TL;DR: Auditing a deal negotiated by an AI agent is not the same as auditing a traditional programmatic transaction. The record exists, but accessing it and reconciling it against your own books requires asking different questions. This piece sets out the practical audit checklist buyers need now: what documentation to request, what reconciliation to run, and where the mandate trail must be visible. If your current vendor cannot answer these questions, that is information.
When a human negotiates a deal, the audit trail is often a conversation, an email chain, and an insertion order. When an agent negotiates a deal, the record should be more complete, not less. The agent has operated from a mandate, consulted a deal record, and executed against defined parameters. If you cannot retrieve all three, the deal has not been audited; it has been accepted.
Programmatic buyers are accustomed to discrepancies. The standard tolerance in the industry has historically been a plus or minus 10% variance between buy-side and sell-side impression counts, a number that the MRC's viewability guidelines have helped sharpen but not eliminate. Agent-negotiated deals do not automatically improve on this. Without a shared record and a reconciliation method that both parties can run against the same source, discrepancies may be harder to catch and easier to obscure.
This is the audit checklist buyers need before, during, and after an agent-negotiated deal.
What documentation should a buyer be able to retrieve from any agent-negotiated deal?
Four documents should be retrievable for every agent-negotiated deal. If any of them cannot be produced on request, the deal lacks an auditable record.
The first is the mandate record. Before an agent acts, it should operate from a buyer-approved mandate: the parameters within which it is authorised to negotiate, the approval thresholds it must not exceed without human sign-off, and the constraints on inventory, audience, and price. The mandate record is the document that establishes whether the agent acted within scope.
The second is the deal record itself. This is the outcome of the negotiation: the agreed price, the inventory specification, the term, the parties, and the timestamp. In a bilateral agent-to-agent transaction, the deal record should be accessible to both buy-side and sell-side without either party being able to edit it after execution. Any platform that cannot produce this document on demand does not have an auditable deal.
The third is the approval log. Where the agent escalated to a human, or where a human overrode the agent, that handover point should be timestamped and retrievable. A clean approval log is also the primary defence against liability disputes: it shows who authorised what and when.
The fourth is the delivery reconciliation. Post-campaign, the delivered impressions, the agreed impressions, and any variance should be reconcilable against the deal record. If your platform produces only a delivery report without the underlying deal record to reconcile it against, you are auditing delivery, not the deal.
How should buyers reconcile an agent-negotiated deal against their own books?
Reconciliation in traditional programmatic often means matching a DSP delivery report against a publisher's impression log. In agent-negotiated deals, there is an additional layer: the deal record itself. Buyers should run three reconciliation passes.
The first pass is mandate-to-deal reconciliation. Compare the agent's executed deal against the parameters in the mandate. Did the agent stay within the approved price range? Did it respect the inventory constraints? Did it escalate correctly where the mandate required human sign-off? Any deal that falls outside mandate parameters is a compliance event, not just a variance.
The second pass is deal-to-delivery reconciliation. Compare the delivery report against the agreed terms in the deal record. Volume, format, placement, and date range should all match. Variances above a pre-agreed tolerance should be logged and queried with the platform.
The third pass is financial reconciliation. Match the invoiced amount against the agreed deal record price. In traditional programmatic, invoice-to-actual discrepancies are a known problem. According to the ANA's 2023 programmatic transparency study, 22% of agencies surveyed reported difficulty reconciling programmatic invoices against actual delivery. Agent-negotiated deals do not solve this problem automatically. If the deal record and the invoice do not share a common reference number, reconciliation requires a manual lookup, which is both slow and error-prone.
What questions should buyers ask vendors before running agent-negotiated deals?
The following questions should be asked of any platform before committing budget to agent-negotiated deals. They are not hostile questions; they are operational questions. A vendor that cannot answer them is not ready.
Can you produce the deal record on request? In what format, within what timeframe, and via what access method?
Is the deal record immutable after execution? Can either party edit it? Who holds the authoritative copy?
What is your reconciliation method? Do you provide the tools to run mandate-to-deal reconciliation, or does that require a manual process on the buyer's side?
Where is the approval log held? Can the buyer retrieve it independently of the vendor's reporting interface?
What happens to deal records if the platform ceases to operate? Is there an export mechanism or an independent escrow?
These questions parallel the ones that buyers have been asking of ad servers and verification vendors for years. The difference is that agent-negotiated deals are a newer category, and some platforms are selling the capability before they have built the audit infrastructure to support it.
What does a failure to produce an audit trail mean for liability?
In traditional programmatic, disputes about delivery are common and the resolution mechanisms are reasonably well established: the MRC provides a framework, TAG certification reduces fraud exposure, and most standard agency contracts include a dispute resolution clause. In agent-negotiated deals, the liability question is less settled.
If an agent negotiates a deal that falls outside its mandate, whether because the mandate was ambiguous or because the agent failed to escalate correctly, the question of who is liable for the resulting expenditure is not resolved by existing ad industry norms. The IAB Tech Lab's Agent Communication Protocol guidelines, published in early 2026, note that deal-level liability depends on the existence of a complete mandate record and a corresponding approval log. Without both, the buyer has limited recourse.
The practical implication is that buyers who cannot retrieve a mandate record and an approval log for every agent-negotiated deal are carrying undocumented liability. That is not a risk associated with agents specifically; it is a risk associated with any deal where the audit trail is incomplete. Agents make it easier to generate many transactions quickly, which makes the risk larger in aggregate if the audit infrastructure is not in place before the campaign runs.
A practical audit checklist
Before the deal runs: confirm the mandate is documented, approved, and accessible. Confirm the platform can produce a deal record on request and that the record is immutable after execution.
During the deal: monitor for escalations. Any human approval or override should be logged at the time it occurs, not reconstructed afterwards.
After the deal: run mandate-to-deal reconciliation, then deal-to-delivery reconciliation, then financial reconciliation. Log every variance above tolerance and query it formally.
If any step produces a gap in documentation, do not accept the next deal from the same platform until the gap is closed. A single undocumented deal is an operational risk. A pattern of undocumented deals is a structural one.
The questions above are not difficult to ask. They should not be difficult to answer. The standard buyers should hold themselves to is the same one they would apply to any other part of the supply chain: if you cannot audit it, you cannot trust it.