A publisher can spot a low-quality AI buying agent by checking five things before it accepts any terms: who the agent acts for, what evidence of authority it carries, whether its operator has registered it, how it behaves when a proposal is countered, and whether it can produce a record of what it agreed. An agent that fails the first two checks should not receive a price. An agent that fails the last three should not receive a deal. Refusing such an agent is not lost demand. It is ordinary counterparty risk management, applied to software.

TL;DR. A buying agent is only as credible as the principal behind it, so the first check is which advertiser or agency the agent represents and whether that principal will confirm the request. Mandate evidence, such as a budget ceiling, permitted deal types and a named human approver, separates an agent with authority from one fishing for prices. A listing in IAB Tech Lab's Agent Registry shows that a company has registered an agent; it does not prove a given request is authorised. A publisher can write all five checks into a sell-side policy its own agent enforces.

What makes a buying agent low quality from a publisher's side?

For a publisher, a low-quality buying agent is one whose commitments cannot be relied on. That covers three problems. An unaccountable agent has no identifiable principal behind it, and can be used to harvest rate cards and floor prices. An unauthorised agent belongs to a real company but has no evidence it may spend that company's money, so its deals may later be disowned. A badly built agent negotiates erratically, ignores the publisher's terms or keeps no usable record, and produces disputes at invoice time. None of the checks below depends on how sophisticated the agent's model is.

Who is the buying agent's principal?

The principal is the advertiser, agency or other organisation on whose behalf the agent acts, and which carries responsibility for what the agent commits to. An agent is not a legal entity. Without a named principal, a publisher has no one to invoice or hold to the terms.

A credible request names two parties, which are often different: the operator, meaning the company that built or runs the agent, and the principal it acts for. The publisher should be able to match the principal to an existing commercial relationship or open one, with a contact who can confirm the request and billing details. An agent that will not name its principal, or names one that will not confirm, has failed the most important check.

What evidence of a mandate should a publisher ask for?

A mandate is the scope of authority a human principal gives an agent. A publisher does not need to see an advertiser's whole mandate, only evidence that this request sits inside one.

Useful evidence includes a budget ceiling for the deal, the deal types the agent may agree (programmatic guaranteed, preferred deal or private marketplace), the threshold above which a human must approve, and the name or role of that human. For example, in a hypothetical case, an agent proposing a £60,000 programmatic guaranteed deal that states a human approval threshold of £25,000 should be able to say who approves the commitment and when. If it cannot, the publisher should treat any agreement as provisional until a person confirms it.

Some protocols carry part of this evidence inside the transaction: AdCP, the Ad Context Protocol published by AgenticAdvertising.org, describes a governance task that checks budget limits before a buy executes and escalates to a person when an action exceeds the agent's authority. Otherwise, the publisher should ask for the evidence directly.

Does an Agent Registry listing prove a buying agent is legitimate?

IAB Tech Lab runs an Agent Registry, part of its Tools Portal, where companies register the MCP and A2A agents they bring to market, categorise them and state their purpose and maturity. Each registering company is tied to its GPP ID, and IAB Tech Lab reviews entries for the required validation information. The registry is free for members and non-members.

A listing is a useful signal of operator identity: a named company has put its agent on public record and accepted review. It is not proof that a particular request is authorised by a particular advertiser. A registered agent can still exceed its mandate, and an unregistered one may belong to a legitimate small operator. A sensible policy treats the registry as one input: an unlisted agent faces more questions about its operator, and a listed agent still faces the principal and mandate checks.

How does a low-quality buying agent behave in negotiation?

Negotiation behaviour is the most direct evidence a publisher gets. Warning signs include:

requesting rate cards or floors across many products with no stated budget or flight dates;
accepting the first offer on every product instantly, which suggests price discovery rather than buying;
reopening agreed terms after acceptance, or sending duplicate requests that could create a second deal;
proposing again terms the publisher has already excluded, such as a blocked category or an unsupported creative format;
being unable to escalate to a person when a counter-offer falls outside its limits.

A well-built agent states its constraints, counters within them and hands the decision to a human at its limit. Pausing for approval is a sign of quality.

What record should exist before a publisher accepts a buying agent's terms?

Before a deal goes live, both sides should hold the same written record of what was agreed: price, flight, inventory, audience, measurement basis, any viewability standard, and who approved the deal. An agent that can only point to its own internal logs is asking the publisher to trust them. A proposal that leaves the measurement basis unstated is an early warning of a later dispute, because delivery will be judged against terms nobody wrote down. The same standard applies to the publisher's own sell-side agent.

How should a publisher refuse a buying agent?

Refusal works best as written policy rather than case-by-case judgement. A publisher can turn the five checks into sell-side rules: no pricing without a named principal; no deal above a set value without mandate evidence and a named human approver; extra questions for operators not listed in the Agent Registry; automatic referral to a person after anomalous behaviour; and no activation until a shared record exists. Each refusal should state which item is missing, so a legitimate buyer can fix it and come back.

Alkimi, a neutral agentic advertising marketplace working within IAB Tech Lab's AAMP framework, runs the authority check before negotiation starts: it verifies that each connecting agent is authorised by the buyer or publisher it represents, and agreed terms are written to a DealSheet that both parties hold and either can retrieve independently.

Frequently asked questions

Can a publisher refuse to deal with AI buying agents altogether?

Yes. A publisher can decline to negotiate with agents, as it can decline any buyer. Most will find a policy more useful than a ban: accept agents that pass principal, mandate and record checks, and refuse those that do not.

Is an agent listed in the IAB Tech Lab Agent Registry safe to deal with?

A listing shows that a named company registered the agent and that IAB Tech Lab reviewed the entry for validation information. It does not show that a specific request is authorised, so the principal and mandate checks still apply.

Who is responsible if a buying agent agrees a deal its advertiser did not authorise?

Not the agent, which is not a legal entity. Whether the operator, the agency or the advertiser carries the loss depends on the contracts between them, which is why publishers should confirm authority before a deal activates rather than after.

What is the quickest single test of a buying agent?

Ask it to name its principal and the person who approves spend above its threshold. An agent that answers neither should not receive pricing.

How Can Publishers Spot and Refuse a Low-Quality AI Buying Agent?