On Monday morning the management team sits down together. A colleague has tried out a chat tool, a competitor is supposedly doing "something with AI", and the director asks whether you should get started too. Nobody knows exactly with what. Meanwhile, two employees spend hours every week retyping invoices and order emails, and the sales figures live in three systems that each give a different answer. It's a familiar place to be: the pressure to do something is high, but the first concrete step is missing.
You don't start with artificial intelligence by looking at the technology, but by picking one process that eats up a lot of time or often produces errors. Then you check whether the data for that process is in order, set up a small pilot with a clear goal, and let a human keep the final decision. Only once that's clear do you choose the model or tool that fits.
What artificial intelligence in a business actually is
Artificial intelligence, or AI, is an umbrella term for software that performs tasks that would normally require human judgment. Think of recognizing patterns, sorting information or predicting what is likely to happen. In a business it's rarely anything spectacular. Most of the time it's a piece of software that makes an existing work process faster or more accurate.
The applications we see most often in practice are:
- Document processing: pulling data from invoices, contracts or forms and putting it into your system.
- Classification: automatically placing incoming messages, tickets or requests in the right category.
- Forecasting: using past data to estimate what will be needed next month.
- Anomaly detection: spotting deviations in figures or measurements before they become a problem.
- Smart routing: making sure a request lands directly with the right person or department.
It's more than a chat window. A chat tool is one application, and by no means always the most useful one for your process.
You're not too late
If you feel like everyone else is already further along, look at the numbers. According to CBS, Statistics Netherlands, one in six companies used AI in 2025, double the figure of two years earlier. According to werkenmetai.nl, Eurostat calculated that in 2025 well over a third of Dutch companies with ten or more employees put AI into practice, while the EU average is around 20.0 percent. Two counts, two results, but the same direction: usage is growing, and there is still plenty of room to start carefully.
So being in a hurry is not a good reason to launch a poorly chosen project. A company that has bought a tool in a month that nobody uses is no further along than a company that has calmly picked the right process.
Start with a process, not a model
The most common mistake is starting with the question of which AI to buy. Start instead with the question of where your team currently loses time or makes mistakes. To find out, spend a few working days alongside the people who do the work and write down what they do over and over.
A good first process usually has a few characteristics:
- It happens often. A task that comes back every day pays off faster than a rare exception.
- It mostly follows a pattern. If two employees do it the same way, software can learn it too.
- A mistake can be fixed. Don't start with a process where a single miss has serious consequences.
- You can measure the outcome. You should be able to say whether things are going better than last month.
We go into how to spot opportunities in your own work in Finding AI use cases in your own processes. Make a shortlist of five candidates and cross out four. That feels harsh, but one focused first project teaches you more than five half-finished experiments.
Look at your data first
AI is only as good as the data it works with. That's often where things get tricky. If the same question gives a different answer depending on the system you consult, you first have a data issue to solve and only then an AI question.
Ask yourself a few down-to-earth questions:
- Where is the data for this process right now: in a software package, in spreadsheets, in mailboxes?
- Who owns that data and who is allowed to use it?
- Is it correct, or does someone keep fixing it by hand?
- Can other systems access it through an integration?
Sometimes the answer is a data platform that brings the data from separate systems together into one reliable source. It doesn't have to be big: a data warehouse, a data lake or something simpler, depending on where your organization stands today. Sometimes a clean integration between two systems is enough. You can read how ready your applications are for this in APIs for AI: is your application landscape ready?.
Set up a small pilot
A pilot is a clearly scoped experiment with an end date. You test whether AI really works in your situation before putting a lot of money and time into it. Define three things in advance:
- The scope: which process, which data, which group of users. The smaller it is, the faster you learn.
- The success criterion: for example fewer manual hours or fewer errors, expressed in something you can measure.
- The stop criterion: at what point do you say it isn't working and stop? That keeps a pilot from taking on a life of its own.
The stop criterion is the part most teams skip. It's also the most important, because it makes it acceptable for a pilot to fail. Put the three points on paper and have everyone involved confirm them before the start. That way nobody remembers it differently afterwards.
Keep people in the driver's seat
AI makes suggestions, people make decisions. Especially for first applications, that's the wisest setup. The model sorts, reads or flags; an employee checks what matters and can step in.
A few agreements go with that:
- Explainability: you need to be able to see why the system arrived at a result.
- Testability: you can check its behavior with examples before it goes live.
- Maintenance by your own team: a solution that only its builder understands makes you dependent.
- A clear exception route: what happens when the model isn't sure?
These are also the principles we build AI on: practical, explainable and with people in the driver's seat.
What a partner can do for you
You don't have to do this alone, but you don't have to outsource everything either. A good partner helps with choosing the process, organizing the data, and building and testing the solution. After that, your own team can maintain it.
At Data & AI Solutions, Isatis builds AI solutions and the data platforms they run on, such as document processing, classification, forecasting and anomaly detection. We use AI where it saves time or catches errors. The team that builds the software can also keep it running for years, so nothing falls between two suppliers. Isatis is ISO 9001 and ISO 27001 certified and works for healthcare, education, manufacturing, government and energy companies, among others.
The first step is a small one: a 15-minute conversation in which you tell us what you want to achieve. We'll tell you honestly whether we can help. If we're not the right party, you'll hear that too.
Frequently asked questions
What is the difference between ChatGPT and AI?
AI is the broad term for software that performs tasks that would normally require human judgment. ChatGPT is one specific application of it, focused on language and conversation. For many business processes, such as reading documents or flagging deviations, you need something other than a chat window.
What types of AI are there?
In a business you mainly come across applications such as document processing, classification, forecasting, anomaly detection and intelligent routing. So it's about what the software does, not an abstract label. Choose the type that fits your problem.
What does AI stand for?
AI stands for artificial intelligence, or kunstmatige intelligentie in Dutch. In Dutch texts you'll see both terms used interchangeably, and they mean the same thing.
Do I need a data platform before I start with AI?
Not always. A data platform is mainly needed when the same question gives a different answer depending on the system you consult. For a small, clearly scoped process, a clean integration or a simpler solution may be enough. First look at the data for the process you want to start with.
A first step with no strings attached
Not sure which process is suited to AI, or whether your data is ready for it? We'd be happy to think along with you. You can get in touch and schedule a 15-minute conversation. We respond to a message sent through the form within 24 hours.
About the author: Jeroen Penninx, Director and co-founder of Isatis.





