What Evidence You Need if a Customer Alleges Your AI Agent Discriminated Against Them

Quick Answer

If a customer alleges your AI agent discriminated against them based on a protected characteristic, you need four categories of evidence ready before the claim arrives: the exact conversation transcript and the inputs the agent actually received, the version of the prompt and policy logic in effect at that moment, any structured decision or eligibility logic the agent applied, and evidence of your broader bias-testing practice, not just this one conversation. This is a higher-stakes version of general evidentiary record-keeping for agent decisions, because a discrimination claim can carry legal exposure that an ordinary customer dispute does not.

Why This Is a Distinct Category of Risk

Most AI agent disputes are about whether an answer was correct. A discrimination allegation is about whether the agent's behavior differed based on who the customer was, which is a legal question, not just a quality question. Proving a negative (the agent did not discriminate) is much harder after the fact than establishing the practice to prove it before the fact. Companies that only keep general logs discover, when a real claim arrives, that they cannot reconstruct what actually happened with enough precision to mount a defense.

The Four Records to Keep

Full transcript and structured inputs. Not just the conversation text, but any structured data the agent had access to at decision time: the customer's account tier, location, prior history, or any field that could plausibly correlate with a protected characteristic even if it was never intended to be used that way. If the agent's behavior can be explained entirely by a legitimate factor visible in this record, that is your strongest defense.

Prompt and policy version at the time. Prompts change. If you cannot prove which version of the system prompt and policy logic was active during that specific conversation, you cannot demonstrate what rules the agent was actually following, which matters both for legal liability analysis and for showing the behavior was not an intentional design choice.

Any structured decisioning logic. If the agent's response involved eligibility rules, pricing tiers, or escalation logic beyond free-text generation, you need the actual rule set applied, not a description of it. Structured logic is easier to defend or disprove than free-text reasoning, so preserve it in a form that can be re-run against the same inputs.

Evidence of your bias-testing practice. A single conversation is a data point; a demonstrated practice is a defense. This means retained records from your pre-launch red-teaming that specifically included protected-characteristic scenarios, plus any ongoing sampling that tests for disparate treatment across customer segments, ideally overlapping with work you are already doing to catch underperformance in specific segments.

What Legal and Compliance Should Confirm Before Launch

Before an agent goes live in any context where it makes decisions that could affect protected classes, whether it is pricing, eligibility, or service prioritization, get explicit confirmation as part of your legal and compliance sign-off that discrimination-specific logging and testing is part of the launch checklist, not an afterthought added after the first claim.

FAQ

Is this different from general customer dispute record-keeping? Yes. General dispute records need to show the agent acted reasonably. Discrimination claims need to show the agent's behavior did not vary based on a protected characteristic, which requires structured inputs and decisioning logic, not just a transcript.

How long should we retain these records? Match your retention period to the relevant statute of limitations in your jurisdiction for discrimination claims, which is often longer than your default support-ticket retention policy, and confirm the specific period with legal counsel.

Does anonymized testing data count as evidence of non-discrimination? It helps establish practice and diligence, but it does not replace the specific conversation record for an individual claim. You need both the general testing evidence and the specific transcript.