The tension
The most valuable work AI can help with often involves the most sensitive information: client files, financial records, contracts, personnel matters. Keeping AI away from all confidential data limits its usefulness. Letting it reach everything creates obvious risk. The goal is controlled access, not no access.
Why AI changes the picture
Confidential data has always needed access control. AI adds some new features:
- Search is very effective. AI can find documents people forgot were over-shared.
- Data moves to the provider. What the AI retrieves is sent to the AI provider to process.
- Instructions can be manipulated. Content the AI reads can try to redirect it.
- Output can leak. An AI can include confidential detail in a draft meant for someone else.
- Scale. An agent can read far more, far faster, than a person.
A practical framework
1. Classify
A simple scheme is enough to start: public, internal, confidential, restricted. Restricted might include payroll, health information, legally privileged material, security details and anything with special contractual protection.
2. Default restricted to excluded
Keep restricted data out of AI access entirely until there is a specific need and a specific control.
3. Limit by person, agent and task
The AI’s access should not exceed that of the person it acts for, and should be narrower still where the task needs less. An agent preparing a client reminder does not need that client’s full file.
4. Share portions, not wholes
Give AI the fields or records a task needs, not the whole system. Views, summaries and projections are safer than full access.
5. Check where the data goes
Understand your AI provider’s terms for business data: retention, use for training, location of processing and subprocessors. Check your provider’s current settings; terms differ by plan.
6. Control actions, not only reading
Confidential information can leak through what AI does, such as an email sent or a file shared. Require approval for external communication.
7. Record access
Keep a record of what confidential data each agent retrieved, for whom and when.
UK considerations
Personal data within confidential information is subject to UK GDPR, including its security and data minimisation principles. A data protection impact assessment may be appropriate for new AI processing that is likely to be high risk. Regulated professionals also have confidentiality duties under their professional rules. This is general information, not legal advice.
How Kroy approaches it
Kroy enforces access at the point where AI reaches your data. Each request is checked against the person, the agent, the delegation and your policies, and Kroy returns only the projection that request is entitled to. Restricted systems can be excluded entirely. Credentials never leave Kroy, and every retrieval, allowed or refused, is recorded.