An LLM, or large language model, is the kind of AI behind the chat assistants everyone has tried by now. It reads text and writes text. Owners keep asking the same thing: could one answer the phone, write the estimates, send the invoices and run the shop while I'm on a roof?
The honest answer is no, not the whole shop. It can take a real share of the office work off you. The line between those two answers is where owners get burned, so this guide draws it plainly and gives you four questions to ask before you trust any AI tool with your customers or your books.
A test of an AI running a real (tiny) business
In 2025, Anthropic, the company that makes Claude, ran a test it called Project Vend. It put Claude Sonnet 3.7 in charge of a small automated shop in its San Francisco office for about a month. The AI picked what to stock, set prices, emailed wholesalers and chatted with customers.
It did some things well. It found suppliers by searching the web and it turned down requests for sketchy items. It also did things no owner would accept. It priced some items without checking what they cost, so it sold some below cost. For a time it gave customers a payment account it had made up. It gave out discount codes when people asked nicely, including after an employee pointed out that nearly all its customers were co-workers. Anthropic's write-up says the shop did not make money.
Two things to keep in mind. This was one experiment on an earlier model, and the people who ran it are the ones who published it. Models get better. But the failure pattern matters more than the model name: an AI that is agreeable, confident and not tied to hard numbers will give away your margin.
A second, independent test points the same way. Researchers built a pretend software company and asked AI agents to do its everyday jobs. The best agent finished 30% of the tasks on its own. That was a software company, not a plumbing shop, and the paper has been updated since it first came out. Still, "does most of a whole job unattended" is not where the evidence is today.
What it does reliably today
Look at the pattern in what works. These are jobs where the input is words, the output is words, and a person can check the result in seconds.
Answering and sorting. A customer texts "my furnace is making a clicking noise." An AI can reply politely, ask the right follow-up questions (is there heat, is there a smell, is it the only unit), and put the answers on a request for your dispatcher. This matters because a lot of calls never reach a person. Invoca's July 2026 report says that across home services, 52% of inbound calls are answered by a person. That figure comes from Invoca's own call data, and vendors count "answered" differently, but it shows how many calls go unanswered.
Summarizing. Turn a tech's rambling voice note or a long thread of texts into three clear lines on the job. This is one of the safest uses, because the source is right there to check against.
Drafting. First drafts of an estimate description, a follow-up text, a review request or a "we're on the way" message. A draft saves typing. A person still reads it and sends it.
Follow-up and reminders. "This estimate was sent nine days ago and nobody has answered." An AI is good at noticing that kind of gap and preparing the nudge.
Answering questions about your own business. "Which jobs from last month are still unpaid?" An AI can turn that plain question into a lookup of your records and put the result in a table. The key is that the answer comes from your records, not from the AI's memory. More on that below.
What it can't do, or shouldn't do alone
It can make things up. NIST, the US federal standards agency, calls this confabulation: confidently stated content that is false. It says confabulations are a natural result of how these models are built, since they predict likely words rather than look up facts. A made-up part number, a wrong warranty term or an invented price sounds exactly as sure as a correct one. That is why "unknown" has to be an allowed answer, and why Project Vend's invented payment account is a warning, not a curiosity.
Don't let it do your math. Totals, tax, margins, payroll and what a customer owes should be calculated by software that is built to add, from numbers in your records. The AI's job is to ask for the number and show it, not to produce it from its head. If a tool shows a figure it can't trace to a record, don't trust it.
It shouldn't move money alone. Refunds, charging a card, writing off a balance, changing a price, paying a vendor. These are the actions that are hard to take back and easy to get wrong. A person confirms them, every time.
It shouldn't message customers without your say-so. There are two reasons.
First, consent. In February 2024 the FCC ruled that calls using AI-generated human voices fall under the Telephone Consumer Protection Act, which means they need the called party's prior express consent, unless there is an emergency or an exemption. That ruling is about voice calls. Texting has its own consent rules, and they are strict enough that you should check your setup with your phone provider and a lawyer. Either way, keep proof of the customer's consent on the customer record.
Second, you own what it says. In 2024 a British Columbia tribunal ruled against Air Canada after its website chatbot gave a customer wrong information about a refund. The airline argued the chatbot was a separate entity responsible for its own actions. The tribunal rejected that, and noted that the airline is responsible for all the information on its website. That was a Canadian small-claims decision, not US law. But a customer holding a text from your company that says "yes, that's covered" will not care which software typed it.
It can't replace licensed judgment. Whether a gas line is safe, whether a panel needs replacing, whether a repair meets code, what a permit requires. Those are decisions for a licensed person who can see the equipment, and your license is on the line. An AI can pull your notes, past jobs and a checklist. It should not sign off.
It's too agreeable. Project Vend showed an AI talked into discounts again and again. A customer who pushes, or a tech who types a wrong number in a hurry, can pull a polite AI along with them. Rules about price, discounts and approvals belong in the system, not in the AI's good manners.
Why grounding in your real records matters
"Grounded" is the whole difference between a toy and a tool. A general chatbot answers from what it absorbed from the internet. A grounded one answers from your price book, your customers, your jobs, your invoices and your calls, and it points to the record it used.
For a trades business this is not a nicety. Your business is a chain: a call becomes a job, a job becomes an estimate, an estimate becomes work, work becomes an invoice, an invoice becomes money. If the AI sits outside that chain, it has nothing true to say about it. If it sits inside it, it can say "the Hendersons' invoice for the water heater is 12 days old, unpaid, and here it is," and you can click straight to the invoice.
Grounding also gives the AI permission to say "I can't find that." In a business, a clear "I don't know" is worth more than a smooth guess.
How to judge any AI tool: four questions
Use these on any product, ours included.
- Does it show its sources? Every answer that involves a number, a date or a customer should link to the job, invoice, call or note it came from. If you can't click through to the record, treat the answer as a guess.
- Does it ask before it acts on a customer or on money? Texts, emails, estimates sent to a customer, charges, refunds and price changes should wait for a person to approve. Ask the vendor to show you the exact screen where the approval happens, and what it says.
- Can you undo it? For small internal changes, like moving a job or editing a note, you should get an undo button, not a pile of "are you sure?" boxes. For things that can't be undone, see question 2.
- Is there a trail? You should be able to see what the AI did, when, on whose request, and what it was looking at. When a customer says "your system told me something different," you need to find out what happened in minutes.
A fifth check is cheap: ask it something you already know the answer to, and something it should not know. A good tool gets the first right and says it doesn't know the second.
A sensible way to start
Begin where a mistake costs the least and checking is fastest.
- Start with summaries and drafts. Have it summarize jobs and draft follow-ups, and read everything before it goes out.
- Move to questions about your own records, and open a few of the records behind the answers each week to see that they match.
- Let it prepare customer messages, but keep a person on the Send button until you have watched it for a month or more.
- Keep money, price changes and anything with a license attached with a person. Move them only if you can show the tool's approval and audit steps work.
Watch your own numbers, not the vendor's. Count how many calls get answered, how fast estimates get a follow-up, and how many drafts you had to fix. If the draft fixes aren't dropping, the tool isn't saving you time.
Where Trades Operator fits
Trades Operator, which publishes this guide, builds Operator, an assistant that works inside the same record as your calls, jobs, invoices and payments. Our own website describes it as answering from your jobs, calls and invoices, drafting an estimate from your notes, and suggesting next steps and then waiting for you. Judge it by the four questions above, like anything else.
FAQ
Can an AI really run my HVAC, plumbing or electrical company?
Not by itself, and the evidence says so. In Anthropic's own Project Vend test, an AI that ran a small office shop for about a month priced items below cost and did not make money. What it can do today is take on well-defined office work (answering, summarizing, drafting, reminding) while you keep the decisions. Think of it as a fast assistant, not a manager.
Will an AI make up prices or numbers?
It can. NIST calls this confabulation: confidently stated content that is false. The fix is to never let the AI do the arithmetic or recall the price from memory. A good tool pulls the number from your price book, your invoices or your jobs, shows you where it came from, and says 'I don't know' when the record is missing.
Do I need a customer's permission before an AI texts or calls them?
Yes, treat consent as required. The FCC ruled in February 2024 that calls using AI-generated human voices fall under the Telephone Consumer Protection Act and need the called party's prior express consent unless an emergency or exemption applies. Texting has its own consent rules, so ask your phone provider and your lawyer, and keep the consent on the customer's record.
Who is responsible if the AI tells a customer something wrong?
You are. In Moffatt v. Air Canada (2024), a British Columbia tribunal rejected the airline's argument that its chatbot was a separate entity and held the company responsible for the information on its website. That was a small-claims decision in Canada, not a US ruling, but it shows how a tribunal looked at it. Have the AI answer from your own records and send customer-facing messages only after a person approves.
How do I tell a trustworthy AI tool from a risky one?
Ask four things. Does it show the record behind each answer? Does it ask before it messages a customer or touches money? Can you undo what it did? Is there a log of who, or what, did what and when? If a vendor cannot show you all four, keep the tool away from customers and your books.