is just fast mistakes."

Neither standard removes the need for checks on the buyer's side. Structured fields narrow the space in which a model can improvise, while a shared record and a human approval point catch what still gets through.

What should buyers do about hallucination now?

Buyers should treat hallucination as an operational risk with known controls, not a reason to avoid agents or trust them blindly. The six checks work whichever model or platform sits underneath, because they test the record rather than the agent's confidence. The accountability question is simple: can either party show, without asking the other, exactly what was agreed?

Rehearsal is the cheapest place to find invented terms. In the Alkimi Gym, buyers and sellers build their own agents, set the rules those agents work to and spar against counterparty agents before any real money moves, while Concourse is a live benchmarking product for AI media-buying agents. Whatever the tool, the test stays the same: every term an agent reports should trace back to a record both sides hold.

Frequently asked questions

Do AI agents hallucinate in media buying?

Yes, because the language models behind them can produce plausible but unsupported output when information is missing or ambiguous. In deals this usually shows up as terms that depart from the brief, the seller's offer or the agent's own earlier statements.

What is a hallucinated deal term?

A hallucinated deal term is any price, volume, date or condition that appears in an agent's output but was never agreed by both parties. It can be invented outright, such as a CPM filled in when pricing data was missing, or it can creep in through paraphrase.

How is hallucination different from semantic drift?

Hallucination produces content with no basis in the inputs, while semantic drift happens when two models interpret the same agreed words differently. Both end with records that do not match, and both are caught by checking against a shared record.

Can hallucination in buying agents be eliminated completely?

No current method guarantees that a language model never produces unsupported output. The practical goal is detection, using structured records, price provenance and human approval to make invented terms visible before they cost money.

What is the quickest check a buyer can add today?

Compare every term in the agent's confirmation, field by field, with the structured record of the deal rather than the chat transcript. Any field that cannot be traced to that record should be treated as unagreed until it is confirmed.

Do AI Buying Agents Make Things Up? How Invented Terms Get Into Deals and How to Catch Them