AI implementation for SMEs: where to start

An AI implementation doesn’t start with a tool, but with one concrete business question and a goal you can measure. Then come a baseline measurement, getting the data in order, the lightest solution that works, a short pilot and a decision based on the figures. How fast that goes depends on the complexity: a simple application often works within a week, a custom build on your own data within four to six weeks.

A licence for everyone. A lunch session with handy prompts. Three months later, four colleagues are using it, mostly to make their emails sound nicer.

Sound familiar?

That is how AI starts in a lot of companies: with the technology, in the hope that the use case will follow. It rarely does. Not because the tools fall short, but because nobody wrote down which work was supposed to get faster or better. If you want AI for business to pay off, reverse the order: the question first, then the data, then the tool.

It’s no surprise that AI is on so many management agendas. Its use is growing fast: according to Statistics Netherlands (CBS, in Dutch), 33 percent of Dutch companies with ten or more employees used one or more AI technologies in 2025, and another 10 percent had considered it without starting. Whether you want to get started with AI or move on after a first experiment, this is how. Plus two things most step-by-step plans skip: whether you need AI at all, and what “data in order” actually means.

What is an AI implementation?

An AI implementation means building artificial intelligence permanently into the way your organisation works, so that a task becomes structurally faster or better. The difference from an experiment: there is an owner, a fixed way of working and a figure you judge the result on.

When implementing AI, you choose between roughly three kinds of AI solutions for businesses:

• An AI assistant. Microsoft 365 Copilot, ChatGPT or Claude helps with writing, summarising and searching documents. The gain depends on how well your people use it.
• AI in software you already have. More and more accounting, CRM and ERP packages read invoices or suggest a reply to a customer question. Often it’s a matter of switching it on.
• Having an AI solution built. An application of your own on your own data, such as a workflow that reads incoming emails and routes them. Only worth it if the first two routes don’t do the job.

Which AI use cases fit you depends mostly on which work takes the most time. If it’s AI in business processes such as invoicing or order handling, the approach is close to business process automation: improve the process first, then automate it.

Why AI projects in SMEs stall

Almost never on the technology. They get stuck on everything around it:

• Starting with the tool. The licence was bought before anyone knew which problem it should solve.
• No owner in the business. Without an owner outside IT, nobody decides on scaling up or stopping.
• Starting too broad. Start everywhere at once and you have nowhere enough to judge.
• Data that isn’t right. A model running on contradictory data gives confident answers that are wrong.
• No measuring point. Without a baseline, nobody knows after three months whether it delivered anything.

A successful AI implementation is therefore mainly an organisational question. The technology is the easy part.

Do you actually need AI?

With AI for SMEs, one test before every project helps: is this a question for AI, or for something simpler? A large share of what is sold as AI automation is ordinary rule-based automation with a new label.

AI is strong at work with unstructured input: reading text, summarising, sorting, drafting. “Send a reminder when an invoice has been open for thirty days” is a rule, not an AI question. And “we want to see how the margin per customer develops” is a question for a dashboard, not for a chatbot.

• Does the work follow fixed rules? Then a workflow or integration is cheaper and more reliable.
• Does someone have to read and judge text or images? Then AI is a serious candidate.
• Do you want to see figures and steer on them? Then you need a reliable data model and a dashboard. AI can sit on top of that later, but doesn’t replace it.

That also answers the question many managing directors ask: what can AI do for my business? Less than the demos promise, and more than you think once you know where the reading and typing work sits.

Get the data in order first

Say your revenue figures live in three places: the ERP, the CRM and an Excel list the sales director maintains personally. For years, the management meeting has been about which of the three is right. Put an AI assistant on top and that discussion doesn’t go away. You just get an answer faster, from one of the three, without knowing which.

That’s why the foundation matters more than the model. It isn’t the smartest AI that delivers the most, but the AI that works on unambiguous data. Before an application uses your own business data, check four things:

• One source per data item. For customers, products and revenue, it’s agreed which system is leading.
• Recorded definitions. Everyone means the same thing by revenue, margin or an active customer.
• An owner. Each source has someone responsible, and errors are fixed at the source.
• Access and permissions. An AI assistant allowed to search everything an employee can reach will also find the payroll list that ended up in a shared folder by mistake.

How to fix errors structurally is covered in our article on data quality. The agreements on sources, definitions and ownership belong in a data strategy.

Implementing AI in your business: six steps

1. Pick one business question. A question with a number in it: processing purchase invoices takes 20 hours a week and has to drop below 8. “Something with AI” is not a question.
2. Take the baseline. Record how the work performs today: hours, lead time, error rate. How to set up such a figure is explained in our KPI examples.
3. Check the data. Go through the four points above for the data the application needs.
4. Choose the lightest solution that works. First AI in software you already have, then an AI assistant, and only after that custom AI solutions.
5. Run a pilot. One to two weeks for a simple application, two to four with integrations or your own data. Always with a small group of users who actually do the work and a success criterion agreed in advance.
6. Decide, scale up and embed. If the pilot meets the criterion, roll it out and assign an owner. If not, adjust or stop. That is a good outcome too, as long as the decision rests on figures.

Whether you still have to start with AI or already have a failed experiment behind you, the order stays the same. If you want to use AI in several places, go through the steps per application.

Proof of concept or pilot: what’s the difference?

A proof of concept tests whether something is technically possible. A pilot tests whether it delivers anything in real work.

Building a proof of concept usually takes a few days. You might have a model read a hundred invoices from different suppliers and count how many fields it gets right.

So what is an AI pilot? The working solution runs for a fixed period in day-to-day work, and you measure the same figures as in the baseline. Only then do you see whether the time the model saves isn’t lost again to checking and correcting.

Agree the success criterion in advance, with a number: at least 40 percent fewer hours, or fewer than 2 percent errors. A pilot without a baseline is always a success, after all. Nobody can prove otherwise.

What does an AI implementation cost?

The costs follow the route. AI in your existing software usually costs only set-up time, an AI assistant a fixed amount per user per month. A custom application adds the implementation of AI models: development hours, integrations and testing. For most AI solutions for SMEs you count four cost items: licences, your own hours, outside support and maintenance after go-live.

Set the return against that, in hours. If five employees each spend four hours a week checking purchase invoices, that is 920 hours a year over 46 working weeks. If the pilot removes half of that, you know what the solution may cost. That’s ROI-driven automation: the return sets the budget, not the other way round.

For a standard AI assistant you usually don’t need a partner: agreements, a training session and a few examples per department are things you can arrange yourself. Outside support pays off once the application uses your own business data or touches several systems.

Bringing people along, and what the AI Act requires

When implementing AI, success lies with the people who have to work with it. Involve them from the pilot onwards: they know where the exceptions are and are the first to see where the model gets things wrong. And say up front what happens to the time that is freed up.

Bringing people along now has a legal side as well. Since 2 February 2025, the European AI Act requires every organisation that uses AI systems to take measures for the AI literacy of the people working with them. According to the European Commission’s questions and answers, that also applies if your staff only use ChatGPT to write texts: they need to know the risks, such as made-up answers. No certificate is required, an internal record of training is enough, and supervision started on 2 August 2026.

So agree which tools your team is allowed to use and which data must never go into them, such as personal data covered by the GDPR. Where the data in those tools ends up and whose law applies there is covered in our piece on data sovereignty.

How does it work at your company: does AI start with a question someone wants solved, or with a licence someone bought? Want to know which application would pay off first for you? Get in touch.

Frequently asked questions