# Data sovereignty, data control and AI access: what changes? Canonical: https://opendataroom.com/guides/data-sovereignty-ai-access Updated: 2026-10-17 Choosing which AI can access your files gives you an access-control boundary. It does not, by itself, establish data sovereignty. Storage location, processing location, provider access, key custody and applicable agreements are separate questions. ## How OpenDataRoom helps Market: Companies choosing how staff and partners use AI The source collection and access decisions can stay organised independently of a particular AI conversation. Keep source documents organised in drops, like folders, and select named packs for each audience. OpenDataRoom manages access and request records; the recipient uses their own AI for analysis. Current AI tools return extracted text, not complete original files. - [Reusable private packs](/guides/control-data-access-ai): Keep the selected documents together for repeat questions and later reviews, within the displayed retention period. - [Named recipients and separate packs](/guides/separate-packs-per-audience): Choose who can connect. Prepare separate packs when people are allowed to see different documents. - [Your choice of compatible AI](/compatibility): Use the connection settings in an app that supports the required MCP and authorisation flow. The public sample lets you try service information first. ### Example Fictional example: The owner has prepared a three-file pack. ### Workflow impact The source collection and access decisions can stay organised independently of a particular AI conversation. ### Bring these documents - Pack inventory.md - Known-answer test.md ### Prompt for your own compatible AI Using only this approved pack, the source collection and access decisions can stay organised independently of a particular AI conversation. Start with this source question: What should happen before inviting someone else? Cite the file and passage. Distinguish supplied facts, assumptions, missing evidence and contrary evidence. Do not infer human reading or investment intent from request records. Check the sources and the receiving AI app before using the pack for the real assignment. ## Frequently asked questions ### What would this change in my workflow? The source collection and access decisions can stay organised independently of a particular AI conversation. The practical task is to the source collection and access decisions can stay organised independently of a particular AI conversation. ### Which documents should I include? For the example on this page, start with Pack inventory.md, Known-answer test.md. Use the approved versions relevant to the task and include source dates, definitions and any known gaps. Review the prepared text and disclosure scope before sharing. The fictional examples illustrate the structure; OpenDataRoom does not supply or verify your underlying research. ### How do I keep the information current for staff or customers? Assign a document owner, include version and effective dates, and review the content before sharing it. When a source file changes, use the drop update flow and review the affected permissions. After an update is accepted, fresh requests to existing packs serve the current versions of their selected files. Previously retrieved material in an AI chat does not refresh automatically. Newly added files stay private until selected, and changed excerpt selections need review. Access-change notices are queued for affected recipients; provider acceptance is not proof of inbox delivery. Tell recipients which effective date applies. ### How does the recipient connect, and who gets access? Choose your own AI or invite the intended email recipient. For a private invitation, the recipient verifies the authorised email and completes the required approval flow, then adds the connection in a compatible AI app’s settings. Forwarding an invitation alone does not authorise a different email. Permission covers the selected pack; create separate packs when audiences may see different documents. ### Can I keep the files and end someone’s access? Your wallet contains source drops, like folders. Each named pack shares selected files or passages from one drop. The source drop and its access packs have separate deadlines. You can keep the source until its displayed retention deadline while ending a recipient’s future access. Choose a sharing window, revoke it earlier when needed, and approve an extension deliberately. Expiry cannot recall material that an AI app or recipient has already retrieved. ### What happens when I update or remove a file? After you review and accept a drop update, fresh authorised requests receive the current versions of files already selected for their packs. This does not force an AI to fetch the files again or refresh its existing chat context. New files do not join those packs automatically. Review changed selected passages explicitly. Removing a file or its permission stops future retrieval, and access-change notices are queued for affected recipients. Provider acceptance is not proof of inbox delivery. This cannot erase content already retrieved. ### What if my files or AI app are not supported? Check the compatibility guide and try the public service-information sample first. An app must support the required MCP connection and authorisation; pasting a URL into ordinary chat does not install it. Use supported text-based files and inspect the preparation report. Originals are retained in encrypted storage, but the current AI tools return extracted text, not complete original files. Charts, images and layout are not faithfully delivered by those tools. Image-only scans and standalone images are refused; no OCR, audio transcription or spreadsheet recalculation is provided. Current pilot limits are listed on the pricing page. ### What information is recorded about use of the pack? The ledger records requests received by the service, authorised connection references, requested source information and preparation outcomes. Recipients can also deliberately submit a request, comment or note after approving the exact text. The owner sees that message, the verified connection identity and time in Recipient inbox for up to 90 days, unless deleted sooner. Received does not mean read or answered; no agent or automatic research is started. It does not record the entire AI conversation. Search text is supplied by the AI and may differ from the human question. An AI can reuse retrieved material without making another request. Literal search can miss relevant material; neither a search nor a file request proves that the AI reviewed the entire pack. Owners can investigate no-match searches before deciding whether to clarify wording, repair extraction or add information. A no-match result alone is not proof that a document is missing. ### Who can read my files in the hosted connection? Private-library file names and contents are encrypted in the browser with your wallet key. A hosted AI connection uses a separate access copy that you approve and that the service can decrypt for authorised retrieval. The receiving AI gets the content it retrieves and applies its own data policies. An optional desktop connector can decrypt locally, but it requires installation and is not a universal mobile connection. Review the security and compatibility pages before choosing a route. ## Before you share 1. Identify every provider receiving content 2. Separate storage location from processing location 3. Document who controls decryption 4. Check access expiry and deletion separately 5. Keep legal or regulated requirements explicit ## Fictional example documents These are invented teaching documents, not customer evidence. ### Pack inventory.md The owner has prepared a three-file pack. ### Known-answer test.md A test question refers to a paragraph that is visible in the source preview. ## A source question What should happen before inviting someone else? Illustrative answer: Connect with your own compatible AI and check the source question, permissions and response references. ## Ask six separate questions Who holds the source files? Who may request them? Where are they stored? Where is retrieved content processed? Who can decrypt it? What can you export or delete? A single claim of “private” does not answer all six. ## What model independence means here OpenDataRoom is designed to let compatible AI apps consult a permissioned pack. You supply the AI. Named client compatibility must still be qualified. This is freedom to choose the consuming application, not a promise of independence from storage infrastructure. ## Hosted encryption and blind encryption differ The private library encrypts contents and file names in your browser with your key. Hosted MCP access uses a separate copy that you approve and the service can decrypt to answer authorised requests. Provider-blind hosted retrieval would need a trusted recipient-side decryption path; it is not an available option for arbitrary cloud AI clients. ## Be precise about geography OpenDataRoom runs on Cloudflare. The current setup does not offer a contractual country-specific residency or sovereign-cloud commitment. If your project requires those commitments, resolve the storage, gateway and receiving-AI arrangements before supplying sensitive material. ## Use the ledger for the question it can answer The request record can identify a connection, a requested file, excerpt ranges and response references. It does not establish where a downstream AI processed a response or whether the AI retained it. ## Sources and further reading - [Microsoft: digital sovereignty controls](https://learn.microsoft.com/en-us/azure/azure-sovereign-clouds/digital-sovereignty): Further reading; review the source in its own context. ## Related workflows - https://opendataroom.com/guides/control-data-access-ai - https://opendataroom.com/guides/ai-document-connection - https://opendataroom.com/guides/time-limited-ai-access Admission is by invitation. Join the waiting list: https://opendataroom.com/signup Check compatibility and the separate public service-information sample: https://opendataroom.com/compatibility ## Try the fictional example https://opendataroom.com/example-demo.html?example=use-your-own-ai-first