Last Wednesday the IAB Tech Lab released a draft framework called Programmatic AI Disclosure and Measurement that almost no one in the industry has properly read. By the time the comment period closes in July, most of the buy-side will discover that they cannot answer the questions the framework asks. The sell-side will be in worse shape.
The trade press treated the release as a routine standards announcement. It is not. It is the first time a sanctioning body has codified what "AI-driven programmatic" means in a way that audit teams can apply, the first time inventory will have to be classified as either AI-decisioned or not, and the first time the difference between the two will affect what a buyer is allowed to claim it bought.
What the framework actually mandates
The draft sets four reporting obligations on any platform that wants to call its supply or demand AI-driven.
- Inventory disclosure — the share of impressions decisioned by AI models, by impression and by spend, declared per campaign
- Decisioning lineage — which signals fed the model decision (audience inputs, contextual inputs, brand-safety inputs, supply-path inputs), with the share of weight each carried
- Output verifiability — a quarterly audit-trail sample of decisions that exposes the model output, the human override rate, and the suppression-list propagation latency
- Bias and fairness — a quarterly self-report on demographic skew of audience inclusion, plus a third-party-attestable test suite
Two things are immediately obvious. The first is that the IAB has chosen to make the standard about the decisioning system, not the outcome. The second is that the standard treats AI-driven inventory the way GAAP treats off-balance-sheet finance — as something the audit must be able to see, even if the buyer cannot directly interrogate it.
The three things stacks were never built to do
Most modern programmatic stacks were built to optimise CPM, win rate and reach. They were not built to attribute decisioning lineage. They were not built to expose the share of weight each signal carried inside a bid model. They were not built to maintain a quarterly self-report against a bias test suite.
- Stacks built before 2022 generally cannot expose lineage because the bidder was designed as a black box
- Stacks that bolted on ML in 2022–24 added it as a layer rather than rewriting the bidder, and lineage exposure was not in the spec
- Even the small number of bidders that can expose lineage do not do it at the per-impression granularity the framework requires
The result is that on the day the framework finalises, somewhere between sixty and eighty per cent of the supply that today bills itself as AI-optimised cannot lawfully describe itself that way to a buyer that is conducting due diligence.
“Most stacks will fail their first IAB AI audit not because they are bad, but because they were never built to be audited.”
Why this is harder than viewability was
Viewability — the previous IAB-led measurement reform — was hard. It was also one-dimensional. Was the pixel rendered, for how long, in the active viewport. Stacks took two years to comply. Adoption took five.
AI measurement is multi-dimensional. The framework asks for a decisioning lineage that holds across audience-input changes, supply-path changes, creative-input changes and brand-safety-input changes. It asks for a record of the human override rate per campaign. It asks for a propagation latency on suppression lists — meaning, from the moment an opt-out is recorded, how long until that opt-out actually stops the spend.
Stacks that pass viewability tomorrow will not pass AI measurement next quarter. The audit cost will be material. The opportunity cost will be larger.
Who wins and who loses
The early winners are the small number of independent bidders that have already invested in lineage-exposure tooling. Their cost of compliance is low, their marketing line (audit-pass on day one) is strong. The early losers are the walled gardens. The walled gardens have the technology to comply. They have a structural reason not to comply at the granularity the framework requires: the lineage they would have to expose is, in commercial terms, the source of their bid-side advantage.
- Independents with lineage tooling win on disclosure positioning, and on premium CPMs that follow
- Independents without must build lineage exposure or be reclassified
- Walled gardens comply at the minimum, with a fight over what "AI-driven" means at the edge of their disclosure
- Smaller DSPs face a squeeze, with attendant consolidation pressure inside 2027
- Buy-side measurement vendors are net winners. This is a new mandatory product line for them
The 90-day audit plan
For an advertiser, the right move this quarter is simple and unglamorous. Bring your three biggest programmatic partners into a room. Ask each one a single question: when the IAB Tech Lab framework finalises, what share of our spend will you be able to certify? Get it in writing.
- Ask for the share of impressions decisioned by AI, by spend
- Ask for a sample decisioning lineage report for one week of one campaign
- Ask for the suppression-list propagation latency, in seconds
- Ask for the bias test suite results from last quarter
- Ask which questions they cannot answer, and when they will be able to
The partners that answer cleanly are the partners that survive the next eighteen months. The partners that hedge — "we are working on it" — are the partners whose contracts you should plan to consolidate or sunset. The IAB will eventually attach this framework to MRC accreditation. When that happens the difference between an answer and a hedge will be the difference between a partner and a former partner.
Programmatic media has lived in a measurement gap since 2009. The gap is closing. The buyers who use the next six months to require their supply chain to ready itself for the audit will end the year buying audited supply at a fair price. The buyers who do not will end the year buying the inventory no one else wanted.