TL;DR: The failure modes of AI-powered advertising are: agents acting outside the scope of their mandate, mandate parameters that are too vague for the agent to evaluate correctly, deal records that are not audited after the fact, and governance gaps where no human is responsible for reviewing the agent's decisions. These are not failures of AI technology; they are failures of mandate governance, and they are addressable with the same discipline that governs any other delegated decision process.

The risks of AI in advertising get discussed in two unhelpful ways. One framing treats AI as an uncontrollable black box that will make unpredictable decisions. The other dismisses all risk as theoretical until something goes wrong. Neither is useful for a buyer or publisher who needs to govern an agent deployment responsibly.

The realistic risk analysis starts from a specific question: in what circumstances does an AI agent make a decision that the human principal would not have approved, and how would we know?

Mandate gaps: the most common failure mode

The most common cause of agent decisions the principal would not have approved is a mandate that did not specify what the agent needed to know. An agent evaluates proposals against its mandate parameters. If the mandate does not address a parameter that comes up in a real proposal, the agent either defaults to a built-in behaviour, which may not match the principal's intent, or rejects the proposal and escalates, which is correct but may slow down the campaign.

Mandate gaps appear in several forms. A buyer's mandate might specify a CPM ceiling but not specify which inventory categories are excluded from the brand's adjacency requirements. The agent agrees a deal within the CPM parameters on inventory that a human would have excluded on brand safety grounds. The problem is not agent failure; it is mandate incompleteness.

A publisher's mandate might specify a CPM floor but not specify which categories of buyer the publisher does not want to deal with. The agent accepts a proposal from a buyer in a category the publisher would have declined if a human had reviewed it. Again: mandate incompleteness, not agent failure.

The remedy is mandate completeness: specifying all the parameters the agent needs to evaluate, not just the most obvious ones.

Approval thresholds set too high

A related risk is approval thresholds that are set at a level that rarely triggers human review. If the approval threshold for a deal is set at a value that no realistic deal would exceed, the threshold provides no governance value. The agent operates with effective autonomy, and the human oversight that the threshold was designed to provide never activates.

Approval thresholds need to be set at realistic values: the level at which a human would genuinely want to review a deal before it is agreed. For a buyer deploying an agent for the first time, this typically means setting low thresholds initially, reviewing which deals trigger escalation, and raising the thresholds gradually as confidence in the agent's decision quality builds.

This is the earned autonomy model: the agent's autonomous scope is expanded incrementally as the evidence supports it, rather than set high from the start on the assumption that the agent will behave correctly.

Deal records not reviewed post-campaign

An agent that negotiates bilateral deals produces a deal record for every deal it completes. If no human reviews these records post-campaign, the oversight function that the deal record is designed to support does not operate. The agent may have made decisions consistent with the mandate, or it may have made decisions that a human reviewer would flag as out of scope. Without post-campaign audit, neither situation is detectable.

Post-campaign deal record review is a governance requirement, not an optional step. Buyers who are running agents should have a defined process for reviewing the DealSheet for every deal above a defined threshold, comparing the agreed terms against the campaign mandate, and investigating any deal that appears inconsistent with the mandate parameters.

This does not need to be a line-by-line review of every deal. It should be a systematic sample at minimum, with automated flagging of deals that trigger any of the mandate parameters at the boundary of the authorised range.

Accountability gaps: no human responsible for agent oversight

A structural risk that is easy to underestimate is the accountability gap: a situation where the agent is deployed, but no human has clear responsibility for reviewing its decisions, responding to escalations, and updating the mandate when gaps appear.

In a trading desk with a conventional programmatic workflow, every campaign has a named campaign manager who is accountable for performance. In an agent deployment, if the accountability is diffuse, "the agent handles it" becomes a way of describing a situation where nobody is specifically responsible for the agent's decisions.

This is a management problem, not a technology problem. Every agent deployment should have a named human responsible for the mandate, the approval workflow, and the post-campaign audit. If that responsibility is not assigned explicitly, the governance function does not operate regardless of how well the technology works.

What good governance looks like

The failure modes of AI advertising are all addressable. A complete mandate, realistic approval thresholds, a systematic post-campaign audit, and clear human accountability for agent oversight: these four elements together constitute a governance framework that makes AI agent use in advertising controllable and auditable.

The technology is not the risk. The governance is the requirement.