Matching AI controls to the data at stake

Why start with data rather than tools?

Two companies can use the same AI tools and need very different controls. A marketing team drafting copy with a chat assistant and a payments team using a coding agent next to production credentials face different consequences if something goes wrong. Asking what data is in reach, and what an agent could do with it, sorts the risk faster than listing products.

What fits each level?

The finder uses three levels. They are rough on purpose: most companies have all three somewhere, and the useful answer is the level of the people and teams you are deciding for.

• Mostly public or general internal information: the finder adds three points to policy. An approved-tools list and clear guidance on what not to paste may be enough.
• Customer data or source code: one point each to the browser extension and on-device positions. You want to see what goes into AI tools and to warn or mask before it leaves.
• Regulated data, secrets or production access: two points to on-device control, one each to the gateway and network positions, and the finder starts checking whether products record who or what took each action.

Why does an action record matter for sensitive data?

When something goes wrong near sensitive data, the first questions are who did it and whether a person or an agent took the action. Products describe this differently. Bay states that it records who triggered an action, whether a human was involved, and the chain from prompt to system action. Several others describe audit reports, registries of allowed agents, or integrations with identity and logging tools, which the finder counts as partly matching. An auditor or a customer may ask for this record before they ask about anything else.

What about masking?

For customer data in browser AI, blocking every request with a customer number in it frustrates people and pushes them elsewhere. Masking replaces the sensitive part before the request leaves. Onyx Security lists mask among its five actions, alongside alert, block, steer and ask. Ask other vendors whether they can do the same.

What should you decide before talking to vendors?

1. List the data classes that AI tools can reach in each team.
2. Mark the teams where agents can reach secrets or production.
3. Decide which actions near that data should always need a person's approval.
4. Decide what record you need, and who will read it.

With those four answers written down, the finder's data question becomes easy to answer for each team, and vendor conversations move faster: you can ask each vendor to show what it does for the specific data classes and actions on your list, rather than walking through a general demo. Teams with different levels can also get different settings from the same product, so you do not have to pick one stance for the whole company.

Sources