IAB Tech Lab Ships AAMP 2.3 for Autonomous Programmatic Buying

IAB Tech Lab Ships AAMP 2.3 for Autonomous Programmatic Buying

The IAB Tech Lab published version 2.3 of AAMP, its protocol for letting AI agents buy programmatic advertising without a human in the loop. The update adds hooks into Amazon Bedrock (Amazon's cloud AI platform) and Databricks (the data and AI cloud), extends buying to Meta's inventory, and pulls Google Ad Manager reporting into the agent layer. It also bakes brand-safety and privacy rules into the protocol itself.

Buried in the spec is a clause requiring that inventory be "real, not derived." Somebody had to write that down, in 2026, after thirty years of RTB. A spec addendum does not fix a supply chain that still cannot reliably prove an impression exists. It just gives the old fraud a new place to hide, moving at machine speed.

The platform war hiding inside a standards release

The Bedrock and Databricks integrations are the part that actually matters for market structure. This is the header-bidding wrapper fight replayed at the agent layer: whoever owns the trusted agent runtime owns the toll booth. The question stops being "which DSP has the best UI" and becomes "which cloud runs the robot that spends the budget." Google wiring Ad Manager into the reporting layer is a defensive move, keeping measurement tied to its pipes while buying logic drifts toward the clouds. Amazon and Databricks sit on the winning side of that shift if AAMP gets real adoption.

Independent SSPs like Magnite and PubMatic are exposed. Their PMP and direct-deal relationships assume a human buyer who values those relationships. An autonomous agent has no loyalty to either.

Why this probably sits on a shelf

The IAB has shipped standards the industry then ignored. AAMP needs coordinated, honest implementation across buy-side platforms, sell-side platforms, two clouds, and two walled gardens simultaneously. Meta and Google have every incentive to adopt the parts that favor their own inventory and slow-walk the rest.

On the operator side, the tooling gap is real. DSP integrations, brand-safety checks, and reconciliation workflows all assume a human closes the loop. Autonomous buys generate discrepancy volumes no AdOps team can eyeball in real time, and nobody has yet answered who eats the loss when an agent buys at the wrong CPM or on junk inventory. That liability question will keep procurement and legal blocking production deployment until someone puts an answer in writing.

Our call: By the IAB's ALM in early 2027, no top-10 agency holding company will have moved a disclosed, material share of programmatic spend through fully autonomous AAMP agents without a human approval step in the loop. The supply chain integrity problem the spec itself confesses, combined with an unresolved liability gap and the IAB's own track record on adoption timelines, makes supervised pilots the base case. Real autonomous spend at scale would require the walled gardens to implement consistently and the liability question to close. Neither is close.

If you are evaluating a pilot: hard spend caps, mandatory reconciliation, and legal sign-off on liability before an agent touches real money.

Your draft

The IAB Tech Lab shipped version 2.3 of AAMP, its protocol for letting AI agents buy ads on their own, no human in the loop. It adds hooks into Amazon Bedrock and Databricks, extends buying to Meta, pulls in Google Ad Manager reporting, and bakes brand-safety and privacy into the agent layer. Buried in the spec is a line requiring that inventory be "real, not derived." Somebody had to write down, in 2026, that the ad should actually exist.

Reversibility: Type 2 for any single operator. Nobody has to adopt AAMP today, and pulling out of a pilot costs little. But the platform positioning it enables (Amazon and Databricks sitting inside the buyer agent layer) is Type 1 once budgets route through it. What's actually being decided: not "do we use AAMP," but "whose agent infrastructure do we trust to spend our money." Forcing function: none yet. This is a standards release, not a deadline. The pressure will come from whichever walled garden makes agentic buying the path of least resistance first.

---

The Market Analyst follows where the integrations point. Bedrock and Databricks inside the buyer agent layer is the header-bidding wrapper fight again: whoever owns the trusted agent runtime owns the toll booth. In plain terms, the money question stops being "which DSP has the best dashboard" and becomes "which cloud runs the robot that spends the budget." Google wiring in Ad Manager reporting is a defensive crouch, keeping measurement tethered to its pipes while buying logic drifts elsewhere. The losers if this sticks are the independent SSPs, Magnite and PubMatic, whose PMP and direct-deal relationships mean nothing to an agent with no loyalty. The winners are the clouds and the walled gardens that host the agents.

The Skeptic notes the IAB Tech Lab has shipped standards the industry then ignored. Adoption, not publication, is the hard part, and AAMP needs buy-side, sell-side, two clouds, and two walled gardens to all implement it honestly and consistently. Meta and Google have every reason to adopt the parts that favor their own inventory and slow-walk the rest. The "real, not derived" clause is the confession. If thirty years of RTB could not reliably prove an impression exists, a spec addendum does not fix it. Plainly: they automated the negotiation before they fixed the thing being negotiated over. The premium here funds pitch decks, not production traffic.

The Operator on Tuesday morning watches your trading desk read the spec and hit a wall. Your DSP integrations, brand-safety checks, and reporting all assume a human closes the loop. Bedrock and Databricks compatibility is plumbing, not a shortcut past the six months of wiring agent decision logic into your campaign systems. First thing that breaks is reconciliation: autonomous buys throw off discrepancy volumes no AdOps team can eyeball. By day 90, procurement and legal are asking who eats the loss when an agent buys at the wrong CPM or on junk inventory. Nobody has that answer. In plain terms: the robot can spend faster than your people can check its work.

The CFO wants the payback, not the feature list. The pitch is efficiency, fewer humans clicking buttons, but the real cost is the new dependency. Route spend through an agent runtime you do not control and you have handed your negotiating leverage to whoever hosts it. That is the Amazon and Databricks position, and it is not free even when it looks free. Add the tail risk: an agent transacting autonomously on non-compliant or fictitious inventory is a fraud-loss and brand-safety exposure with no clear owner. Not funding autonomous buying until I know who pays when it buys air.

---

Where they part ways

The Market Analyst and the Skeptic disagree on whether this matters at all. The Analyst treats AAMP as a live platform war worth positioning against now. The Skeptic says show me adoption first, because IAB specs have died on the vine before. That gap is the whole call: if AAMP gets ignored like past standards, the platform-war framing is premature. If a walled garden makes agentic buying the default path, the Skeptic's caution costs you a seat at the table.

The Operator and the CFO agree on the risk but split on tempo. The Operator sees a tooling gap that closes with engineering time. The CFO sees a structural dependency that engineering time makes worse, because the better your agent works, the more spend you have surrendered to someone else's runtime.

---

What it hinges on

Two beliefs. First, whether the walled gardens implement AAMP honestly or carve it up to favor their own inventory. Nothing in a Meta or Google incentive structure suggests they play it straight, and the "real, not derived" clause tells you the chain has an integrity problem the spec cannot enforce. Second, whether budget actually migrates to autonomous agents fast enough to matter in the next year, or whether this is another standard that sits on a shelf.

The council leans skeptical on near-term impact and wary on long-term positioning. Stacking autonomous negotiation on a supply chain that still cannot prove an impression is real just gives the old fraud a new place to hide, moving at machine speed. Before committing anything: run a walled pilot with hard spend caps and mandatory reconciliation, and get legal to answer the liability question in writing before an agent touches real money.

---

Prediction: By the IAB's ALM in early 2027, no top-10 agency holding company will have moved a disclosed, material share of programmatic spend through fully autonomous AAMP agents without a human approval step in the loop.

Confidence: Medium. Standards ship years before real adoption, and the liability question is unanswered.

Why: The spec itself confesses the supply chain cannot prove inventory is real, which is exactly the condition that makes buyers keep a human on the approve button. No CFO or general counsel signs off on autonomous spend when nobody can say who pays for a bad buy, and that answer does not exist today. The IAB has a track record of publishing standards that take years to reach production, so the base rate favors slow, human-supervised pilots over hands-off automation. The opposite outcome, real autonomous spend at scale within a year, would require the walled gardens to implement consistently and the liability gap to close, and neither is close.

Revisit by 2027-03-01: We're right if agencies are running AAMP only in supervised pilots with humans approving buys. We're wrong if a holding company publicly reports material spend flowing through fully autonomous agents with no human in the loop.