Article · Operations
Using AI safely in a licensed trade
A working boundary for AI in a business where some decisions legally require a licensed person — what to use it for, what to keep it away from, and what to put in writing before anyone starts.
In short
AI is safe in a licensed trade when it drafts and a qualified person decides. Use it for field notes, scope language, customer messages, procedure documentation and message triage. Keep it entirely out of pricing, code interpretation, permit determination and safety judgements. Put that boundary in a written policy, require human review before anything reaches a customer, and never enter customer payment details or personal data into a public AI tool.
A licensed contractor carries accountability that does not transfer. If a wrong answer goes out under your business name, “the software said so” is not an explanation. That constraint shapes every sensible use of AI in the trades, and it is the reason generic advice about AI productivity does not survive contact with a plumbing company.
The good news is that the constraint is workable, and the line falls in a fairly obvious place once you look at it directly.
Where the value actually is
The work AI absorbs well is written and repetitive, and there is more of it in a small trade business than most owners consciously account for.
| Task | Why AI suits it |
|---|---|
| Field notes into job records | Restructuring fragments into a consistent form |
| Scope and exclusions drafting | Applying the same structure to every job |
| Customer messages | One voice regardless of who is typing |
| Procedure documentation | Extracting what is in your head by questioning |
| Message triage | Sorting a backlog by stated urgency |
| Objection responses | Preparing consistent answers in advance |
What these share: a person supplies the facts, the model supplies the structure, and a person checks the result. None of them asks the model to know something only a licensed person knows.
The four hard boundaries
Pricing
AI does not set, estimate or imply a price. Your pricing depends on your costs, your market and your margin. A model has none of that and will still produce a confident number if asked, because producing plausible text is what it does. An invented price in a customer-facing document is a representation you may be held to.
Code compliance
AI does not state whether work is or is not code-compliant. Codes vary by jurisdiction, are amended locally, and are interpreted by inspectors who do not read your chat history. The correct output is a routing instruction: permit and code requirements to be confirmed by the company prior to scheduling.
Permit requirements
Same reasoning, same routing. Concealed conditions are the most common trigger for a permit scope change, which is exactly the moment nobody wants a guess.
Safety determinations
AI does not decide whether a condition is safe, and never tells a customer to operate, shut off, reset or bypass anything. This one is not negotiable in either direction: it protects the customer, and it protects you.
What must never be entered
Independent of the boundaries above, some information should not go into a public AI tool at all:
- Payment card numbers, bank details, any financial account information.
- Government identification numbers, dates of birth, medical information.
- Employee records, disciplinary records, pay information.
- Confidential business information or anything under a non-disclosure agreement.
- Photographs showing a customer’s home interior, documents or belongings.
- Passwords, access codes, alarm codes, key box combinations.
Where a customer’s name or address is genuinely needed, use the minimum. Most drafting tasks work perfectly well with the customer described as “the customer”.
Also worth checking once, before your first real job note goes in: most tools have a setting controlling whether your conversations are used to improve their models. For business use, turn it off.
The review step
The most common failure is not bad output. It is staff pasting output without reading it, because it looks finished.
A one-minute review before anything is sent catches nearly all of it:
- No price appears unless an authorised person put it there.
- No statement that anything is code-compliant, permitted, safe or unsafe.
- No instruction telling a customer to operate or shut off anything.
- No promise of an outcome, lifespan, or that a problem will not recur.
- Nothing appears that was not in the source material.
- Warranty terms quoted match the written warranty word for word.
That last one deserves emphasis. Warranty language is the highest-risk sentence you send, and paraphrase has a way of becoming more generous than intended.
Putting it in writing
A written policy is worth more than an informal habit, for two reasons. It gives staff a standard to point at instead of guessing, and it gives you something to review when something goes wrong.
It does not need to be long. What it needs to contain: what AI may be used for, the four prohibited decisions, the information that must never be entered, the mandatory review step, and the conditions under which someone stops and escalates rather than proceeding.
Then have people read and sign it — not as bureaucracy, but because the act of signing is what makes staff read it once properly.
Where this leaves you
AI in a licensed trade is neither a transformation nor a threat. It is a drafting tool that removes a specific, well-defined category of work: the writing you currently do after dinner.
The businesses that get value from it are the ones that decided in advance what it is not allowed to do, and then held that line when it would have been convenient not to.
Common questions
- Do I need a paid AI subscription?
- No. Free tiers handle the drafting tasks described here. Paid tiers generally produce better output on longer tasks like reconstructing a field note or extracting a procedure by interview.
- Should I tell customers AI was involved?
- There is no general obligation to disclose that a draft was AI-assisted, in the same way you would not disclose which word processor you used. What matters is that a qualified person reviewed and approved what was sent, and that it is accurate.
- What is the single most common mistake?
- Pasting output without reading it. Polished text reads as verified text. It is not.