Where AI should never touch your service business: a no-fly list for owners

The useful question about AI in a service business is not where it can help. It can help almost everywhere: drafting, sorting, matching, summarizing, finding the invoice line the tech forgot. The useful question is where AI should never be the one that decides. Get that list right and you can turn on the rest with a clear conscience. Get it wrong and one bad answer can cost you a customer, a lawsuit or worse.

This post is that list. Seven decisions where AI should stay out of the driver's seat, why each one is on the list, what law or case backs it up, and what AI can still safely do in each area.

Three tests for the AI no-fly list

Before you let any AI tool act on its own, put each decision through three questions.

1. Could a mistake hurt someone physically? If yes, AI can collect facts, but the decision follows fixed rules and goes to a trained person.
2. Would it commit the company to money or a promise? Prices, refunds, credits, guarantees and contract terms. If yes, AI drafts and a person approves.
3. Does it judge a person? Hiring, firing, discipline, pay, performance ratings. If yes, AI can summarize data, but a manager decides and owns the reasoning.

These tests line up with the National Institute of Standards and Technology's AI Risk Management Framework, a voluntary framework released in January 2023 whose first function, Govern, is about deciding who is accountable for an AI system's outcomes.¹ NIST's 2024 profile for generative AI names two risks that matter most here: confabulation, meaning confidently stated but false content, and automation bias, people's tendency to defer to a machine's answer.² The no-fly list exists because both are guaranteed to happen eventually.

The seven no-fly zones at a glance

1. Safety triage: gas, carbon monoxide, water and sparks

The Consumer Product Safety Commission says more than 200 people in the U.S. die every year from accidental non-fire carbon monoxide poisoning associated with consumer products.³ Your phones will eventually take a call from one of those houses. The caller may not say "carbon monoxide." They may say "the alarm thing keeps beeping and my head hurts."

A language model is good at understanding that sentence most of the time. "Most of the time" is the problem. The right design is boring on purpose: a fixed list of trigger words and phrases checked by plain code before the AI ever answers, a fixed safety script (leave the building, call 911 or the gas utility), and an immediate transfer to a person. The model can help the dispatcher afterward by summarizing the call. It should never be the thing deciding whether a gas smell is urgent.

What AI can do safely here: transcribe, summarize, and fill in the job details after a human has handled the safety part.

2. Prices, policies and promises

In February 2024, British Columbia's Civil Resolution Tribunal ruled against Air Canada after its website chatbot told a grieving customer he could apply for a bereavement fare refund within 90 days of buying his ticket, which contradicted the airline's actual policy. Air Canada argued, in effect, that the chatbot was responsible for its own actions. The tribunal called that "a remarkable submission," found negligent misrepresentation, and ordered the airline to pay $650.88 in damages plus interest and fees.⁴ It is a small-claims decision from Canada, not U.S. law, but the reasoning travels: the chatbot was part of the company's website, and the company was responsible for what it said.

Apply that to a trade business. If your AI texts "we can do that for $89" or "that's covered under warranty," you own the sentence. So the rule is simple:

• AI may repeat prices you have published, such as from-prices on an online booking page.
• AI may draft an estimate from your price book for a tech or comfort advisor to review.
• AI may not create a price, a discount, a warranty term, a payment plan or a policy exception in a message to a customer.

The same applies to what you say about your own AI. The Federal Trade Commission's 2024 "Operation AI Comply" sweep included a company marketing itself as "the world's first robot lawyer" that, the FTC alleged, never tested whether its output matched a human lawyer's. The proposed settlement included a $193,000 payment.⁵ If your website says your AI "never makes mistakes" or "guarantees" anything, you have a claim to substantiate.

3. Money movement: refunds, credits, write-offs and charges

Any action that moves money out of your business, or charges a customer, should require a person to press the button. This isn't because AI does math badly, although language models are unreliable at arithmetic. It is because money mistakes are expensive to unwind and the customer always notices.

AI can do the preparation well: find the duplicate charge, calculate the prorated membership refund, flag an invoice that doesn't match the job notes, draft the credit memo. A person approves, and the system records who approved what and when.

4. Collections and payment reminders

Collections combine two risks: tone and consent. A reminder that sounds threatening, goes out at 10 p.m., or keeps coming after the customer asked you to stop is how a $400 balance becomes a complaint.

The consent part has teeth. FCC rules let a person revoke consent to calls and texts "by using any reasonable method," and say that if a reply uses words other than STOP or similar keywords, the sender must still treat it as a revocation "if a reasonable person would understand those words" that way. Requests must be honored within ten business days at most.⁶ "Please quit texting me about this" is an opt-out, even though it doesn't say STOP. A good AI can recognize that sentence, which is a real help. It should not be the one deciding whether to send the next reminder.

The safe pattern: AI drafts a polite reminder per overdue invoice and sorts the list by priority; a person approves each send; opt-outs are enforced by code, not by the model's judgment.

5. Hiring, firing, discipline and pay

This is the zone with the most regulation and the most risk. New York City's Local Law 144, enforced since July 5, 2023, bars employers from using an automated employment decision tool unless it has had a bias audit within the past year, requires a public summary of the audit, and requires notice to candidates.⁷ Other states and cities have passed or proposed similar rules. Even where no law applies yet, you need to be able to explain why you let someone go or turned an applicant down. "The software ranked them low" is not an explanation.

What AI can do: summarize a tech's callbacks, first-time fix rate and customer ratings for a review conversation; draft a job posting; organize applications. What it shouldn't do: rank candidates, reject applicants, recommend discipline, or set commission and bonus amounts. Pay rules for technicians are complicated enough without a model in the loop.

6. Safety findings on equipment

A cracked heat exchanger, a failed flame rollout switch, an undersized breaker: these are judgments a licensed technician makes and signs. AI is useful in turning a tech's spoken notes into a structured finding with a severity, and in flagging that a note mentions something worth a closer look. It should not declare equipment safe or unsafe, red-tag a furnace, or tell a homeowner it's fine to keep running something.

The test: if the finding would end up in a liability claim, the name on it should belong to a person who was in the room.

7. Reviews and testimonials

The FTC's rule on consumer reviews and testimonials, in effect since October 2024, makes it a violation for a business to write or create a review that misrepresents that the reviewer exists, used the service, or had the experience described.⁸ An AI-written "review" posted under a customer's name fits that description. So does asking your office staff to post reviews without disclosing they work for you.

The FTC's case against one AI review-writing tool was later set aside in December 2025, with the Commission saying the original complaint didn't adequately allege a violation by the tool maker.⁹ That changed nothing for the business that posts the fake review. The rule still applies to you.

What AI can do: draft the review request text, time it after a completed job, and route unhappy customers to a manager before anything goes public. What it shouldn't do: write reviews, put words in a customer's mouth, or reply to a complaint without a person reading it first.

What AI is good for in every one of these zones

None of this means "keep AI away from your business." In every zone above, AI can take on most of the work:

The common thread is evidence. Every AI suggestion should show why it was made: the quote from the notes, the overdue balance, the trigger phrase. A person can check evidence in seconds. A person cannot check a black box.

How to write your own AI policy in one page

You don't need a consultant. Put four things on one page and share it with the team.

1. 1

   List the tools

   Every AI feature you use, including the ones built into your phone system, CRM and accounting software. Many shops are surprised how many there are.

2. 2

   Mark the no-fly zones

   Copy the seven zones above, add any specific to your trade, and write who approves in each one: owner, service manager, office manager.

3. 3

   Set the defaults

   New AI features start in suggest-only mode. Moving anything to automatic needs the owner's sign-off and a reason.

4. 4

   Review monthly

   Look at what the AI proposed, what was approved and what was rejected. Read five customer-facing messages it drafted. Adjust.

How TradesWork draws these lines

We built TradesWork around this list. Customer texts from the collections helper always wait for a person, even when other helpers run on their own, and the AI receptionist (in early access) checks gas, CO, flooding and smoke triggers in code before the model speaks, then transfers the caller to your on-call person. The full list of what each helper may and may not do is on the AI overview and the AI office help page.

Whatever software you use, ask the vendor to show you where each of these seven lines is drawn. If they can't, our post on how to judge any AI tool you're sold has the questions to ask.