The AI-Drafted Customer Message Review SOP

Purpose

This procedure guarantees that no customer message drafted with an AI tool leaves the shop without a person checking the claims in it, at a depth matched to what the message can cost if it is wrong. Without it, review drifts to one of two failures: everything gets a glance, so the warranty promise and the safety instruction get the same five seconds as the appointment reminder; or everything gets a full review, so staff stop using the tool or start skipping review to keep up. A tool that sounds equally confident about every sentence it writes needs a review that is not equally deep on every message.

Scope

Covers every outbound customer message (email, text, portal message, letter) where any part of the text was produced by an AI tool, including a feature inside the field software. It does not cover:

• What may be pasted into a tool in the first place. That is the shop's AI use policy (see related: The Shop AI Use Policy SOP).
• Automated messages sent by a rule with fixed, pre-approved wording, such as a reminder template. Changes to those go through the automation change-control procedure (see related: The Automation Change-Control SOP).
• Live phone or chat conversations handled by an AI answering service. Those are reviewed after the fact under the weekly call review (see related: The Weekly AI Call Review SOP).

Roles and responsibilities

The three tiers

A message takes the highest tier any sentence in it reaches. One price in an otherwise routine reschedule makes it tier 2.

Procedure

1. Assign the tier before reading the draft for polish. Read the draft once looking only for tier triggers: numbers tied to money, words like "includes," "covered," "warranty," "guarantee," "free," "safe," "do not use," "turn off," "refund," "damage." Assign the highest tier triggered and write it at the top of the draft. Acceptance: a tier number is recorded before any edits are made. Wrong looks like a tier assigned after editing, because by then the drafter has read for flow and the triggers blend in. Stop rule: if the drafter cannot decide between two tiers, take the higher one. Hazard: none physical at this step, it is a desk decision, but a safety instruction misread as tier 1 here is how one reaches a customer unreviewed, which is why the default is up.

2. Strip invented content before checking facts. Delete every sentence the drafter did not ask for and cannot source: added apologies, added promises ("we will make sure this never happens again"), added policy statements, added sign-off offers ("feel free to call anytime for a free follow-up"). Acceptance: every remaining sentence either carries a fact with a known source or is plain courtesy with no commitment in it. Wrong looks like a draft that grew by a third during generation; that growth is where invention lives. Stop rule: any deleted sentence that the drafter thinks the customer should hear goes back in only after it is sourced. Hazard: none physical, desk step.

3. Tier 1 self-check: every fact against the job record. For a routine message, the drafter points at each fact (customer name, date, arrival window, tech name, address, job type) and confirms it against the job record in the field software, not against memory or the tool's draft. Acceptance: every fact matches the job record exactly, and the count of facts checked is noted at the top. Wrong looks like a date in the draft that matches what the customer asked for in their message rather than what was booked. Stop rule: one mismatch means the drafter fixes it and re-checks every fact in the message, not just the one that was wrong, because a tool that got one fact from the wrong source often got another. Send when clean. Hazard: none physical, desk step.

4. Tier 2: the drafter prepares, the reviewer checks line by line. The drafter completes steps 1 to 3, then attaches the sources: the estimate or invoice, the warranty sheet, the job record. The reviewer reads each sentence with a money or scope claim and checks it against those attachments. Acceptance: every price and total matches the estimate or invoice in the software exactly; every "includes" matches the line items; every warranty or guarantee term matches the shop's current warranty sheet word for word in substance. Wrong looks like a total the tool restated (its arithmetic is not checked arithmetic), or a warranty length that is typical for the trade but not yours. Stop rule: any mismatch goes back to the drafter with the line marked; the reviewer does not fix it and send, because the drafter is the one who learns from the correction. A mismatch involving what the customer was already told goes to the owner. Hazard: none physical, desk step.

5. Tier 2: the reviewer reads for soft promises. Separately from facts, the reviewer reads for commitments the shop cannot guarantee: "definitely," "guaranteed to fix," "first visit," "same day," "will not happen again," "no extra charge." Acceptance: each such phrase is removed or replaced with what the shop actually commits to ("we expect," "the estimate covers," "if more is needed we will call you before doing it"). Wrong looks like a draft that reads warmer and more certain than the shop speaks. Stop rule: if the customer has been promised something like this already in writing, the reviewer flags it to the owner before the new message goes out, so the shop does not compound it. Hazard: none physical, desk step.

6. Tier 3: the owner or designated lead writes or approves the message. For a safety instruction, the wording comes from the shop's own written safety language for that hazard (for a smell of gas: leave the house now, do not touch light switches or use a phone inside, call the gas utility or 911 from outside; for a carbon monoxide alarm: get everyone outside into fresh air and call 911), not from the tool's paraphrase. For a dispute, refund or damage claim, the message states facts and next steps only, makes no admission of fault and no offer the owner has not decided on. Acceptance: the owner has read every sentence and initialled the draft, and any safety instruction matches the shop's written wording with no softening. Wrong looks like "you may want to avoid using the furnace until we can take a look," which a customer reads as optional. Stop rule: if the owner is unreachable and the message is a safety instruction, the drafter does not wait for the email; they phone the customer, read the shop's written safety wording aloud, and log the call, then the written message follows when approved. Hazard: this is the step where the hazard reaches someone outside the shop. A softened shut-off instruction leaves a customer running equipment that can injure them, and the control is the exact written wording, delivered by phone first when the message is urgent.

7. Send, and file the draft with its tier. The drafter sends from the shop's system so the message attaches to the customer's record. Acceptance: the sent message is visible in the customer record with its tier and reviewer noted. Wrong looks like a message sent from a personal phone or a personal email account, which leaves no shop record. Stop rule: if a message went out from outside the system, the drafter copies it into the customer record the same day with a note. Hazard: none physical, desk step.

8. Log every correction a reviewer made. At send time, the drafter records in one line what the reviewer changed and why ("tool stated 2-year labor warranty, ours is 1-year"). Acceptance: every tier 2 or tier 3 message with a correction has a log line. Wrong looks like an empty log over a month in which reviewers were visibly fixing drafts. Stop rule: none; this step is the shop's evidence base. Hazard: none physical, desk step.

The record this produces

Each AI-drafted customer message carries, in the customer record: its tier, the drafter, the reviewer (tier 2) or owner (tier 3), the date sent, and any correction. The correction log is read once a month by the office lead, looking for repeats. A correction that appears three times in a month (the same wrong warranty term, the same invented "free" offer) becomes a fix upstream: a line added to the shop's prompt library (see related: Building a Shop Prompt Library the Office Will Actually Use) or a phrase added to the step 1 trigger list. In a dispute, the record shows what was sent, who checked it, and what it was checked against.

Worked pass

One Tuesday morning, one drafter, three messages.

Message A, reschedule. Step 1: no triggers; tier 1. Step 2: deleted one added sentence ("we apologize for any inconvenience this may cause"), kept the plain apology the drafter wanted. Step 3: five facts checked (name, new date, window 1 to 3, tech, address), five of five match the job record. Sent at 8:40. Pass.

Message B, estimate follow-up. Step 1: triggers "includes" and a total; tier 2. Step 2: deleted "and as always, there is no charge for the follow-up visit," which the shop does not offer. Step 3: four facts checked (name, address, estimate number, tech), four of four match. Step 4: reviewer checks. The draft says the estimate includes "removal and disposal of the old unit"; the estimate's line items include removal and disposal, match. The draft restates the labor as "about 6 hours"; the estimate shows 5.5 hours. Step 4 fails. The reviewer marks the line and returns it. The drafter finds the tool had rounded from the customer's own earlier message, which said "you mentioned around six hours." Corrected to the estimate's 5.5 hours, re-checked, the reviewer approves. Step 5: "we will definitely be able to start Monday" becomes "we have Monday held for you." Sent at 9:25. Step 8 log: "Tool restated labor from customer's message, 6 hours; estimate says 5.5."

Message C, reply to a complaint about a dripping relief valve after an install. Step 1: triggers "complaint" and a safety-adjacent instruction the draft added ("it's safe to keep using it until we come out"); tier 3. Step 2: that sentence is unsourced; deleted. Step 6: the owner reviews. A dripping relief valve can mean a failing valve or excess pressure. The owner rewrites the instruction in the shop's own wording: do not plug or cap the relief valve or its discharge pipe, keep the area clear, and if it begins discharging steadily or the tank makes rumbling or popping sounds, shut off the gas or power to the heater and call. The dispute side states only that a tech is booked for Wednesday and makes no comment on cause. Owner initials at 10:05. Sent from the shop system with tier 3 and the owner noted. Step 8 log: "Tool added a keep-using-it reassurance; replaced with shop wording."

Three messages, one step failure (B at step 4), corrected before sending. The deleted sentence in C is the reason tier 3 exists: the tool offered a safety reassurance nobody had made.

References

• See related: The Verify-Before-You-Send Rule for AI Output
• See related: The Shop AI Use Policy SOP
• See related: Training a New Office Hire to Use AI Tools Responsibly
• See related: Building a Shop Prompt Library the Office Will Actually Use
• See related: The Weekly AI Call Review SOP