Verifiable AI for insurance underwriting: proving a decision was fair

Insurance underwriting is one of the first places AI meets serious regulatory scrutiny. When a model approves, declines, or prices a policy, the decision can be challenged � by a customer who feels wronged, by a regulator checking for unfair discrimination, or by an internal auditor months later. In that moment, the underwriter needs to show exactly what the model decided, on what inputs, using which model version, and prove that the record was not adjusted after the fact. A log the company can edit is not evidence, and reconstructing a decision from a model that has since been retrained is close to impossible.

The fix is to capture each underwriting decision as a signed, self-verifying record at the moment it is made. You record a structured entry � a hash of the input features used, the decision and price, the exact model identifier, and a timestamp � sign it with a post-quantum signature (Dilithium-2, NIST FIPS 204), and chain it to the previous entry. Because these records may need to hold up for the multi-decade life of a policy, a classical signature is the wrong tool: it becomes forgeable once large quantum computers arrive. Tamper with any field and the signature fails; anyone with the signed history can verify order and integrity without trusting the insurer.

Be precise about what this establishes, because overclaiming here is dangerous. It proves the authenticity, ordering, and integrity of what the model decided and on which inputs � a strong foundation for a fair-lending or conduct audit. It does not prove the decision was unbiased or lawful; that is a question of the model and the features, which this record makes auditable rather than answering on its own. It is resistant to known classical and quantum attacks per NIST, not unbreakable. For a regulated underwriter, turning every AI decision into evidence you can hand to a regulator � instead of a black box you have to defend on trust � is quickly moving from nice-to-have to expected.