AI feedback isn't real validation when it comes from a model role-playing your customer — it can only tell you what sounds plausible, not what a real person would do or what a real AI agent will actually recommend.
The synthetic user problem
An AI persona has no real preferences — it predicts a plausible-sounding response, and plausible usually means agreeable. Ask a role-played "ideal customer" if they'd buy your product and it will typically say something encouraging, because agreement is the statistically likely completion, not because anyone real is going to be disappointed later. No one carries any consequence for that answer being wrong.
Two honest uses of AI in validation
Real, identity-verified people give you real stakes and real disagreement — a Human panel. Real AI agents, tested on the actual decision task they'll face in the wild — an Agentic panel — is a different and legitimate use of AI: not simulating your customer, but testing whether the assistant your customer already consults would recommend you.
The tell
If "feedback" always agrees with your framing, it isn't feedback. Both a Human panel and an Agentic panel are built to produce disagreement, objections, and sometimes a flat no — because that's the entire value of running the test.
If your validation never says no, you haven't validated anything.