Who sees what: AI alerts, admin visibility and who owns the API key

Most conversations about AI risk focus on what the model gets wrong. In regulated firms, an equally important question is quieter: who can now see information they could not see before, and who controls the account it flows through?

Start with AI alerts. A growing number of systems offer an AI watch that reads projects, tickets or client messages on a schedule and alerts staff when something looks wrong: a deadline slipping, a client conversation turning sour. These are genuinely useful. They also deliver information to people based on rules you may not have read closely.

In one platform we reviewed, a watch on client conversations alerted ordinary staff only about conversations they were already part of. Account owners and administrators were the exception: they were told about any conversation the watch read, including private ones they were not part of. That is a reasonable design, but it is also a change in who sees private client communication, and it belongs in your access policy.

Next, look at AI assistants that can act inside your systems. The better designs limit an assistant to the permissions of the person using it, and ask for confirmation before changing a record. That is a good starting point, but it means your role permissions now govern what the AI can do on someone's behalf. If roles were set up loosely years ago, the assistant inherits that looseness.

Check the assistant's default tools as well. It is common for delete and deactivate actions to be switched on by default. Switch off every tool an assistant does not need, and decide which roles are allowed to use it at all.

Then there is the account itself. AI features usually run on a model provider account and an API key. In practice, that account is often opened by whoever set up the feature: an IT provider, a software consultant or a tech-savvy employee, on their own card and in their own name. Client data then flows through an account the firm does not control, cannot audit and may lose access to when the relationship ends.

The fix is simple. The firm opens and owns the AI vendor account, on a business plan with appropriate data terms, and grants access to outside help as needed. Keys are issued per system, so one can be revoked without breaking everything else. Spending caps are set in the provider's console.

Finally, keep your outside providers out of the approval chain. Your IT vendor can configure an AI feature. The person who approves what it says to a patient or client should be on your staff.

We help healthcare and financial services firms work through exactly these questions as part of managing their IT and compliance. Our self-assessment covers access, vendor management and the basics around them.