New rules are coming for AI. Here's the plain version.
In one breathRegulation is turning “we think it's fine” into “show the record.” The companies that make proof a habit will barely feel it.
If you use AI in anything that touches people — money, health, hiring, safety — the rules are changing. The EU AI Act is the big one, and its obligations have been phasing in since 2025, with more arriving through 2026. You don't need to be a lawyer to get the shape of it.
What it actually asks for
Strip away the legal language and it comes down to a few plain demands: know what your model is. Write it down. Track what changed. Keep the record. For the first time, evidence about your AI is required, not optional.
Notice what's new here. It isn't "make your AI perfect." It's "be able to show your work."
Why this is hard the old way
Most AI today enters production on a description — a paragraph nobody proved. When a regulator, or a customer's security team, asks "what is this model, and how do you know?", a paragraph isn't an answer. Scrambling to produce evidence after the fact — that's where the cost actually lives.
The routine version
It gets a lot cheaper if the evidence is produced as you go: a reading before you adopt a model, a record of what changed on each release, proof attached to each and kept in one place. Do that, and the compliance ask stops being a fire drill — the paper trail assembles itself.
That's the whole idea behind how this house works: produce the evidence a buyer or a regulator will ask for before they ask, in a form they can check for themselves.
The takeaway
The rules aren't asking you to be certain. They're asking you to be able to prove what you claim. Start treating proof as routine now, and the deadline is just another Tuesday.
The longer argument for why this is a category, not a chore → The thesis