In most companies people already use AI at work, often in personal accounts with company data pasted in. The question is whether that use is visible, controlled, and auditable. Agents raise the stakes, because they send data to models all day without anyone typing it.
Policies are not controls
Policies say sensitive data must not go to public models. In practice a contract clause gets pasted into a chat window, or a customer list is uploaded for a quick analysis.
Another policy document does not fix this. People use these tools because they are useful, and they route around friction. The control that works sits in the path of the data and is easier to use than the workaround.
One control point
We put a gateway between your people and agents on one side and any model on the other. Every request passes through it, so policy is enforced in one place whichever model or provider is behind it.
That is our Privacy Gateway, and it does four things:
- Detect. Scan prompts and attachments for personal identifiers, financial and health data, credentials, and your own internal terms.
- Mask. Redact or pseudonymise what policy says must not leave, so the model sees a safe version and the person still gets an answer.
- Route. Send each request to a model suited to its sensitivity: public models for public data, private or on-premises models for the rest.
- Audit. Log who sent what, what was masked, which model answered, and when.
A simple policy
Most organisations can start with three classes of request:
- Standard. No sensitive data. Sent as is.
- Transformed. Personal or confidential data is masked before sending.
- Restricted. High-risk data. Blocked, or reviewed by a person first.
Add data minimisation (send only what the task needs) and a retention rule for logs, and you have something a privacy team can sign off. For the detail on what to mask, see What to redact before sending data to an LLM.
Deployment options
- Gateway in front of public models. Least friction. People keep the tools they use; what leaves is masked and logged.
- Private models for sensitive work. Models you control, in your cloud or ours, for data that never touches a public API.
- On-premises. For banking, legal, or healthcare: the full stack inside your perimeter.
The right mix depends on your data and your regulators. We usually start with the gateway because it gives visibility within days, and that visibility informs the rest.
Evidence, not intentions
An auditor or a board does not ask whether you have an AI policy. They ask you to show what data went where, who used which model, and what was masked.
The gateway produces that record as a by-product of normal use: usage by team and data class, masking events, a model inventory, and a trace for every request. Under GDPR, that record is the difference between believing you comply and being able to show it.
Limits
No detector catches everything; we tune the balance between false positives and leakage for each organisation. Heavy masking can lower answer quality. And a gateway controls data in transit; it does not undo what already leaked.
To talk through your data map, see private AI and the Privacy Gateway, or book a call.
What goes on autopilot first?
Tell us how you work today. We will tell you where to start.
Book a call