What This Costs You, Both Ways
Here is what we cost, and what the law can cost if you get it wrong. Big fines make the news, but they are rarely what hurts first. Deals get stuck and your team rebuilds proof by hand long before anyone from the government calls.
Price Guide
These are our real prices, not a danger score. We will not make up a dollar number for your risk. Nobody can work that out from three questions. A made up number that looks exact is worse than no number at all.
One job, one price. We put it in writing before any work starts.
Which rules apply to you depends on your job with the AI. Do you make it, use it, bring it in, or sell it on? It also depends on how risky your AI is. We work that out during a check, with your lawyer. A slider cannot do it. This is only a rough guide.
What The Law Says The Fines Can Be
These are the biggest fines under the EU AI Act. They use whichever of the two numbers is bigger. Smaller companies and new companies get lower limits.
Some AI is banned outright. AI that tricks people. AI that takes from people who are weak. AI that gives people a social score. AI that scans faces without asking.
Breaking the rules for risky AI. You did not handle the risk. You did not look after your data. You did not keep records. You did not tell people. Or you did not let a human watch.
Giving wrong, missing, or misleading answers to the people whose job it is to check on you.
For comparison, GDPR tops out at €20M or 4% of the money you make in a year. The AI Act was set higher on purpose. And both apply at the same time. One does not replace the other.
When Each Rule Starts
Some dates were moved in 2026. This is the list as it stands right now.
Rules for the big AI makers
Companies that build the big AI models had to start being open and keeping papers.
Telling people they are talking to AI
You must tell people when they are talking to an AI. This date did not move.
Marking AI work, and banned uses
Words, pictures and sound made by AI must carry a hidden tag a machine can read. The banned uses start on this date too.
Risky AI on its own
The rules for risky AI. Pushed back from August 2026 so the rule writers had more time. The rules themselves did not get any easier.
AI built inside products
AI inside products that already have safety laws. Pushed back from August 2027.
This is a short guide, not legal advice. What applies to you depends on your job with the AI and how risky it is. Check with a lawyer.
What Actually Hurts First
Deals slow down
Big buyers check your safety before they sign. With no proof of your filters, an AI form adds weeks to a deal that was already slow. And it lands at the worst moment, right before they sign.
Your team builds proof by hand
Working out what your AI said, to whom, and with which words, from logs that were never built for it. That is your engineers’ time, not your lawyer’s. And it happens all over again every year.
Good features stay switched off
The most common cost is not a fine. It is a finished AI feature that sits switched off for six months. Nobody will sign off on what it might say to a customer.
Fixing it later costs more
Adding safety to one AI tool is a small change. Adding it to eleven is a whole project. By then each one has grown its own habits. Now you need people and money.
Get a real number, not a guess
Fifteen minutes with your real list of AI tools gives you something you can take to your board.