Draft-and-review: setting up AI customer comms without the horror stories

Every AI customer-service horror story shares one root cause: the AI answered from its imagination and nobody checked. The fix is a pattern, draft-and-review, where the AI prepares grounded, cited drafts and a human clicks send. Here is the full setup.

The pattern in one paragraph.

A customer message arrives. The system reads it, pulls the relevant facts from your knowledge base, and writes a reply in your voice, with citations back to the files it used. You read the draft, fix anything wrong (the fix goes back into the knowledge base), and send. The AI never sends anything. You are always the sender.

Setup order.

Voice first. Collect 10-20 real replies you were proud of. This is the difference between drafts you edit and drafts you send.
Compliance folder second. Marketplace messaging rules, regulated-industry rules, your own no-go topics. The system reads this folder before every draft, every time.
Templates for the repeats. Your top 10 questions get grounded template answers. For most businesses this covers 60-80% of volume on day one.
Live drafts last. Only after the grounding exists do you wire the inbox. Wiring the inbox first is how horror stories start.

The guardrails that make it boring.

No source, no claim. If the knowledge base has no answer, the draft says "I need to check on that" instead of improvising. That sentence is the system working.
Citations on every fact. Review means checking the source, not re-researching the answer. Seconds, not minutes.
Corrections are permanent. Every edit you make flows back into the files. Week 4 drafts are visibly better than week 1 drafts.

How you know it is working.

Three numbers: edit rate (share of drafts you send unchanged; should climb weekly), time per reply (from minutes to seconds), and corrections logged (should fall as the knowledge base converges).

If the edit rate is not climbing by week 3, the knowledge base is missing something the messages keep hitting; the fix is more extraction, not more AI.

Questions

Asked before reading this far.

Can AI reply to customers automatically without review?

It can, but for a small business it should not start there. Full automation is earned: after weeks of draft-and-review show a high unchanged-send rate on a mature knowledge base, the most routine template replies can graduate to auto-send with sampling. Regulated and marketplace messages should keep human review.

What is a compliance folder?

A folder in the knowledge base holding every rule that carries a penalty: marketplace messaging policies, industry regulations, and your own forbidden topics. The AI reads it before drafting any customer reply, so drafts are grounded in the rules as written, not the model's memory of them.

How fast do AI-drafted replies get good?

Voice-trained drafts are usable in week 1 and converge on send-ready over 3-4 weeks, because every correction becomes a permanent rule. The metric to watch is the share of drafts sent without edits.

Or skip the homework

This is what month one builds, done with you.

The knowledge base, the workflows, the tools wired, in four weeks, yours to keep. Priced by fit · Book a discovery call · See what changes in 30 seconds