Before providing private business data to an AI application, map where the data goes, who can access it, what is retained and how derived copies are handled. Evaluate the complete service rather than the model name alone. Hosting location, access controls, provider terms, support access, logging and deletion arrangements all need evidence appropriate to the intended workflow.
Let’s plan your private-data AI project.
Bring your workflow and the decisions you need to make. We’ll help you define the requirements, examine the trade-offs and plan a practical next step.
Draw the actual data path
Include upload storage, preprocessing, extraction, retrieval indexes, model requests, outputs, logs, backups and support tools. Mark which organisation operates each component and which data leaves the business-controlled environment.
Derived data matters. Extracted text, document fragments, embeddings and summaries can carry information from the original records. Ask how these copies inherit access rules and what happens when the source is corrected, access-restricted or removed.
This is a technical procurement questionnaire. It does not determine which legal requirements apply to your organisation or establish that any vendor is compliant.
Ask for evidence at each boundary
| Question | Evidence to request | Unresolved issue to record |
|---|---|---|
| Which data is transmitted? | Field-level flow and example request | Unnecessary information included |
| Who operates each component? | Service and subprocessor inventory | Undocumented external dependency |
| How is data used? | Applicable product terms and configuration | Training or other secondary use unclear |
| Who can retrieve records? | Access design and representative tests | Cross-user or cross-account exposure |
| What is retained? | Storage, logging and backup policy | Different retention across copies |
| How are changes propagated? | Correction and removal workflow | Stale derived records remain available |
| Who can support the service? | Support access and audit arrangements | Broad or persistent privileged access |
Do not accept an unrelated product's terms as evidence for the exact service being proposed. Record the applicable product, configuration and contract version.
Compare a concrete document workflow
Suppose an internal assistant uses commercial proposals. A document may include customer contacts, pricing and internal notes. The architecture review should explain which fields reach a model, which stay in retrieval storage and which appear in logs.
Now remove one employee's access to a proposal. Test whether that person can still retrieve a cached answer, shared conversation or derived summary. Source-folder permissions alone do not prove that every copy follows the new boundary.
This is a hypothetical test scenario. It describes evidence to request, not a claim that a particular provider exposes or protects these records.
Separate privacy, security and operating choices
Encryption, isolated infrastructure and restricted access address parts of the technical design. They do not by themselves answer whether data use is appropriate, retention is justified or staff understand the workflow.
The NIST Privacy Framework offers privacy risk-management context. Use it to organise questions alongside your organisation's policies and relevant professional advice, not as a substitute for reviewing the actual data flow and terms.
Plan correction, removal and exit
Name the owner of a removal request and the systems it affects. Ask how backups, audit records and contractual retention are treated rather than promising instantaneous erasure from every copy without evidence.
Specify usable export formats, index rebuild responsibilities and how service access ends. Test the procedure with authorised sample records. A written promise becomes more useful when the team can demonstrate the path and explain its limitations.
Questions procurement teams ask
Does private hosting answer every data concern?
No. Access, secondary use, retention, monitoring and internal operating practices still matter. Evaluate the complete arrangement.
Can we rely on a model provider's public statement?
Use it as one input. Confirm the statement applies to the specific product, terms and configuration being purchased.
Prepare the vendor discussion
Bring a data inventory, permitted users, example records and required retention decisions. Pair the questionnaire with AI supplier selection and discuss AI development using the complete service boundary.
