Where should a small business start with AI?

Start with one task, not with a tool. Pick something that comes up every week, takes a good deal of your time and has a result you can check in a few minutes: replies to typical inquiries, first drafts of proposals, meeting notes. For a week, try AI on real work. The customer chatbot and the “AI strategy” can wait.

Why not start with a tool?

According to Eurostat, in 2025, 8.4% of Polish firms with at least 10 employees, in the industries the survey covers, used AI technologies. The EU average was 20.0%. If you have fewer than ten people, you’re not in this statistic at all, because the survey covers firms with 10 or more employees.

In my view, a late start has one upside. You can skip the stage where a business buys licenses for everyone and waits to see what comes of it. That scenario usually ends the same way: someone does something in ChatGPT, someone else tests a third tool, and after a quarter nobody can say what changed.

Choose the tool last, once you know what it’s for.

You first, for a week

Ethan Mollick, a management professor at Wharton who studies AI at work, advises in his book Co-Intelligence that you bring AI into everything you do, as long as the law or ethics don’t rule it out. No manual will tell you what AI does well in your industry, so you have to find out on your own work.

Mollick also describes something he calls the jagged frontier. Some tasks that seem hard, AI does very well. Others, which look easy, it gets wrong. An example from the book: it will write a sonnet with no trouble, but usually not a poem of exactly 50 words. You can’t see this frontier from the outside. You only see it after a few dozen tries.

That’s why the first week belongs to you, not to your team. Every day, give AI three things you do anyway. For each one, write a single line: did it help or not, was it faster or slower.

How do you choose the first task?

After a week you have a list. The first task to hand over to AI should meet five conditions:

it comes up at least a few times a week;
it takes a noticeable amount of time, ideally yours;
you can check the result in a few minutes, without extra digging;
a mistake doesn’t go straight to a client, because someone reads it before it goes out;
one person is responsible for the result.

Mollick advises starting with tasks you dislike but can easily check. I add one question before that: does this task need to exist at all? A report nobody reads can simply be crossed off.

Take a fictional renovation firm with eight people. Every evening the owner answers inquiries from the website and by phone: how much a bathroom costs, when there’s a free slot, whether they do facades. The answers are similar, yet each one is written from scratch. That’s a good candidate. AI drafts a reply from the price list and the calendar; the owner reads it, corrects it and sends it. A bad candidate from the same firm is quotes sent to clients automatically. There, a mistake costs money and reputation, and checking it takes a site visit.

Before you decide it works, test AI on 10 real cases from the past few weeks, including 3 hard ones. Judge the result as strictly as you judge your own emails to clients.

Where not to start

with a chatbot for customers on your website: it talks to people without you in control, and that’s the worst place for a first lesson;
with automating a process that doesn’t work today, because AI will only speed up the mess;
with licenses for the whole team and no specific task;
with a project that needs a connection to your accounting or inventory software;
with decisions, because AI can prepare them but shouldn’t make them;
with pasting client data into a free account before you check the provider’s terms.

When this doesn’t work

This method assumes your business has repeatable tasks. If every job is different, look for repetition one level down: in the emails, notes and documents you write for every project.

If your problem is a lack of clients, AI in admin won’t solve it. The time you win back only helps once you know what you’ll spend it on.

The frontier of what AI can do moves with every new model. A task it does badly today may come out well in six months. Go back to your list every quarter.

What to do tomorrow

List ten tasks from last week that repeated. Next to each one, mark whether you can check the result in five minutes. Pick one of the marked tasks.

Business X-ray in 5 minutes

How this text was written: the research, the choice of sources and the final version are mine. AI helped me prepare a draft; I checked every fact, and I am responsible for this text. — Maciej Chmura

Sources

Eurostat, news release of December 11, 2025 on the use of AI technologies in EU enterprises (enterprises with at least 10 employees, NACE sections C–J, L–N and 95.1; Poland 8.4%, EU 20.0%): https://ec.europa.eu/eurostat/web/products-eurostat-news/w/ddn-20251211-2 (accessed October 15, 2026).
Ethan Mollick, Co-Intelligence: Living and Working with AI, ch. 3 “Four Rules for Co-Intelligence” (the principle of bringing AI into every task, the jagged frontier, the example of the sonnet and the 50-word poem).
Ethan Mollick, Co-Intelligence: Living and Working with AI, ch. 6 “AI as a Coworker” (start with tasks you dislike but can easily check).