Most of what you read about AI is written for companies with a data team. You don't have a data team. You have ten people, a phone that rings, a spreadsheet that three people edit, and a to-do list that never gets shorter.
So here's the version for you. What AI is genuinely good at right now, in a business your size. What it still gets wrong. And how to find out for yourself in about a week, without signing anything.
What it does well
Turning messy input into structured output. This is the big one, and it's boring, which is why nobody writes headlines about it. A voicemail transcript, a rambling web form, a text message, an email thread. AI can read that and pull out: name, phone, what they want, how urgent it is, where they are. Reliably. That's the unlock behind most useful small-business automation, because almost every process you run starts with a human saying something imprecise and a staff member turning it into a record.
Drafting the message you were going to send anyway. Follow-ups, quote summaries, appointment confirmations, review requests, the "sorry we missed you" text. AI writes a solid first draft in your voice if you give it examples of your voice. You still read it before it goes out, at least at first.
Answering questions from documents you already have. Your pricing sheet, your service area, your warranty terms, your onboarding checklist. Point AI at those, and your team can ask plain-English questions and get plain-English answers with the source attached. This is the least glamorous and highest-ROI use I set up for people.
Classifying and routing. Is this a new lead, a current customer, a vendor, or spam? Is it an emergency? Which tech should get it? AI is good at this and it's the difference between a lead sitting in an inbox for six hours and a lead getting a callback in six minutes.
Summarizing. Ten sales calls into a Monday brief. A week of support tickets into three recurring problems. Long email thread into a two-line status.
What it does badly
Anything where being wrong is expensive and nobody checks. AI will produce a confident, plausible, incorrect answer. Not often, but often enough that you never put it in a spot where a wrong answer goes straight to a customer or a bank account without a human in the loop. Quotes with real numbers. Legal or medical language. Commitments about dates.
Knowing things you never told it. It doesn't know your prices, your capacity next Tuesday, or that you fired that subcontractor. If the information isn't in a system it can read, it will guess. Guessing looks exactly like knowing.
Judgment about people. Whether this customer is about to churn and needs you personally to call them. Whether the angry review is worth a refund. Whether the guy who's been with you eight years gets a break on the invoice. Don't hand that over. Not because AI is dangerous, because that's your actual job.
Being the whole system. AI is a component. It's not a CRM, it's not a phone system, it's not a database. Bolting a chatbot onto a broken intake process gets you a fast, articulate broken intake process.
The shape of a good first use
Look for a task with four properties:
- It happens a lot. Ten times a week minimum, or the savings never show up.
- It has a clear input and a clear output. Voicemail in, structured lead out. Ticket in, category out.
- The cost of being wrong is low or a human catches it. Draft, not send. Suggest, not commit.
- You can tell if it worked. You can count the result.
That last one eliminates most of the "AI strategy" conversations people want to have. If you can't count it, you can't tell whether you bought a tool or a story.
Concretely, in a ten-person business, the first thing worth automating is almost always one of these:
- Every missed call turns into a text within a minute, and a lead record within five.
- Every web form and inbound email is read, categorized, and routed to a named person with a deadline.
- Every quote that goes out gets a follow-up sequence that stops the moment the customer replies.
- Every completed job triggers a review request at the right time, to the right channel.
None of that is exciting. All of it makes money.
How to test it in a week
You don't need a pilot program. You need one week and a stopwatch.
Monday. Pick one process. Write down what happens today, step by step, including who does it and how long it takes. Don't design the future yet. Just write down the present. Most owners are surprised here, because the real process isn't the one in their head.
Tuesday. Count. How many times did this happen last month? How long did each one take? What did it cost when it went wrong? Multiply. Now you have a number. If the number is small, stop and pick something else. That's a win, not a failure.
Wednesday and Thursday. Do the AI part by hand, badly, with a human. Seriously. Have someone paste last week's voicemails into a chat tool and ask it to extract name, need, urgency, and callback number. Look at the output. Is it right? How often is it wrong? What kind of wrong? You've just run the exact experiment a vendor would charge you for.
Friday. Decide. If the manual test worked, the automation will work, because automation is just doing that reliably without a person. If the manual test produced garbage, no amount of software will fix it, and you saved yourself a subscription.
The thing nobody tells you
The hard part was never the AI. The hard part is that most small businesses don't have the plumbing for the AI to plug into. The lead lives in an inbox. The customer lives in the phone. The job lives in a notebook. Nothing talks to anything.
You can add AI to that, and you'll get a smart component in a dumb system. What actually moves revenue is connecting the pieces so the information flows, then letting AI handle the messy translation at each seam. That's a systems problem, and it's mostly not glamorous, which is exactly why it's still sitting there unfixed and still leaking money.
Start with the one process you can count. Get it working. Then do the next one.
If you want a second set of eyes on which process to start with, that's most of what I do. Have a look at what I build, or just tell me what's broken and I'll tell you if it's worth automating.