Private AI
Three things to check before staff ask AI about internal documents
An answer that sounds right is not enough. Before a question-answering tool goes near your policies and files, three checks will tell you most of what you need to know.
1. Is the source shown?
An answer that points to the policy and section it came from can be checked in seconds. An answer that does not asks the reader to trust it.
Ask to see this in practice, with a real question: does the tool show which document and which section produced the answer, and can the reader open it? If the tool cannot show its source, treat every answer as unverified.
Showing the source does not make the answer correct. The tool can still misread a document. What it does is make errors visible to the person who is about to rely on the answer.
2. Does access follow permissions?
Internal documents sit under different permissions for good reasons. A question-answering tool should not become a way around them.
- Can a person receive an answer drawn from a document they could not open themselves?
- How are permissions set: by group, by document set, by individual?
- What happens when a person changes role, and who maintains the mapping?
The test is simple. Pick a document that one group can see and another cannot, ask a question that only that document answers from each group's account, and compare the results.
3. Is there a record?
If something goes wrong, or if someone simply asks how a decision was informed, you will want to know who asked what, and which documents were used to answer.
- Are questions and the sources used logged, and for how long?
- Where are the logs stored, and who can read them?
- Do the logs themselves contain sensitive text, and are they protected accordingly?
Logging has a cost as well as a benefit. Questions can reveal sensitive intent, so decide deliberately what you keep and who can see it.
A checklist you can copy
- Each answer shows the document and section it came from.
- A person cannot see an answer drawn from a document they cannot open.
- Permissions can be reviewed and changed by your own team.
- Questions and sources are logged, and you know where and for how long.
- It is clear to users that answers must be checked against the source.
- You know where processing runs, and whether anything leaves your network.
Where this leads
These checks apply to any tool, whoever supplies it. They are also the three behaviours we built the private AI pilot around: sources shown, access controls, and a local record.
James Ma runs Maandoh, which deploys private and air-gapped AI for regulated organisations in Singapore and Malaysia. He has more than 15 years of experience in fintech and secure software engineering and is the author of three books.