Where Does Your Business AI Send Its Data? A Buyer's Checklist
Questions French and German businesses can ask AI vendors about hosting, subprocessors, retrieval permissions, retention and practical exit plans.
Business owners, IT and procurement teams
6 min read
2 November 2026
3 sources
Map
Sources + purpose
→Locate
Providers + regions
→Check
Reuse + retention
→Test
Permissions + deletion
→Exit
Export + ownership
Hosting location is one part of the processing and access picture.
A server location is one useful answer. It is not a complete explanation of who can access the information or what happens to it next.
- 01Map prompts, retrieved documents, logs and backups separately.
- 02Ask who processes data and whether it is reused.
- 03Test access, deletion and export before committing sensitive work.
Follow one document through the system
Imagine a consulting firm in Lyon asking an AI assistant to summarise a confidential client contract. The document may be stored in one region, indexed by another service, sent to a model endpoint, logged for monitoring and copied to a backup. A single 'hosted in Europe' badge does not describe that whole path.
Draw the route for one actual task. Identify the organisation responsible for each stage, the data sent, where it is processed and retained, who can access it and how deletion is propagated. Include subprocessors and support access, not only the main application vendor.
Ask precise questions about reuse
Can the provider use prompts or uploaded files to train or improve its models? Are those settings different for a trial, a paid account and an API? What metadata remains in logs after a file is deleted? What happens when the provider changes subprocessors? Ask for contractual terms and operational evidence, not just a reassuring slide.
CNIL's guidance discusses the different considerations for on-premise and remotely hosted generative AI, particularly when personal or strategically sensitive information is involved. On-premise deployment may reduce some third-party exposure but still requires the organisation to secure, patch and operate the system.
Check who can retrieve what
A knowledge assistant should apply a user's permissions before it retrieves a client file. Test a person who can access one project but not another, a former employee whose access has been revoked and a document moved to a restricted folder. Verify that stale search indexes and cached answers do not preserve access that should have ended.
Agentic systems add another boundary: permission to read a record does not imply permission to email it, change it or send it to a third-party tool. CERT-FR recommends restricting tools and isolating sensitive actions because untrusted content can attempt to redirect an agent.
Plan the exit before the pilot
Ask for usable exports of source files, metadata, permissions and activity logs. Test how an administrator revokes access and how long data remains in backups. Agree who owns the index, any fine-tuned artefacts and the cost of switching providers.
The right hosting and contract depend on the data, purpose and applicable obligations. A small trial using approved non-sensitive material can test the workflow while the organisation completes its broader legal and security assessment.
Use this before approving the work.
- 01Map processing locations, providers and subprocessors.
- 02Confirm retention, reuse and deletion terms.
- 03Test retrieval permissions and revoked access.
- 04Practise export and exit with real administrators.
Research and references supporting this article. Interpretation and recommendations are Thalenor’s.
The firm and contract are constructed. The article is procurement guidance, not a determination of compliance.
Published and reviewed on 2 November 2026. This article provides general technology and operational information, not legal, tax, medical or regulatory advice.