What to check before AI sends a customer reply

Before you let AI send a customer reply, decide who owns the promises in that email. A friendly sentence can commit your business to a delivery date, a refund, or work you never agreed to do.

Our starting recommendation is simple: let AI prepare the reply, then give a named person the information needed to approve it. Sending can remain a human action while you learn which requests the system handles reliably.

Give the draft a source

Choose a narrow kind of message, such as a customer asking for an update on an existing job. Define the records the drafting tool may use and who maintains them. An approved job record and current customer policy give the reviewer something concrete to check.

Keep the source beside the draft. If the record has no confirmed completion date, the reply should leave that date open and flag the missing information. Fluent wording cannot settle an unanswered scheduling question.

Check the commitment before the wording

The reviewer should be able to answer these questions without searching through another inbox:

• Do the customer and job reference match the incoming request?
• Which record supports each date, price, or status in the draft?
• Does the message offer anything that needs separate approval?
• Is every attachment intended for this recipient?

Read for tone after those checks. A polished reply with the wrong delivery date still creates work for someone else.

Hold an uncertain reply

Illustrative example: a customer asks whether an installation will finish on Friday. The job record says “awaiting parts,” and an older email mentions Friday as a possibility. The draft should ask the coordinator to confirm the schedule before offering a date.

Write the hold rule into the process: conflicting dates, missing records, complaints, and requests outside the agreed scope go to a named reviewer. Give that person a place to record the answer so the next draft uses the corrected information.

Count the time spent reviewing

Try a small set of representative messages with sending disabled. Record drafting time, review time, corrections, and the cases that required a colleague. Compare the whole task with your current way of replying.

If checking a draft takes longer than writing the response, simplify the request type or improve its source records. Expand only when the reviewer can verify the result consistently. For everyday AI assistance, that review step belongs in the work from the beginning.