What Will an Advertising Agent Be Allowed to Believe? Part 1
By Arian Senior, Director, Operations & Innovation at Utiq
For fifteen years, identity in programmatic advertising has been an execution-layer utility: recognise an audience, manage frequency, reconcile events and, where the same identifier is available across environments, understand incremental reach.
Agentic advertising changes the question. Buying agents will not just bid on impressions. They will interpret briefs, compare proposals, negotiate terms and move approved plans towards execution.
That means an agent will have to do something humans have always done, often imperfectly: decide what information to trust. That may be the most important identity question of the agentic era.
From buying impressions to evaluating proposals
In September, IAB Tech Lab released AAMP 3.0 and its OpenProposal specification, extending agentic workflows upstream into the RFP and proposal stage. OpenProposal gives sellers a standard, machine-readable way to describe what they offer – placements, formats, audiences, volumes and prices – and to respond to buyer briefs. In the Ad Context Protocol (AdCP), the unit is already the campaign: a brief written in plain language becomes one structured buy across several sellers and channels.
What an agent does next depends on how the campaign is bought. In direct deals – whether IO-based or programmatic guaranteed – an agent can increasingly handle the brief, negotiation, booking and reporting, while delivery remains optimised by the publisher or platform executing the buy. When buying across publishers through platform-level agents, such as PubMatic’s AgenticOS, it can also shift budget between deals and manage frequency within the platform that executes the buy.
In both cases, the agent starts by comparing proposals – and that is where the difficulty begins. One publisher forecasts reach on one identifier, another on a different population. One can explain exactly how its audience was counted; another provides only a modelled estimate.
Humans compensate for those differences. Agents cannot, unless the differences are made explicit. An agent can compare twenty proposals in seconds; that does not mean it knows which one is right.
Speed is not judgement. Standardisation is not truth.
The first agentic problem is not automation. It is trust.
Protocols such as AAMP and AdCP make it far easier for systems to exchange information and execute workflows. But standardising how an action is requested does not improve the quality of the decision. A standard interface can activate a weak audience definition just as efficiently as a good one.
If an agent is told that Publisher A can deliver ten million people and Publisher B eight million, it needs to know what those numbers mean. Are they:
- unique people or identifiers?
- observed or modelled?
- deduplicated or gross?
- measured on the same population, across the same environments?
The agent does not need every answer to be identical. It needs to know the difference. Hence a simple principle:
Before an agent can optimise a proposal, it must be able to understand the basis on which that proposal was measured.
Identity becomes part of the proposal
Today, identity is discussed in relation to the impression: who can be recognised, targeted or frequency-capped. In an agentic market, it also becomes part of the metadata around the transaction. The question is no longer only “who is this impression?” but:
- What population does this seller’s forecast represent?
- How much of it overlaps with another seller’s?
- How much incremental reach does this proposal really add?
- Which parts of that number are observed, and which are modelled?
An identifier alone cannot answer those questions. They require identity, measurement and governance to work together.
There will not be one answer
A common identifier is not the only way to make audiences comparable. Clean rooms, first-party identity graphs, probabilistic measurement and contextual signals all play a part, and IAB Tech Lab’s Agentic Audiences initiative defines how agents can exchange identity, contextual and reinforcement signals.
That diversity is healthy. The objective is not one identity system for everyone. It is to make the basis of every claim visible and comparable to the agent making the decision.
For agents to make decisions with confidence, understanding the basis of each claim is only the beginning. The harder question is what happens when those claims must be combined into a single media plan.
Read Part 2 of Arian’s essay here on Utiq.com
