AI governance that executes
Follow a question that should not be answered.
Someone in sales asks for the payroll. See exactly where the policy steps in, and what is left on record after it does. For the concept, see the AI governance guide.
The question arrives with an identity
It is not an anonymous request. It comes with a.ribeiro, area comercial, and the stated purpose.
The agent resolves what was asked
"Payroll" becomes the concrete resource rh.folha_pagamento. It has not touched any data yet.
The policy answers before the database
The resource is outside the scope of the sales area. The query is never even built. It is not the model choosing to be polite: it is the layer stepping in.
The denial becomes a line in the trail
DENIED · rh.folha_pagamento · limit by area · Aug 19 14:38
Without that line, an audit cannot tell a rule that worked from a rule that was never tested.
Personal data comes out hidden, with its type in plain sight.
When the query is authorized but the result carries personal data, it does not vanish: it comes out marked. The reader understands there is a value there and why they cannot see it, which is different from a field disappearing with no explanation.
In the original document
A. Ribeiro
CPF 123.456.789-00
CREA-SP 123456
How it reaches the person who asked
[personal data hidden: name]
[personal data hidden: CPF]
[personal data hidden: registration]
It applies to documents, not just database columns. Contract excerpts, reports and attachments go through the same filter, and those with authority still see the value.
Does your policy stop anything, or does it only describe?
Context Diagnostic: a conversation with your rules applied at runtime and, if it makes sense, a day with your team.
Talk to an AI expert →