Standards & regulation

IAB Tech Lab Ships AAMP 2.3 — Agentic Buying Gets Its Guardrails

APH Regulation Desk ·4 min read Share Print
In this piece
    Figure The four points the piece sets out Why this matters for publishers →
    1. 01

      Your reporting is now machine-readable sales collateral.

    2. 02

      The guardrails only protect implementers.

    3. 03

      Agents extend no benefit of the doubt.

    4. 04

      The standard is your refusal framework.

    Regulation Desk

    Agentic advertising has spent two years arriving as speculation — panels, prototypes and press releases about a future in which software buys media from software. This week it arrived as a spec you can implement. The IAB Tech Lab shipped AAMP 2.3, and the release notes read less like a vision document and more like plumbing.

    The headline item: Agentic Audiences v1.0 is finalised and ready for programmatic transaction — audience definitions that buyer agents can act on are no longer a draft. Around it, the release adds Google Ad Manager reporting support and Meta buying support, brings full OpenDirect spec validation, and wires privacy diligence directly into buyer agent workflows via the IAB Diligence Platform and SafeGuard Privacy. For sell-side teams, though, the substantive parts are the two guardrails: verification against real inventory rather than derived numbers, and deterministic, provable price commitments with human approval gates.

    Adform CTO Jochen Schlosser framed the intent precisely: agentic advertising should build “on, not around, programmatic controls.” That sentence is the whole battle, stated politely. The alternative — agents transacting around the control layer the industry spent fifteen years constructing — is the disintermediation scenario every publisher should be pricing.

    01The GAM hook, and why the guardrails need implementing

    For a Google Ad Manager operator, the concrete change in 2.3 is the reporting support. Buyer agents can now read the delivery data your stack produces — which quietly converts your reporting hygiene into sales collateral. Naming conventions, clean key-values, accurate inventory mappings and consistent line-item structures were previously internal virtues; under AAMP they are what a buyer’s software sees when it evaluates you as a counterparty. A GAM instance that a human trafficker can navigate but a machine cannot parse is now a commercial liability with a specification number attached.

    The guardrails deserve a closer reading than a standards release usually gets, because both are protections that exist only for sellers who wire them in. Verification against real inventory means an agent can check your avails against reality instead of trusting derived or modelled numbers — protection against both overselling and being undersold by someone else’s model of your inventory. Deterministic, provable price commitments with human approval gates mean an agent cannot silently negotiate your pricing into places no human on your side approved. Neither happens by default. Schlosser’s “on, not around” is an intent, and intents bind nobody; the spec becomes protection at the moment of implementation, not publication.

    There is also a floor-setting function here that is easy to miss. Once a standard exists specifying verified inventory, provable pricing and human approval gates, any agentic buyer unwilling to transact within those controls is telling you something — and the standard gives you the vocabulary to refuse. Before AAMP 2.3, declining an opaque agentic integration was a judgement call; now it is a compliance position.

    02Why this matters for publishers

    Your reporting is now machine-readable sales collateralWith GAM reporting support in the spec, buyer agents evaluate you through your delivery data. Stack hygiene has moved from operational nicety to revenue-relevant surface area.
    The guardrails only protect implementersInventory verification and provable price commitments are opt-in engineering, not ambient safety. Sellers who skim the release get the risks of agentic demand without its protections.
    Agents extend no benefit of the doubtA human buyer tolerates a discrepancy and calls you; software deprioritises you and moves on. The tolerance band for sloppy avails and asserted pricing is about to narrow sharply.
    The standard is your refusal frameworkAny agentic buying platform that resists transacting within AAMP's controls has told you how it intends to treat your pricing and your data. "We transact within the standard" is now a complete answer.
    Standards releases are easy to ignore and this one should not be, because it marks the moment agentic buying stopped being a thing that happens to the industry and became a thing the industry specified.

    03What publishers should do

    04The bottom line

    Standards releases are easy to ignore and this one should not be, because it marks the moment agentic buying stopped being a thing that happens to the industry and became a thing the industry specified. The controls in AAMP 2.3 are genuinely publisher-protective — verified inventory, provable prices, humans in the loop — but they are protections in the way a lock is protection: only once installed. The sellers who implement early become safe counterparties for automated budgets as those budgets arrive; the sellers who wait will discover that software, unlike a buyer with a relationship to protect, simply routes around whatever it cannot verify.

    Sources & caveats

    Sources: IAB Tech Lab press release, “IAB Tech Lab Releases AAMP 2.3, Bringing Enterprise-Grade Infrastructure and Privacy Diligence to Agentic Advertising” (July 30, 2026), for the release contents — Agentic Audiences v1.0 finalisation, Google Ad Manager reporting and Meta buying support, OpenDirect validation, the IAB Diligence Platform and SafeGuard Privacy integrations, the inventory-verification and price-commitment guardrails — and Jochen Schlosser’s quote. This is a standards-body announcement; adoption levels and real-world enforcement of the guardrails are not yet observable, and characterisations of how agents will behave toward non-compliant sellers are analysis.

    The weekly

    One letter a week, from the desk that runs the auctions.

    What actually moved in yield, CTV and curation across our publishers — written by the people who saw it, not a content team. No digests, no roundups, one email.

    One email a week. Unsubscribe in one click. We never share or sell the list.

    More from this issue

    Ran alongside this piece in the Weekly of 2 August 2026 — read the whole issue →