TL;DR: When a buy-side agent executes a deal outside its mandate, the immediate priority is to determine whether the deal can be unwound, who needs to be notified, and how the mandate failed to prevent it. The root cause is almost always a mandate gap rather than an agent failure: the agent acted consistently with its instructions, but the instructions did not cover the situation it encountered. Fixing the agent without fixing the mandate produces the same problem on a different impression.

Out-of-mandate agent behaviour is a governance event, not just a campaign anomaly. When an AI agent commits to a deal that falls outside the parameters its buyer authorised, it has taken a financial action without authority. The appropriate response is the same as for any governance event: contain, investigate, remediate, and prevent recurrence.

The instinct to treat it as a technical problem, to be fixed by adjusting a platform setting, misses the point. The question is not only how the agent made the decision it made; it is why the mandate did not prevent it.

Step one: contain

The first question is whether the deal can be unwound. This depends on the marketplace infrastructure and the deal terms. In a bilateral deal recorded in a DealSheet, the deal terms are agreed and the DealSheet is binding. Unwinding requires a mutual agreement between both parties, which may or may not be possible depending on the marketplace's deal management capabilities and the publisher's willingness to release the commitment.

If the deal cannot be unwound, the buyer needs to decide whether to proceed with delivery and reconcile the out-of-mandate cost, or to pause delivery and accept any break fees or relationship consequences. This is a business decision, not a technical one, and it should involve the person who holds budget authority for the campaign, not only the trader who manages the campaign.

While the decision is being made, the agent should be paused. Allowing an agent to continue operating after an out-of-mandate event, before the mandate gap has been identified and closed, risks further out-of-mandate activity.

Step two: investigate

The investigation has two objectives: understanding the specific decision the agent made, and identifying the mandate gap that permitted it.

For the specific decision, the agent's decision log should contain a record of the proposal the agent received, the parameters it evaluated, and the decision it reached. In a well-designed system, this log is sufficient to reconstruct the decision: what the agent was told, what options it evaluated, and why it selected the course of action it did.

For the mandate gap, the question is: what aspect of the mandate was absent or ambiguous enough to permit this decision? Common gaps include: inventory scope that did not explicitly exclude the publisher or domain where the out-of-mandate deal was made; a CPM range that did not account for the specific inventory format; an approval threshold that was set at a level that the deal fell just under; and audience parameters that did not exclude the specific segment combination the agent used.

In most cases, the mandate gap is a specification error, not a system failure. The agent behaved consistently with its instructions; the instructions did not include the constraint that would have prevented the behaviour.

Step three: notify

Who needs to know depends on the severity and the commercial context.

Within the agency or trading desk, the account lead and the person with budget authority need to know. If the deal creates a cost overrun beyond a stated tolerance, the finance team needs to know. If the deal involves a publisher or inventory category that the client has specifically excluded, the client account manager needs to know and may need to inform the client.

The notification does not need to be alarming. Out-of-mandate events at low financial magnitude are governance events, not crises. The appropriate tone is factual: this happened, this is how it happened, this is what we are doing to prevent recurrence.

If the out-of-mandate event is material (significant cost overrun, brand safety breach, data protection concern), the notification may need to be escalated to legal or compliance, and the client may need to be informed directly.

Step four: remediate the mandate

The mandate must be updated before the agent is reactivated. The update should close the specific gap that permitted the out-of-mandate decision, and it should be reviewed by someone other than the person who originally wrote the mandate.

The updated mandate should include an explicit record of the change: what was added, what was changed, and why. This change log is part of the audit trail for the event: if the same buyer has a future out-of-mandate event, the reviewer needs to know whether the previous event informed the mandate revision.

The agent should be reactivated only after the mandate has been updated, reviewed, and approved by the person holding budget authority. Reactivating an agent on an unchanged mandate after an out-of-mandate event is a governance failure, not a standard operating procedure.