What your AI can access and what information it should never use
Confidential information reaches an AI in two ways: what the AI can read (the documents and systems it is connected to) and what people give it (what they paste into the chat). Governing both starts with one question: what data can each tool use?
The two doors information leaks through
Imagine the AI is a new employee. Through one door it reads what you connect: shared folders, email, databases. Through the other it receives what colleagues tell it. If the first door is wide open or the second has no rules, confidential information ends up where it should not.
Step 1: classify data into a few levels
Four levels are enough to start:
- Public: what is already on your website or in press releases.
- Internal: working documents that are not sensitive.
- Confidential: contracts, pricing, strategy, customer data.
- Restricted: sensitive personal data, trade secrets, financial information before publication.
Step 2: decide which level each tool may use
- Public: any tool.
- Internal: only tools approved by the company.
- Confidential: only tools under a contract that guarantees your data is not used for training and defines where and for how long it is stored.
- Restricted: never in external tools; only in controlled internal environments, with a legal basis and an impact assessment where applicable.
Step 3: tighten the permissions of what the AI reads
An AI connected to your documents usually inherits the permissions of whoever asks, and that is a trap: if the payroll folder is open “by mistake”, the AI will find it in seconds. Before connecting anything, review permissions, apply least privilege and exclude sensitive collections. That is what “not loading the whole SharePoint” means.
Step 4: set clear rules for people
A short AI usage policy answers three things: which tools are approved, which data can never be entered and whom to ask when in doubt. And it must offer an approved alternative: if all there is is a ban, people will use tools on their own (shadow AI).
Step 5: review what you sign with each vendor
Before contracting a tool, ask whether they use your data for training, where it is stored, how long it is kept and which sub-processors are involved. Get it in writing.
Step 6: log and review
Enable logs of what is accessed and by whom, review access periodically and define what counts as an incident and who is notified.
Common mistakes
- Banning without an alternative: AI keeps being used, just out of your sight.
- Connecting the AI to everything “so it works” without reviewing permissions.
- Over-classifying: twelve levels nobody remembers. Four that get applied are better.
- Forgetting personal data: GDPR applies just the same when an AI processes the data.
Start here
- List the AI tools in use today (including unofficial ones).
- Define the four classification levels and which tool accepts each.
- Review the permissions of the folders and systems you plan to connect.
- Publish the usage policy with the approved alternatives.
Frequently asked questions
Is banning ChatGPT and similar tools enough?
No. Without an approved alternative, use continues unchecked. It is better to offer an authorized tool and clear rules.
What should I do first?
Classify data and review permissions. Without knowing what is confidential, you cannot decide what the AI may see.
What data should never be entered into an external AI?
Sensitive personal data, trade secrets, credentials and any information under contractual confidentiality, unless the tool is approved for it.
Informational content, not legal advice.
This gives you the judgment. If you want the finished document: Data Classification and Retention Policy.
What's your Data Governance maturity?
Free assessment with your priority gaps, plus the self-assessment quiz and savings calculator on the Data Governance path.