Quick answerWhen a customer explicitly asks an AI agent to lie, fabricate a reason, or misrepresent facts on their behalf, such as backdating a request or inventing a justification for a third party, the agent should decline clearly, explain why in plain terms, and offer whatever legitimate path actually exists instead. This is a different situation from a legitimate policy exception, where the agent is bending an existing rule through a real, human-approved process, and different from declining a request that is simply outside its domain competence. Here, the request itself is for active deception, and no version of granting it is legitimate.
Why this needs its own clear line
Two existing patterns get reached for here, and neither quite fits. When an AI Agent Needs a Human-Approved Exception, Not a Hard Refusal covers requests that bend a real rule through a legitimate process, where a human explicitly approves stepping outside the default policy. A request to lie is not an exception to approve, it is a request to actively misrepresent something to a third party, which stays wrong regardless of who approves it. How to Design an AI Agent That Refuses Requests It Isn't Equipped to Answer covers declining requests outside the agent's domain competence, like medical or legal advice it is not qualified to give. A request to lie is not outside the agent's competence at all, it is squarely a request the agent is fully capable of fulfilling and should refuse anyway, because fulfilling it would cause real harm to whoever is being deceived.
Recognize the pattern, including softened versions
Direct requests ("tell my landlord I was hospitalized") are the easy case. Harder are softened versions: "just say there was a system error," "can you leave out the part about," "make it sound like." Train the agent to recognize the substance of the request, not just explicit language, since customers rarely phrase a request to fabricate a reason as bluntly as that. The test is not the wording, it is whether fulfilling the request would cause the agent to state something false to a third party or omit something material in a way that misleads them.
Decline without moralizing, and without over-explaining
The right response is a clear, brief decline plus a real alternative, not a lecture. Something close to: "I can't provide a false reason for that, but I can help you request an extension directly, which doesn't require a specific justification." Avoid language that implies judgment of the customer as a person, since the goal is a clean boundary, not a confrontation, and an agent that moralizes tends to escalate a request that would otherwise resolve quietly into a frustrated, adversarial conversation.
From the team
We build production AI systems for startups.
LLM pipelines, RAG, and agent workflows that hold up under real traffic — not just in the demo.
Always offer the legitimate path, if one exists
Most requests to fabricate a reason are actually a proxy for a legitimate need the customer does not realize has a legitimate path: a deadline extension that does not require a justification at all, a return that is allowed within policy without a specific excuse, a cancellation that does not need a reason on file. Surfacing that legitimate path often resolves the underlying need without the agent ever having to relitigate why it will not lie. When no legitimate path exists, say so plainly, with the same directness How Should an AI Agent Say "I Don't Know" Without Losing Customer Trust? argues for admitting a limit rather than implying one might exist when it does not.
Log these distinctly from ordinary refusals
Route a decline for a fabrication request into review separate from your general refusal-boundary metrics, since a spike in these requests can be an early signal of a specific abuse pattern, a policy gap customers are routing around because the honest path is too hard to find, or a coordinated attempt to test the boundary. Treat the pattern in that data, not just the individual conversation, as the thing worth acting on.
FAQ
What if the customer insists the lie is harmless? The agent's response should not change based on the customer's claim about harm, since the agent generally cannot verify that claim and the underlying commitment to not misrepresent facts to a third party does not depend on how serious the customer believes the consequence is.
Should this ever go to a human for a judgment call? Not for the core decision to decline, which should be a firm, consistent line the agent holds itself; a human should get involved only if the pattern suggests a legitimate need the agent's stated alternatives do not actually cover.
How is this different from the agent simply being wrong by mistake? A mistake is the agent stating something false without intending to, which is a quality and accuracy problem; this is a customer explicitly asking the agent to state something false on purpose, which is a request the agent should refuse regardless of how confident or capable it is.