What is an “audit-defensible” AI answer?
In short.
“Defensible” here renders the French legal term “opposable”: able to be presented to a third party with probative force. An audit-defensible AI answer is therefore one that an auditor, a client or a judge can reconstruct and verify without taking your word for it. Five attributes define it: an identified source (document, page, version), a timestamp, an explicit confidence score, a named human approver, and the replayability of the processing. Without them, an answer remains a claim, however well written.
Where the word comes from, and why it matters
“Opposable” (defensible) comes from law: able to be presented to a third party with probative force. Applied to document AI, it shifts the question from “is the answer right?” to “can I demonstrate that it is right, and how it was arrived at?”. It is decisive in three situations: audit (the certifier asks for evidence of every statement), disputes (the client contests a claim in your bid), internal review (management wants to understand how a commitment was made).
The five attributes in practice
The source: not “our internal documents” but “Security policy v4.2, page 12, in force”; the version matters as much as the document. The timestamp: a company’s truth has a date. The score: the system qualifies its own confidence, which organises the review and documents due diligence. The approver: a name, a role, a date; human oversight becomes a fact. Replayability: reprocessing the same file and explaining any difference, which requires controlled versions and complete logs. Quick checklist: take an answer sent last quarter and try to produce the five attributes in less than an hour. The result is your real level of defensibility.
The Optivalue.ai approach
The five attributes are Optivalue.ai’s native format: every answer comes out sourced, dated, scored, approved and logged. Defensibility is not extra work, it is the standard output.
Is a correct but unsourced answer defensible?
No: without a source, timestamp and approver, it remains a claim a third party cannot verify.
Why does the version of the source document matter?
Citing a superseded policy is an error that looks flawless, the worst kind.
On the same topic