Skip to content
Articles / data10 min read

Getting your data ready for AI:where to start and what to leave alone

Most businesses that get stuck with AI get stuck on their own data. You don't have to clean up everything: start with one question. An action plan in six steps.

Published
By
Luuk
Reading time
10 min read
About
data · base · privacy

Most businesses that get stuck with AI don't get stuck on the AI. They get stuck on their own data. In a survey of 104 Dutch organisations, 54% named poor data as a barrier, right after a lack of knowledge1.

What I see at clients is almost always the same: the data is there, but there's a lot of noise in it. Customer details that look slightly different in three systems. Folders full of old versions. And knowledge that only lives in a colleague's head. An AI can't do anything with that, or worse: it confidently gives you the wrong answer.

The good news is that you don't have to clean up your whole business first. You start with one question you ask every week, and you make sure the answer is right in one place.

  • After knowledge, data is the biggest barrier to AI, in the Netherlands too.
  • Your data doesn't have to be perfect. It has to fit the question you ask.
  • Start small: one question, the sources that go with it, and someone responsible for each source.
  • Write down knowledge that now only lives in people's heads.
  • Decide who can see what before an AI can read everything.
  • Keep your data in the EU, also when the AI reads it.

We built Base for exactly this problem. Base connects the systems you already use and puts its own AI on top, running in the EU. We do the connecting. But which source holds the truth and who can see what, only you can decide. That's what this article is about.

01

Why data is the brake

The problem usually isn't too little data. It's that nobody knows anymore which data is right. In practice we see three things come back again and again.

Mess. The same customer is in your accounting, your CRM and your email, and slightly different everywhere. A CRM is the system where you keep track of customers and deals. Of over 600 companies that use one, 76% say less than half of that data is accurate and complete2. Missing and wrong data come up more often than duplicates.

Noise. Businesses keep far more than they use. In a study by Omdia for storage company Everpure, 99% of organisations say they keep data nobody uses anymore. 58% don't even have a view of everything that's stored3. For an AI that's tricky. It doesn't know the quote from 2022 no longer applies.

Forgotten knowledge. A lot of things you just know. Which customer prefers a call to an email. Why that one discount was ever given. Research by Panopto found that 42% of the knowledge in a company sits with that one employee only4. That study is older and looked at larger American companies. I expect a small business with ten people has more of it rather than less, because less is written down.

What the research says

Name poor data as a barrier

54%

Of 104 Dutch organisations. Only a lack of knowledge scores higher.
CRM less than half in order

76%

Less than half of the customer data is accurate and complete.
Knowledge sits with one person

42%

That knowledge is gone when that colleague isn't there.
Of working time goes to searching

25%

Knowledge workers lose a quarter of their time searching for answers.

So this is what you get. According to Atlassian, knowledge workers lose about a quarter of their time searching for answers5. Put an AI on top that reads everything, and it searches faster. But it finds the same mess.

Large companies are no different. Of the IT leaders in the Everpure study, 68% name data management as the biggest brake on getting AI from a trial to daily use. So it plays out everywhere, large and small.

02

You don't have to clean up everything

The reflex is: clean up everything first, then AI. That project never starts, or it never ends.

Research firm Gartner puts it very clearly: "There is no way to make data AI-ready in general or in advance"6. Whether your data is ready depends on what you want to do with it. Gartner even adds that errors and outliers are allowed, as long as the data fits the question.

So you turn it around. You start with a question. For example: which quotes are still open and who should I call back? For that question you need your CRM and your email. Not your accounts from five years ago, and not the shared drive with old presentations either.

Gartner also warns that it isn't a one-off job. They expect organisations to abandon 60% of their AI projects through 2026 if the data isn't ready for it7. Getting your data ready is maintenance. Like your bookkeeping: you do a little every week, not everything once a year.

Whether that's equally easy for every business, I don't know. A business with one accounting package and a tidy mailbox is ready for the first question quickly. A business with ten years of shared drives has more work. But there too, you start with one question.

03

How it works: a head office with rooms

Think of your business as an office building. Every system has its own room: the accounting, the CRM, the email, the planning. Your team walks in and out and just keeps working.

The problem is that nobody has the overview. If you want to know where you stand, you walk past every room and write it down on a note yourself.

Base is the hallway between those rooms. From every room a fixed line runs to a screen in the hallway. In the server room sits its own AI that reads what comes up on that screen, and answers your questions. That server room is in the EU.

What does that mean for your data? Your rooms don't need rebuilding. You don't switch to another package. But it does have to be clear which room holds the truth. If a customer's address is in your accounting and in your CRM, someone has to say which of the two counts. Otherwise the screen in the hallway shows two answers.

Which room holds the truth?

Two rooms with the same address give two answers. Pick one, and the screen is right again.

04

Who can see what

An AI that can read all your data also shows how messy your permissions are. That sounds like a detail, but large companies got badly stuck on it.

When Microsoft rolled out its AI assistant Copilot, Gartner asked 132 IT leaders how it went. 40% delayed the rollout by at least three months8. The reason: the AI showed employees documents they weren't supposed to see, like salaries. According to Gartner that wasn't Copilot's fault. It came down to folders and permissions nobody had kept up for years.

So you also have to decide who the AI knows things for. Someone on the shop floor doesn't need to see the margins. The bookkeeper doesn't need to read every customer conversation. You settle that before the AI starts reading, not after.

And then there's the question of where your data goes. Of the Dutch companies that considered AI but don't use it, 49% name privacy as the reason9. I think that's fair. What ChatGPT and Claude do with your data, you can read in Claude, ChatGPT and your business data. How to run an AI model in the EU yourself is in Running your own AI model.

Who can see what?

Pick a role and see which rooms get locked, and what shows up on the screen.

05

Action plan: how to get your data ready

Six steps you can take yourself, with or without us.

  1. Pick one question you ask every week. Where do we stand? Which invoices are still open? Which quotes should I follow up? The more concrete, the better.
  2. Find out where the answer lives now. Write down which systems, folders and people you need to answer that question. Often it's fewer than you think.
  3. Decide per source what the truth is. If something is in two places, pick one. And put a name on it: who keeps this source up to date?
  4. Get knowledge out of heads. What does a new colleague need to know to answer this question? Write it down, in plain sentences. An AI reads a note just as well as a table.
  5. Decide who can see what. Go through the permissions of the sources you picked. Whatever is open to everyone now, everyone will soon see through the AI.
  6. Keep it up. Ask the AI the question every week and check whether the answer is right. If it isn't, you know straight away which source needs attention. Then you take the next question.
06

What we see at clients

It differs per business, but the picture is often the same. The data is there. It's just messy, there's a lot of noise in it, and part of the knowledge is forgotten or sits with one person. Two examples.

A notary's office. There, an AI assistant had to be able to answer clients' questions. The answers did exist, but they were all over the place. The office put together a document for it, and they spent quite some time on that. The owner did part of it and divided the rest among colleagues who know a different subject. That knowledge only lived with them.

A wedding and event venue. A lot of information is on their website, and an AI can simply read that. But that's only what's public. A couple getting married wants to know more: roughly how such a day goes, what happens when and what it costs. That was in a run sheet that wasn't online anywhere. They added that run sheet, and now the AI can answer those questions too.

What I take from this: most of the work was in gathering and writing down knowledge, and in choosing what the AI may and may not tell. The technology was the smallest part. That's exactly steps 3 to 5 of the action plan.

That's actually reassuring. It means you're not starting from zero. You don't have to introduce a new system. Mostly you have to make choices you've been putting off for years.

07

When it's not worth it

I'd generally advise against this if:

  • You have no question that comes back

    If you only want to use AI to write a text now and then, you don't need to get your data ready. A business account with one of the big providers will do.

  • Your business runs on one system

    If everything is already in one package, and that package has a good AI feature of its own, try that first.

  • Nobody wants to keep the sources up

    An AI doesn't clean up a mess. It just shows it faster.

Frequently asked questions

Does my data have to be completely clean before I can use AI?

No. Your data has to fit the question you ask, not be perfect. Gartner itself says you can't make data AI-ready in general or in advance. Start with one question and get the sources for that question in order.

What does "AI-ready data" mean?

Data that fits the task you want done. Gartner describes it as data that is representative of that one use case, including the errors and outliers that come with it. An overview of open invoices needs different data than a customer service assistant.

Why do so many AI projects get stuck on data?

Because the same details are slightly different in several places, because a lot of stored data isn't used anymore, and because knowledge lives in heads. In the Netherlands, 54% of organisations name poor data as a barrier, right after a lack of knowledge (InSpark, 2025).

Do I have to switch to another system?

No. With Base your systems stay where they are. Base reads along through a connection and only writes back when you approve something.

What if my system has no connection?

If a system has an API or an export, it can be connected. An API is a fixed connection programs use to exchange data. If neither works, we'll look together at what does.

Will the AI see everything in my systems?

Only what you allow. In Base you set per role who can see what, and what a connection may do. Do go through your permissions first: whatever is open to everyone now, everyone will soon see.

Does my data stay in Europe?

With Base it does. The language model runs on our own server in the EU, next to your database, and isn't trained on your data.

Rather not figure it out yourself?

That's fine. With Base we connect the systems you already use and put its own AI on top, running in the EU. You don't install or manage anything.

What we do ask of you are the choices from this article: which question first, which source holds the truth, and who can see what. We make those together, in a half-hour conversation.

Talk through which question you want AI to answer first.

Sources
  1. 1InSpark, 2025 · dutchitchannel.nl
  2. 2Validity, 2025 · validity.turtl.co
  3. 3Omdia and Everpure, 2026 · futurecio.tech
  4. 4Panopto, 2018 · hrdive.com
  5. 5Atlassian, 2025 · atlassian.com
  6. 6Gartner · gartner.com
  7. 7Gartner, 2025 · gartner.com
  8. 8Computerworld, 2024 · computerworld.com
  9. 9CBS, 2025 · cbs.nl