Almost every company I talk to has the same blind spot. Shadow AI does not come in through IT. It comes in through the expense report, and the person who approves it owns the consequences.
I sit in a lot of meetings where someone says "we do not really use AI yet". Then I ask who has a ChatGPT subscription on their company card. Half the room. Then I ask what they paste into it. Customer emails, contracts, CVs, meeting recordings. Nobody in the room thinks of that as "using AI". The regulator does.
Here is the part that surprises people every single time: the company is the data controller for whatever its employees process at work, whichever tool they picked. When a manager approves that forty-euro receipt, or just looks the other way, the company has sanctioned the processing. The fine, the breach notification and the broken customer contract all land on the company, not on the person who pasted.
So the question is never "do we use AI". It is "which of the forty subscriptions we approved this year is already a breach". Usually several.
I have watched the same eight failures repeat across companies of every size. They map straight onto the rules that matter.
No processor agreement (GDPR Art 28). A tool that receives work data is legally a processor. Without a written agreement it is unlawful processing, full stop. The free tier of anything has no such agreement. The "team plan" usually does not either.
Data leaving the EU without a safeguard (GDPR Chapter V). "The tool was free and handy" is not a transfer mechanism. Most people cannot say which country their AI note-taker stores audio in.
Special-category data in a general tool (GDPR Art 9). Sick notes, union membership, anything about health or ethnicity. Pasting it into a chat is a serious breach even with a perfect contract.
Keeping everything forever (GDPR Art 5(1)(e)). No retention period, no deletion job. The tool remembers every customer detail across every conversation, for years.
Letting the AI decide (GDPR Art 22). Auto-rejected applications, auto-denied refunds, performance scores that feed straight into pay. Anything with legal or similar effect on a person needs a human in the loop and a right to an explanation.
Pretending it is a person (EU AI Act Art 50). A chatbot with a human name and a stock photo that never says it is an AI. Since August 2026 that silence is a breach, not a design choice.
Shared logins and hard-coded keys (GDPR Art 32). One AI account for the whole team, an API key in a shared script. No traceability, no isolation.
No governance at all (ISO 42001 / 27001). No list of tools, no owner, no training, no incident plan. This is the one that lets all the others happen.
Not a policy document. Three concrete moves, in this order.
The Pocket Operator exists partly because I got tired of telling companies what not to do without being able to show what "right" looks like. So the data page on this site is the working: every rule, what we actually do about it, and the thing itself behind every row. Three companies touch your words, all under agreements. The chatbot tells you it is an AI before it is created. Delete means files erased from disk. Logs go at 90 days, your notes stay until you delete them.
None of that is clever. It is just the right move, done on purpose, and written down.
The compliance page stopped being a claim and became a dataset. If you want the same 120 examples in your own tooling, they are one fetch away as JSON, linked at the bottom of the examples page.
Eight rules, fifteen examples each. Every card opens to show what the right move would have been. Five at a time, so you can stop when you have seen enough.
Join the launch list and you get your personal start link the day it opens.