Why is an acceptable-use policy not enough for AI data governance?

A policy cannot intercept a paste. Only a control in the path can classify the content and mask or block it before the prompt leaves the browser.

What is Named Entity Recognition in this context?

Machine reading that identifies people, organizations, and locations in ordinary prose, catching regulated content that pattern matching on formats like card numbers would miss.

Does logging AI traffic create a second copy of sensitive data?

It should not. A well-built control records the classification and the enforcement decision, not the sensitive payload itself.

AI Data Governance: Buyer's Guide | RAIC

THE PROBLEM IS A PASTE, NOT A BREACH

Nobody exfiltrates your customer database. Someone pastes a support ticket into a chatbot to get a faster reply. Someone drops a spreadsheet of patient identifiers into a summarizer. It takes four seconds, it feels helpful, and there is no record that it happened.

An acceptable-use policy does not intercept that moment. Only a control in the path does.

WHAT IN-LINE ENFORCEMENT MEANS

Classify what the traffic contains: before a prompt leaves the browser, inspect it for regulated content — health identifiers, payment data, credentials, personal data under GDPR.

Decide by policy, not by vibes: allow, mask, or block, chosen per data class and per destination. A model your legal team approved gets different treatment than a free tool nobody has reviewed.

Mask instead of blocking where you can: blanket blocking teaches people to route around the control on their phones. Replacing an identifier with a token preserves the employee's answer and removes your exposure — the outcome you actually want.

Keep the payload out of the open: the control must prove it ran without creating a second copy of the sensitive data in your logs. Record the classification and the decision, not the content.

WHY PATTERN MATCHING IS NOT ENOUGH

Regular expressions catch card numbers and national ID formats. They do not catch a patient named in a sentence, a contract counterparty, or a project codename that is confidential by context. Named Entity Recognition — machine reading that identifies people, organizations, and locations in ordinary prose — closes the gap between structured identifiers and how people actually write.

Buy both. Patterns are precise, entity recognition is contextual, and regulated data arrives in both shapes.

THE FRAMEWORK TIE-IN

EU AI Act Article 10 puts obligations on the data used with AI systems. HIPAA cares about disclosure of protected health information regardless of the channel. SOC 2 confidentiality criteria ask what stops information reaching parties who should not have it. In every case the auditor's question is the same: what enforces it, and how do you know it ran?

An in-line control with a decision log answers both halves. A policy PDF answers neither.

THE BUYER'S CHECKLIST

Coverage: which browsers, and what happens on unmanaged devices?
Timing: does classification happen before the prompt leaves, or after?
Actions: allow, mask, and block — or only block?
Detection depth: patterns only, or entity recognition as well?
Privacy of the control itself: is the payload stored anywhere to make the log work?
Destination awareness: can policy differ by tool and model?
Evidence: does each decision produce a record an auditor can trace?

WHAT GOOD LOOKS LIKE SIX MONTHS IN

Employees still use AI, because you did not take it away. Your masked-event count is high and your blocked-event count is low, which means the controls are working with people instead of against them. And when your auditor asks how you keep regulated data out of AI tools, you show decisions, not intentions.

AI Data Governance: A Buyer's Guide