The founder on shadow AI: everyone does it, no one talks about it

Rickard Eriksson is the man who in 1996 built what became LunarStorm, the world’s first social medium, with millions of users long before Facebook existed. Since then, he and his company have kickstarted AI development for more than 7,000 Scandinavian businesses and public sector organisations, seeing up close what happens when technology arrives before the rules do.

Today, he runs Skräddarsydd AB in Varberg and is the founder behind EuroWork. We sat down to talk about why he built an AI workplace for European organisations, and why he believes the most important feature is the one that shows what an answer costs.

You say that shadow AI is the real problem. What do you mean?

— It’s already being used. At every workplace I visit, there’s someone pasting a draft into ChatGPT to make it look better. Often several people. They do it on their personal login because that’s what they had, and they do it because they want to get their work done. It’s not disobedience, it’s ambition.

The problem is that no one knows what went out. And when no one knows what went out, no one can answer the question when it comes.

Who asks the question?

— Sooner or later, someone other than you. A customer wondering where their data went. An auditor. A safety representative. Or a journalist, if you’re unlucky. And then it’s not the employee who has to answer, it’s management. Saying "we have a policy" isn’t enough if the policy is a document no one has read and nothing enforces.

But most places must have some kind of rule?

— Many have a rule that says no. That’s the worst version, because then usage continues but becomes invisible. You’ve moved the problem from being unmanaged to being unmanaged and secret.

I saw it clearly at one place where they had banned AI completely. Everyone still used it. The difference was that now they did it on their phones instead of their work computers, so it wouldn’t show. They’d managed to make it just as risky and harder to see.

And work and private life get mixed up?

— Yes, and that’s the part people underestimate. When someone uses their personal account, the work material is in an account the employer doesn’t own. The person changes jobs, and the material goes with them. The person loses their password, and no one can access anything. And in that history, you have both the customer quote and the question about their child’s school choice, in the same thread, with a provider on the other side of the Atlantic.

It’s not malice from anyone. It’s just what happens when you don’t give people a tool that’s their work tool.

So EuroWork is the alternative?

— EuroWork is an attempt to make the legal option the easiest one. That’s the whole idea. If the secure way is more complicated than the insecure way, the secure way loses, every time, in every organisation, no matter what the boss said in a meeting.

What specifically makes it easier?

— It looks like what people are already used to. You write a question, you get an answer. No one should need a course to get started. The difference is in what surrounds it: answers are processed in Europe by default, the content is never used to train any model, and it’s displayed under every answer so you don’t have to dig through a contract.

And the cost?

— That’s my personal hobbyhorse. No one else shows what an answer costs. You buy a licence, and then it’s silent, and only the following month does the finance department see a lump sum.

With us, the price is shown before you send the question. A balanced answer from a European model costs a few pence. If you choose a top US model, the same question costs forty-four pence. There’s nothing wrong with choosing the expensive option—sometimes it’s worth it—but you should know beforehand, not after.

Why do you care so much about that detail?

— Because it’s the honest version of what you’re selling. Everyone says they’re transparent. Very few show a number that can be used against them. I want a finance director to be able to sit with our service and calculate, not guess.

You’re also building an agent mode. Isn’t that the opposite of control?

— That’s exactly what I was afraid of. That’s why the default setting is that the AI asks before it does anything. You give it a goal, it suggests a step-by-step plan, and you approve one step at a time, seeing which tool will be used and why. Everything ends up in the audit log.

I think that switch is the whole difference between a municipality being allowed to use agents and not being allowed to. The capability exists for everyone now. What’s missing is proof of who said yes.

What’s the hardest part of building this?

— Resisting. Every week there are ten more things that could be built, and nine of them make the product worse because they make it harder to understand. The one who adds everything ends up with something that requires a consultant to get started, and then you’ve built exactly what you wanted to replace.

What do you hope someone will say in a year?

— That they stopped thinking about it. That the question "Can I use AI for this?" is no longer asked at their workplace, because the answer is visible in the tool while you work. And that the finance director knows what it cost, down to the penny.

Give me something concrete. What do people do with it, on a completely ordinary Tuesday?

— The funniest thing is that the answer is rarely what you expect. Everyone guesses "writing texts". What actually saves the most time is not having to read. You get a 300-page annual report, you say who you are, and you get back what’s relevant to your role. Or a procurement request: attach it along with what you’ve delivered before and ask straight out if it’s worth three days of work. We tried this with a document that deliberately didn’t meet two of the mandatory requirements, and the answer pointed them out exactly.

— Then there’s the contract case. Two versions of the same purchase agreement, where the other party has removed a clause about staff responsibility and shortened the warranty period. Both came back. That kind of change slips past when you’re reading through it at seven in the evening.

What about taking photos of things?

— That’s the feature people laugh at first and use the most afterwards. Take a photo of the whiteboard after the meeting, ask for a project plan with responsible parties. Take a photo of a handwritten document from 1897, get it interpreted with the uncertain words marked. Take a photo of a manual in Finnish. It sounds like a toy until you’re standing in a storeroom with a gadget you don’t know the name of and therefore can’t search for.

What’s the most common disappointment?

— When people ask for numbers without providing any material. Then the model can invent a source that sounds completely credible. We tried asking for counterarguments to a debate article, and the answer was based on a report that doesn’t exist. The reasoning was useful, the numbers were made up.

— That’s why I always say the same thing: give it a document. With material, the numbers are correct. Without material, it’s a well-formulated guess. That’s not a flaw in EuroWork, it applies to all AI, but we’re obliged to say it out loud instead of letting people discover it at a committee meeting.

Final question: what should the person reading this do today?

— Ask your colleagues what they actually use. Not in a formal setting, just ask. The answer is usually more than you thought, and it’s always better to know it yourself than to find out from someone else.

---

About Rickard Eriksson

Created in 1996 what became LunarStorm, the world’s first social medium, with millions of users years before Facebook started. Together with his company, he has trained and developed AI usage for more than 7,000 Scandinavian businesses and public sector organisations. Today, he runs Skräddarsydd AB in Varberg and is the founder of EuroWork.

Same principle then as now: technology that ordinary people actually use, and look forward to opening every morning. The company behind LunarStorm was called LunarWorks. This project is called EuroWork.

Contact: rickard@eurowork.ai.