Working safely with AI and business data

The fear around AI and data is usually about the model itself: does the AI learn from my numbers, is the vendor reading along. Under business terms that is actually the best regulated part. Serious AI vendors process business data without using it to train public models.

The real risks are more mundane: a password shared in a chat message, a connected tool that received far more permissions than needed, an export of the customer base lingering in the wrong place, a former employee whose access was never revoked. If you want to use AI safely, you do not start with the AI but with the basics.

One: integrations run through the company’s own accounts, so the owner always controls and can revoke access. Two: every integration gets minimal permissions, read-only wherever possible. Three: no shadow copies of databases; systems read data the moment it is needed. Four: AI processing only through vendors with business terms, and only the data the feature truly needs. Five: irreversible actions, such as payments or sending messages, always require human approval.

None of these rules is technically complicated. They only require discipline at the moment of setup, which is exactly when most companies are in a hurry.

The final step is being able to show how things stand. Which systems are connected, which data they read, who has access. If you can show that overview at any moment, the security conversation takes two minutes. If you cannot, the conversation never ends.

We have published our own agreements on this in our way of working. Not because we have to, but because trust grows faster when you can be verified.