Where AI actually pays for itself in a service business (and where it doesn't)

Most of the business owners we talk to have already tried AI. Not in a pilot program—they opened ChatGPT one evening, asked it to do something real, and watched it confidently produce something wrong.

That experience is worth taking seriously, because it’s teaching the right lesson for the wrong reason. The tool didn’t fail because AI is useless. It failed because it was asked to do a job it had no information for, with nobody checking the result. That’s a setup problem, and it’s the same setup problem behind most of the disappointing AI projects we see.

So here’s our honest read on where this technology pays for itself in a 15-to-150-person service business right now, and where it doesn’t.

A three-question test

Before evaluating any specific tool, run the task through three questions. If you can’t answer yes to all three, the task isn’t ready.

Is it repetitive? Something that happens 20 times a week is worth automating. Something that happens twice a month is usually cheaper to keep doing by hand.

Is it text-shaped? These systems work on language and documents. Reading a PDF, summarizing a thread, extracting fields from a form, drafting a paragraph, classifying a request—all good fits. Deciding whether to take a job is not.

Is there a right answer someone can check? This is the important one. If a person can glance at the output in five seconds and know whether it’s correct, you can use AI safely. If verifying the answer takes as long as producing it did, you’ve saved nothing and added risk.

Where it pays off

Intake and screening. This is the strongest case in this industry, by a wide margin. Every service business takes in a stream of requests—referrals, applications, quote requests, inspection reports, insurance paperwork—and somebody spends their morning reading each one, pulling out the same eight fields, and deciding what happens next. That work is repetitive, text-shaped, and checkable. We’ve seen a review that took 45 minutes per file come down to about four, with a human still making the final call on every one.

Pulling structure out of documents. Scanned forms, faxed orders, field reports, lab results. Extracting the data into your system beats re-typing it, and the output is verifiable at a glance.

After-hours triage. Not a chatbot pretending to be your staff—a system that takes an overnight request, works out how urgent it is and what it’s about, and makes sure the right person sees it first thing. The customer still talks to a human. They just talk to the right one, sooner.

Internal search. Twelve years of proposals, job notes, and specs sitting in folders nobody can search. Asking a question and getting the answer with a link to the source document is genuinely useful, and if it’s wrong, the source is right there to check.

The repetitive parts of writing. First drafts of recurring reports, service descriptions, job posts, and routine correspondence. Not the final version—the draft your team edits instead of starting from a blank page.

Where it doesn’t pay off yet

Anything customer-facing without a human check. A chatbot that quotes a price, commits to a timeline, or gives technical advice will eventually get one wrong, and you’ll own the consequence. If you run one, scope it to answering questions you’d be comfortable printing on a flyer, and hand off to a person for everything else.

Pricing and estimating. These models are good at producing plausible numbers, which is the worst possible failure mode when the number is a bid. Use them to assemble the line items; keep the pricing with whoever is accountable for the margin.

Anything where being wrong is expensive and hard to notice. Compliance filings, safety documentation, legal language, anything with a regulator on the other end. The risk isn’t that it makes an obvious mistake. It’s that it makes a subtle one nobody catches for three months.

Replacing the person who knows how things actually work. The judgment your dispatcher or office manager applies is built on years of context that isn’t written down anywhere. AI can take the typing off their plate. It can’t take over their job, and projects that assume otherwise tend to end badly.

Why the evening experiment failed

Three things were missing, and they’re the same three things any working system needs.

Grounding. The model had no access to your pricing, your history, or your documents, so it filled the gaps with something reasonable-sounding. A system connected to your actual data answers from your data, and says so when it can’t find something.

A defined task. “Write our service tickets” is not a task. “Take this voicemail transcript, pull out the customer, address, equipment, and reported problem, and format it as a ticket” is. Narrow beats clever.

A review step. Someone has to see the output before it matters to a customer. In a well-built workflow that’s a 10-second check on a screen, not a second full pass.

What it costs, and how to think about it

Skip the monthly software fee as your unit of measure. The useful number is hours per week.

Take one workflow. Count how many times it runs in a week and how long each run takes. That gives you hours per week, which converts to a salary cost you can actually compare against a build. A process that eats 12 hours a week of a $60,000 employee’s time is roughly $17,000 a year of capacity—and it’s usually the capacity of your most experienced person, doing your least valuable work.

Build costs vary widely depending on how many systems have to talk to each other. The thing to insist on is that the payback is stated in your terms, not in the vendor’s. If nobody can tell you how many hours a week it saves and how you’d confirm that, it isn’t a business case yet.

Where to start

Pick the one process that everybody in your office complains about. Time it honestly for a week. Run it through the three questions. If it passes all three, that’s your candidate—and it’s almost always intake. If you want to go a step further before building anything, our five-question readiness audit checks whether the process itself is ready to automate.

Start there, keep a person in the loop, and measure the result in hours, not in adjectives.

If you want a second opinion on which of your workflows are actually good candidates, that’s a conversation we’re happy to have. It usually takes about an hour and you’ll leave with a shortlist either way.