Before connecting AI to company data, map where the information goes
Sovereign AI is becoming a practical business discussion about access, continuity, and control. Start with one workflow and follow its information from beginning to end.
An assistant connected to company documents can look simple: ask a question, receive an answer, continue working. Behind the interaction, information can travel through several systems. Understanding that path belongs in the buying decision.
Microsoft's October 5 sovereign-AI article describes a framework developed with NVIDIA around workload-specific control, choice, flexibility, and resilience. Its questions cover data, access, infrastructure, and continuity. It does not establish that every company needs private infrastructure.1
Draw the journey of one request
Take a customer account summary. List the request, customer record, supporting documents, model service, output, and activity log. At each step, ask what moves, who can access it, and how long it is retained and recoverable. Include error logs and support access. A controlled source folder does not by itself prevent sensitive excerpts appearing in a less restricted system.
Separate verified facts from assumptions. A general security page may not describe the exact tier, integration, region, or contract being purchased. Request documentation for the intended configuration. Legal and regulated-data questions need qualified advisers.
Choose controls for the work
A public marketing draft and an internal personnel analysis carry different consequences. Start a trial with approved information and a named owner. Keep retrieval access narrow enough to explain and test.
OpenAI's October 6 Atlassian announcement describes access to organizational project and document context subject to appropriate permissions. Integration design and information access belong in the same conversation.2
Test what happens when an employee changes roles, a document is withdrawn, or a relationship ends. Check revocation across the entire workflow, including cached copies, logs, and retained outputs where applicable.
Check continuity before dependence grows
Which requests can wait during an outage? Which need a manual route? Who notices stopped work? Keep fallback instructions accessible without the failed service and test a realistic interruption.
Four supplier questions
- Where does this workflow process and retain information, including temporary data and logs?
- Who can access it, and how is access inspected and revoked?
- How does essential work continue during interruption?
- What can be exported or reused when changing models or providers?
Anthropic's October 1 Barclays announcement places governance and human oversight alongside operational use cases. Apply that principle proportionately to your organization.3 Start with an accurate map, the next decision, the evidence needed, and its owner.