The confidentiality question to ask any vendor using AI

Most conversations about AI and compliance stall on the wrong question. They ask whether a vendor uses AI, when the question that actually protects you is where your data goes once they do. For regulated work, confidentiality is not a feature of the model. It is a property of the environment the work runs in, which means it is something you can put in a contract and verify.

The exposure most teams miss

When a delivery partner brings AI into regulated work, the sensitive material does not stay still. A document, a dataset, a draft can pass through systems the client never evaluated and never named in an agreement. The content is the same content your contract protects. The path it travels is new, and it is the path, not the tool, that creates the exposure. A confidentiality obligation does not care whether a person or a model read the file. It cares where the file went and who could see it.

Why "we use enterprise AI" is not an answer

"We use enterprise AI" sounds like an answer and is not one. The brand of the assistant tells you nothing about where your data lives, whether your material was exposed to a third-party model, or who kept a copy and for how long. Two vendors can both say it and run completely different risk profiles. The questions that matter sit a layer below the product name. The NIST AI Risk Management Framework makes the same point in plainer terms: trustworthy AI is governed by how data and third-party components are managed, not by how capable the model is.

What a defensible posture looks like

A defensible posture has three properties, and none of them are about the model. The material is handled inside an environment the partner controls, rather than handed to services chosen for convenience. Access is limited to the people accountable for the engagement, so "who can see this" has a short, nameable answer. And the commitments are the kind a client can verify rather than take on faith, because they are written into the agreement and reflected in how the work is actually run. Confidentiality built this way is a property of the practice. It holds whether or not AI is involved, which is exactly why it survives the arrival of AI.

The questions to put to any vendor, including us

If you are evaluating a partner for regulated work, a short list separates a real posture from a reassuring sentence:

• Where does our material live while you work on it, and does it ever leave that environment?
• Is any of our content exposed to a third-party model, retained, or used to improve a service?
• Who, by role, can reach the work, and how is that access limited?
• Which of these commitments are in our contract, and how would we confirm they are being kept?

Ask us the same questions. The point of the list is that the answers should be specific, checkable, and consistent with what the agreement already says.

Compliance as posture, not paperwork

This is not a feature anyone added. It is the shape of how regulated work has to be done, and AI did not change the obligation, it only raised the cost of ignoring it. A partner who treats confidentiality as the environment the work runs in, rather than a promise made next to it, gives you something you can hand to counsel and defend. The work moves faster, and the protection your agreements require is built into how it moves.

Source: NIST AI Risk Management Framework (AI RMF 1.0).