What I've found AI is actually good for in a service business

Most people who ask me about AI have been told it'll save them time, but nobody has told them which part of their week it should touch.

I've spent a few years now looking at that question inside real businesses. Here's what I've found, including the parts where I think it's the wrong tool.

What the useful stuff has in common

Three things, and the ones that pay off tend to have all three. It happens many times a day, not once a week. The input is words: an email, a transcript, a form somebody filled in. And a person was going to read the result anyway, so a wrong answer gets caught in five seconds.

I've come to weight that third one heaviest. If a wrong answer gets caught in five seconds by a person who was reading it regardless, occasional errors are survivable. If a wrong answer flows quietly into an invoice, you have a different problem.

Where I've seen it pay off

Turning conversations into records

Calls and consultations produce information somebody is supposed to type into a system afterward. It often doesn't happen. Pulling the structured parts out of a transcript works well here, because the input is language and somebody reviews it as part of normal work.

First drafts of writing you do over and over

Proposals, scopes, follow-up emails, recurring reports. What I've watched happen is that people stop putting these off. Starting from something rough beats staring at a blank page, and the thing gets sent the same day.

Reading documents you'd otherwise skim

Intake forms, contracts, submitted paperwork. Pull out the fields that matter, flag the ones that look wrong, and a twenty-minute read turns into a two-minute check.

Sorting and routing what comes in

Classifying inquiries by type, urgency, or fit and sending them where they go. This is the boring one. It's also usually the highest return in the building, because it happens dozens of times a day and right now it costs somebody's attention every single time.

One thing I check before building any of this: how often does it run? I've opened up automations that fire once a day when they should fire the moment an email arrives. That's a plumbing problem and it's cheap to fix. Worth ruling out before you spend money on anything smarter.

Where I've seen it fail

• ·Anything that has to be right the first time with nobody checking. Billing, payroll, anything with legal exposure.
• ·Decisions that need context the model can't see, which covers most of the decisions that matter in a service business.
• ·Anywhere a confident wrong answer costs more than the whole thing saves.
• ·Relationship work. Clients can tell, and what it costs you when they notice is more than you saved.

The part people don't want to hear

This is the most common thing I find when I go looking at a business that wants AI.

AI can't fix a process with no system of record. If nobody can say where a client stands today, adding intelligence on top gives you confident answers built on incomplete information.

A lot of businesses asking me about AI can't tell me how many active clients they have right now without someone going and counting. That's the thing to fix first. The AI part is more interesting, but it's sitting on top of numbers nobody can produce.

Building the foundation first is less exciting, but it's where I've seen the money. It also makes everything after it obvious, because once there's a real record of what's happening you can finally see which parts of the week repeat.

I bring back two lists, not one

When I look at a business I come back with two lists. One is what would save the most money. The other is what people complain about most, which is usually something like the weekly report that takes somebody four hours every Friday.

Those come out different almost every time. There's usually some process everybody dreads that isn't costing much, and fixing it changes how people feel about their week more than it changes the P&L. Whether that's worth doing is your call. I've stopped trying to make that decision for people, and I'd rather you see both lists than have me pick one.

How to work out where yours fits

Don't start with AI. Write down where your week actually goes, then ask three things about each block of work that repeats:

1. 1.How often does this happen? Weekly isn't enough. You want daily, or many times a day.
2. 2.Is the input language, or is it a judgment call? Language is workable. Judgment usually isn't.
3. 3.If it came out wrong, when would somebody notice? Right away, because they read it anyway, means it's a good candidate. A month later in an invoice means leave it alone.

Whatever survives all three is worth building. In my experience that's one or two things, not ten, and they're rarely the ones people expected walking in.