Using AI with confidential data: essential controls
Pasting a document into an AI service creates data processing. Before choosing a model, establish what enters, where it travels, how long it remains and who can access it.

The short answer
Do not send sensitive material to a consumer service without an approved framework. Classify and minimise data, use an approved or local environment, check contracts and logs, then test with fictional data before production.
- Minimise before transfer
- Validate the complete data path
- Test with fictional data
Build a controlled data path

AI-generated illustration 1. Classify information
Separate public, internal, confidential, personal and legally protected data. Give each level a clear rule: allowed, anonymise first, dedicated environment or prohibited.

AI-generated illustration 2. Minimise inputs
Send only necessary extracts. Replace names, numbers, addresses and identifiers with fictional values where meaning can be preserved. Remember that reversible pseudonymisation is not anonymisation.

AI-generated illustration 3. Map the journey
Document the application, model provider, extensions, connectors, logs, backups and subprocessors. Risk may sit in a peripheral integration rather than the main model.

AI-generated illustration 4. Check the framework
Read the terms and contractual documents for the exact plan: training use, retention, processing locations, deletion, access control and incident notification. Marketing language is not a binding commitment.

AI-generated illustration 5. Choose the architecture
Depending on risk, use a professionally governed service, private instance, local model or no AI processing. Local operation reduces some transfers but does not remove configuration, logging or endpoint risks.

AI-generated illustration 6. Monitor and delete
Limit accounts, enable useful audit logs, define retention, test deletion and review access. Assign an owner and an emergency stop procedure.
Trace the actual data path
Use a fictional file and record what the service receives at each step: the original, extracted text, passages sent to the model, response and logs. For every copy, check who can access it and what deletion covers. Review the plan terms for the provider’s permitted uses. If the observed behaviour differs from the approved workflow, keep testing with fictional data and record the discrepancy.
Put the method to work

Practical case
A team wants to analyse an internal contract. First prepare a fictional document with the same structure but no real names or amounts.

Evidence to keep
Map who receives the file, retention settings, administrator access and deletion or export procedure.

Make the decision
Move to the real contract only when these points are documented and compatible with your organisation’s rules.
Four checks before deployment

Necessity
Is each data item essential to the expected result?

Journey
Which services, countries, people and logs receive it?

Commitments
What do the contract and selected plan actually guarantee?

Control
Can you restrict, audit, export and delete?
Local, open and professional options
The directory helps explore different architectures. Inclusion is not legal or security approval: verify the exact offer, documentation and your own configuration.
ChatGPT
general assistant
OpenAI · US
Visit official siteClaude
long-document analysis
Anthropic · US
Visit official siteGemini
Google Workspace assistance
Google · US
Visit official siteMicrosoft Copilot
office assistance
Microsoft · US
Visit official sitePerplexity
search with sources
Perplexity · US
Visit official siteMistral Vibe (Le Chat)
European conversational assistant
Mistral AI · FR
Visit official siteHow is this selection produced?
These profiles are linked to the guide’s subject in our editorial catalogue. Inclusion is neither a ranking nor a performance measurement on your documents. Methodology.
Related tool families
Frequently asked questions
Is a local tool always confidential?
No. It can reduce external transfers when processing is truly local, but endpoint security, logs, backups and access controls remain essential.
Can a document be anonymised automatically?
Software can help find identifiers, but context may still permit re-identification. Human review and risk-based policy are required.
What should be tested first?
Use fictional data to verify permissions, logs, exports, deletion, model failures and the shutdown procedure before real data is introduced.
The references below expand on the concepts and checks discussed. Scenarios and trial frameworks remain editorial proposals; provider documentation describes its own product rather than an independent benchmark.
Source note
Write a usage policy before opening a tool to the team
CNIL recommends documenting what colleagues may do with an AI system, how to check results and which confidentiality precautions apply. It also advises against entering confidential, personal or strategic data into a service without an appropriate framework for that use.
Turn this into a procedure: permitted data classes, acceptable prompt examples, an approval owner and a reporting channel for accidental disclosure. Test the policy on a realistic task, such as summarising a customer file, before treating a general instruction not to share sensitive information as sufficient.
Replace real files with fictional examples of the same structure, without names, account numbers or identifying details. Removing only a name may leave many identifying clues. Have the example reviewed before offering it as a reusable team prompt.
Plan what to do after accidental entry: stop further sharing, record the tool and context, and notify the designated owner. Avoid copying sensitive data into a broad reporting channel. Enable prompt assessment without causing another disclosure.
- Example
- Train the team using a representative fictional file.
- Verification
- Have standard prompts checked before reuse.
- Incident
- Provide a reporting process that limits further copies.
Does removing a name automatically anonymise a document?
No. Dates, places, roles and events can still identify someone in combination. Review the whole document and prefer fictional examples for trials without an approved framework.
Try the reasoning
A free tool accepts documents, but internal rules say nothing about customer data. What prerequisite is missing?
Reasoning: Define permitted data, owners and processing conditions before testing on real information.
Primary source : CNIL — AI factsheet 4: usage precautions



