Can AI answer customer emails without making things up?

The three conditions, and why each is load-bearing.

1. Grounding. The AI drafts from a knowledge base: your product facts, policies, and compliance rules as written files. Without it, the model fills gaps from training data, which is where invented specs and imaginary policies come from.
2. Citations. Every claim in a draft points to its source file. This makes review fast (check the source, not the whole answer) and makes gaps visible: no source, no claim. A system that cannot cite is asking you to proofread hallucinations.
3. Review before send. A human approves every outbound message. Early on this catches errors; over time corrections flow back into the knowledge base and drafts converge on send-ready. The review step never disappears, it just gets fast.

The marketplace case, specifically.

Marketplace buyer messages are the sharpest version of this problem: the rules about what a seller may say are enforced by suspension. The fix is a compliance folder the system reads before every reply, kept current, so drafts are grounded in the rules as written rather than the model's memory of them.

What it does not do.

It does not remove you from the loop, and you should not want it to. The honest failure mode of draft-and-review is rubber-stamping: approving without reading.

If you will not budget minutes a day for review, full automation of customer comms is not ready for you yet, and anyone who sells it to you unreviewed is selling the incident too.