What can AI do in a small business today?

If you want to use AI in a small business, begin with one task that keeps coming back, and leave the rest of your business alone for now.

By AI I mean language models here, the technology behind tools like ChatGPT and Claude. That is software trained on huge amounts of text, which lets it read and write text. An Excel formula always gives the same result for the same input. A language model estimates which answer fits, and that makes it usable for work that is hard to capture in fixed rules. It also explains why it sometimes makes mistakes.

What AI can do today

AI can draft messages your team only has to check. Felix, my own product for boutiques, writes a personal message for each suggestion, and the owner can edit it.

Sorting is a second kind of work. Say a contractor gets requests by email every day: AI can divide them by type of job and pull out the details, so someone can follow them up faster.

AI can also flag errors in large amounts of data. Eagl, where I work in the team, catches booking errors before they hit the month-end close. A fixed rule can check whether a booking matches a pattern, and a language model can go further and reason about why it looks wrong.

You can also ask AI questions about your own data. A business owner might want to know which customers have not ordered anything this year, without having to build an overview for it.

Start with one task

Choose something that comes back every day, week or month and takes some time on each occasion. Those are the same signs as in When has your spreadsheet grown too big? A good first task has two further qualities: you can check the result quickly, and a mistake is easy to put right. A draft reply to a standard question meets both, while an automatic payment does not.

With one task you quickly see what AI does well and where it goes wrong, and you can check whether you are really getting time back. If that works, you have a reason to pick a second task.

Work with your own data

A language model knows nothing about your customers. Ask it for a message to a loyal customer and you get a general text. AI gets useful when it works with your own data, like your sales history or your bookings.

Felix is built to connect to the shop’s till and look at who bought what and when. That lets it say why someone is due a message, for example “hasn’t visited in six months” or “a new arrival matches something they bought last spring”. If a customer at the counter asks for something that is not in stock, the owner can log it by voice. Felix watches incoming stock and flags a match.

Your data does have to be reachable for this. How that works is explained in Connect your software to your till and your accounting.

Count on mistakes

AI makes mistakes, so you need a way to catch them. A language model can get something wrong and still write it down with great confidence, so the tone will not tell you.

How you catch mistakes depends on the task. With messages, a person reads along before anything goes out. With figures, you compare part of the results with the source. It also helps if the software shows what a suggestion is based on, the way Felix says why it suggests a customer, because then you see sooner whether the reasoning holds. Keep track of what goes wrong as well. If the same mistake comes back, you can change the instructions or the data.

Have a person approve

With Felix, nothing goes out until the owner has read and approved the message. A customer reads a message like that as something the owner wrote. There is more on this in Why Felix sends nothing without the owner.

The same goes for everything that leaves your business, like a quote or a reply to a customer: a person approves it before it goes out. That approval takes some time, so include it when you work out how much time you get back.

Where your customer data ends up

If you put customer data into an AI tool, you are passing it to another company. The GDPR, the European privacy law, allows that under conditions, and you usually remain responsible for it yourself. This concerns personal data: anything that lets you recognise someone, like a name, an email address or a particular customer’s purchases.

You can ask the supplier these questions, or look them up in its terms:

• Where is the data stored, and does it leave the European Union?
• Does the supplier use your data to train its models? That can differ between a free account and a paid business account.
• How long does the data stay there?
• Is there a data processing agreement, a contract that says what the supplier may do with your data?

Check too whether your privacy statement mentions that you use this service. Terms change, so read the current version. This is a first explanation and not legal advice. For your situation, ask a lawyer, and the Belgian Data Protection Authority has guidance for businesses.

Until that is sorted out, you can run a first trial with made-up sample data.