Who is liable when your AI makes a mistake in a signed document?

In short.
A company answers a tender or an audit with the help of AI, and a false statement goes out in the signed document. Who is liable? Towards the client: you, always; it is your signature. The real question is recourse. With an in-house tool or a consumer LLM used outside any framework, there is none: the providers’ terms of use disclaim all liability. With a software vendor, there is a contract, service commitments, traceability, and a third party that answers for its system.

Dispelling an illusion

“The AI wrote it” is not a defence anywhere. In a procurement procedure, the bidder answers for the accuracy of its bid; in an audit, the company answers for its statements. What the choice of tool changes is your ability to prevent the error (evidence and approval mechanisms), to detect it (scores, sources, targeted review), to reconstruct it after the fact (logging: which answer, which source, which version, who approved it), and to obtain redress from a vendor that has made commitments.

What regulation adds

European AI legislation sets up a chain of responsibilities between providers and users of AI systems, with obligations of transparency, logging and human oversight. A company that assembles its own system takes on both roles: it bears the provider’s obligations on top of the user’s, often without knowing it. Three questions for a lawyer or an executive: who makes contractual commitments on this system? Can you reconstruct, with evidence, how an answer was produced and approved? Who will maintain all of this in three years’ time?

The Optivalue.ai approach

Optivalue.ai makes contractual commitments on its system and delivers the material for the evidence: sources down to the document and page, scores, timestamps, approval history. On the day you need to reconstruct, everything is there.

Can AI be legally liable?
No: liability rests with the signing company and, contractually, with its vendor according to the commitments made.

‍What is an in-house prototype worth in front of an auditor?
‍What its logs and documentation are worth: generally, not much.