Where to start with AI in a small business: pick the team that loses the most time

Start with the team that loses the most hours to repeat work, and give AI one job there. Not a company-wide rollout. Not a new tool for everyone. One team, one job, and one number you can check.

That advice sounds small, and it is meant to. Most small businesses have tried AI. Very few have made it part of how the work gets done.

Most businesses use AI. Few have built it in.

The Goldman Sachs 10,000 Small Businesses survey asked 1,256 owners about AI in early 2026. About three-quarters use it in some form. Only 14% said it is fully embedded in their core operations, and 73% want more training and help putting it to work (Goldman Sachs, Fortune). The top barriers were data privacy and security (50%) and lack of technical expertise (49%).

Government data asks a narrower question and lands lower. The Census Bureau found 19.8% of US firms used AI in a business function as of May 2026. Among firms with 4 or fewer employees, it was under 20% (Census Bureau).

So most businesses sit somewhere between “a few people use ChatGPT” and “AI handles part of a real process.” Getting across that gap doesn’t take a big project. It takes a good first pick.

How to pick the first job

For one week, ask each team to jot down the work they repeat. Then look for a job that fits most of this list.

• It happens many times a week, mostly the same way.
• It follows rules you could write on one page.
• A person can step in when a case gets unusual.
• You can count it, as calls answered, questions resolved or invoices chased.

Pick the team where that job eats the most hours. That is your starting point, even if it isn’t the most exciting option on the table.

Before you change anything, write down the current number. How many calls went to voicemail last month? How many support emails came in? How many invoices went past 30 days? Without that count, nobody will know later whether the AI helped.

Starting points by department

Front desk and sales

The usual first job here is answering calls and chats that would otherwise go to voicemail, and replying to new leads quickly. Response times are often slow. A 2025 secret-shopper study of 1,333 law firms by Hennessey Digital found a median of 13 minutes to first human contact, and 26% never responded (JustCall summary). ServiceTitan’s 2026 survey, a vendor claim, found only 50% of contractors reply to new leads within the first hour (ServiceTitan).

Customer support

Think of the questions your team answers every day. Where’s my order, how do I reschedule, what’s your return policy. AI can answer the common ones from your own policies and hand tricky cases to a person with the whole conversation attached.

Operations

Look at the updates and hand-offs nobody enjoys. “Your tech is on the way” texts, job status updates, and the details people copy from one system into another. Each automation should stop the moment a person steps in.

Finance and admin

Reading and filing invoices and receipts is a common place to start. So are polite reminders on overdue invoices and month-end summaries for your accountant to check.

HR and people

Time-off, pay and benefits questions can be answered from your handbook, and every new hire can get the same onboarding checklist. A person still makes every hiring decision.

Marketing

Asking every customer for a review, and drafting replies to the reviews you get, is a contained first job. Nothing gets posted until you approve it.

Legal and compliance

Summarizing contracts and marking the clauses that need a look is a good fit, with a person making every call. If the files can’t leave your office, this can run on a private server instead of a cloud service.

What to expect in the first months

For customer support, we plan for AI to resolve 30 to 50% of questions on its own at first, and we tune it every month. That is a planning range, not a guarantee. It depends on how complete and current your documents are.

You will see higher numbers in vendor marketing. Intercom reports an average 76% resolution rate for its Fin agent across more than 12,000 customers (AI Agent Index). HubSpot reports 65% (HubSpot) and Tidio 67% (Tidio). These are vendor claims, and how a vendor counts matters. One analysis reports that Intercom changed its count to leave some conversations out (dev.to).

Smaller tests land lower. One review measured 38% across 500 tickets at 4 small-business clients (Superframeworks). In a March 2026 survey of 2,000 consumers commissioned by Ada, itself a vendor, only 24% said their last AI customer service interaction was fully resolved (IrisAgent).

The realistic picture is that AI handles the routine share and passes the rest to your team with the context attached. Your people still handle the hard cases. The aim is that less of their day goes to the same five questions.

Mistakes that stall a first project

• Buying a tool before deciding which job it should do.
• Skipping the before-and-after count, so nobody can say whether it helped.
• Sending automated texts to people who never agreed to them. Texas has treated texts as telephone solicitation since September 2025 (Morgan Lewis).
• Letting the AI answer from general knowledge instead of your own documents.
• Leaving nobody responsible for the cases the AI can’t handle.
• Starting three projects at once, so none of them gets tuned properly.

Next step

Start the chat and tell us which team is stretched thinnest. We’ll suggest a first job, what it would connect to, and how we’d measure it.