Blog · August 7, 2026 · By Rickard Eriksson

The receipt principle: AI that shows its working

We don’t believe in AI you have to trust. We believe in AI that shows its receipts: what the answer cost, where it was processed, which sources it stood on, and who approved it. Here’s the principle that guides everything we build.

This image is AI-generated and approved by a person. The machine-readable AI marking sits inside the image file.

There’s a phrase that keeps popping up in almost all AI marketing right now: "trust us." The model is safe, trust us. The data doesn’t leave Europe, trust us. The cost is reasonable, trust us.

We don’t believe in that phrase. Not because the providers are lying, but because trust isn’t an argument. It’s a hope. And an organisation that has to answer to a board, an auditor, or a regulator can’t show up with a hope.

That’s why we build EuroWork around a different principle, and it’s simple enough to fit on a receipt. We call it the receipt principle:

Every time the AI does something for you, you should be able to see what happened, where it happened, what it cost, and who said yes.

Not in an annual report. Not on request. Right there, in the interface, while the work is happening.

What the receipt principle means in practice

Ask a question in EuroWork and look under the answer. There you’ll see which model responded and why, what the answer cost in your currency, and where the processing took place. Run a Deep research and the receipt shows the number of web searches. Let the AI colleague run a recurring task, and every run leaves its receipt in the list: "cost 4 cents." The lead machine labels every contact detail with whether it actually appears in the source or is a conclusion, because a guessed email address that looks found is a lie in table form.

The same principle holds at scale. The organisation’s owner sees the monthly cost per type and per member, sets limits in their own currency, and those limits cap real costs, not estimates. What the AI creates is labelled as AI-generated according to Article 50 of the AI Act. And where every part of the chain is based, down to who owns the server hall, is reported openly.

What the receipt principle forbids

A principle is only worth something if it costs something. Ours forbids things that would have been profitable.

It forbids credits and consumption pools, because a receipt in a made-up currency isn’t a receipt. It forbids managers from seeing what employees ask the AI about: owners see the number of requests, never the content, because surveillance is the opposite of the safety that gets people to actually use the tool. And it forbids silent learning: when the AI colleague remembers something you said, the memory line appears in a list you can read and delete, nothing else.

Why now

The AI industry is moving toward agents: systems that work independently, return on schedule, and make small decisions along the way. We’re building that too, our AI colleague takes on tasks, remembers what you liked, and asks when it’s unsure. But the more independent the AI becomes, the more important the receipts are. An assistant that answers wrong costs an irritation. An agent that works wrong in silence, week after week, costs trust in the whole idea.

So when someone shows you an AI service, don’t ask if it’s good. Ask what the last answer cost, where it was processed, and how you’d show it to your auditor. If you get a receipt, you’ve found the right one. If you get a "trust us," you know the answer too.

Rickard Eriksson

In 1996 Rickard Eriksson created what became LunarStorm, the world's first social medium, and has since trained people from more than 7,000 companies and public-sector organisations in AI. Today he runs EuroWork.