Quick answerWhen the company itself, not a single customer, needs to determine what actually happened between two of its own customers, such as a marketplace buyer and seller dispute, the AI agent's conversation and transaction records can be a legitimate input to that internal adjudication process, but the agent's own generated summary of events should never be treated as the determination. Route the underlying, unedited records to a human adjudicator with defined authority, let that person form the actual conclusion, and keep the agent's role limited to surfacing the raw data quickly and completely, not interpreting it.
Distinct from a customer asking the agent to vouch for them
When a customer asks the agent to characterize what happened in a dispute covers a customer-facing request: one customer, mid-conversation, asking the agent to take a position on their behalf. This is a different scenario entirely: the company's own operations or trust and safety function needs to rule on a dispute between two separate customers, like a marketplace transaction gone wrong, and is deciding what role the AI agent's data should play in reaching that ruling. Neither customer is asking the agent anything in this scenario, the company is using the agent's historical records as an internal input.
Also distinct from a live contradictory-answer bug
What to do when a buyer and a seller each get a different answer from the AI agent about the same transaction is a data or model consistency bug, the agent told two people conflicting things and that needs to be diagnosed and fixed. The adjudication question here assumes the agent behaved correctly throughout, it answered each party accurately based on what each said, and the company still needs to reconstruct the full picture across both conversations to rule on a dispute that has nothing to do with the agent's own accuracy.
Why the agent's own summary cannot be the ruling
An agent asked to synthesize a single, unified account of what happened across two separate customers' conversations is being asked to make interpretive choices, whose framing to trust when the two parties' statements conflict, which details matter, exactly the risk already identified in the customer-facing characterization scenario, just moved one step upstream into an internal process instead of a customer-facing one. The fix is the same in spirit: surface the actual, unedited transcripts and transaction data from both sides to the human adjudicator, formatted for fast review, rather than having the agent pre-digest them into a single narrative the human then just rubber-stamps.
What good internal tooling for this looks like
Build the retrieval tooling to pull both parties' full relevant history quickly, time-aligned so the adjudicator can see what each party knew and when, without asking the agent to resolve any ambiguity itself. Keep a record of what data the adjudicator reviewed and what they ruled, consistent with the discipline in evidentiary record-keeping for AI agent decisions, so the ruling itself can be defended later if the losing party disputes it further. If the dispute escalates to an actual legal or regulatory process, the same records should already be structured to support a subpoena or legal-hold disclosure process without a separate reconstruction effort.
FAQ
Can the agent flag which parts of two transcripts are relevant to speed up the adjudicator's review? Surfacing relevant excerpts is fine and useful, as long as it is framed as a retrieval aid, not an interpretation, and the adjudicator still has access to the full, unfiltered record rather than only the flagged excerpts.
Who should have authority to make the actual ruling in this kind of dispute? Whoever owns trust and safety, marketplace operations, or the equivalent function for the specific dispute type, with clear escalation to legal for disputes with financial or regulatory weight beyond routine cases.
Should both customers be told the AI agent's records were used in the adjudication? Yes, transparency about the basis for a ruling, including that conversation records were reviewed, is generally good practice and reduces the chance either party feels the process was opaque or unfair.