AI governance · Data handling · Technology selection
What should you establish about data use before choosing an AI provider?
Map the processing route, separate inference from other uses, and verify the controls for the exact service and deployment you intend to operate.
Describe the whole route a request follows
Begin with one useful task and the information it requires. Trace the input through retrieval, application services, the model, logging and the returned result. Include intermediaries with access to the content. A provider name alone does not describe this route.
NIST’s Generative AI Profile recommends examining third-party access and risks across the system’s supply chain. Turn that into concrete selection questions: which organization receives each data type, for what purpose, and under whose operating responsibility?
Identify where the proposed design differs from the customer’s approved environment. Resolve those differences before treating a working demonstration as a deployable product. Keep unnecessary documents and fields out of the initial design.
Separate processing, storage and model improvement
Using data to answer a request, retaining it for application history, inspecting it for abuse and using it to develop a model are different activities. Ask about each one. A statement about one activity does not settle the others.
Microsoft’s documentation for models sold by Azure illustrates the distinction: it separately describes model-training restrictions, storage used by stateful features, abuse monitoring and processing locations. Global and DataZone deployments have different processing boundaries from standard deployments. Preview features can also have different conditions.
For the specific service, record the selected features, deployment type, retention behavior and available deletion controls. Verify the configuration and applicable terms rather than transferring an assurance from another product or deployment. Recheck when the configuration changes.
Make data suitability part of the product decision
Use a purpose-specific review: identify what data the task needs and whether that use is permitted. Have the responsible data, security and legal owners resolve rights and obligations for the intended processing. Technical access to a document does not settle those questions.
NIST also recommends governing third-party intellectual property and training-data use. Keep retrieval and model-development proposals separate so a later experiment cannot silently expand the approved purpose. Reducing detail can reduce exposure, but does not by itself establish anonymity.
Before rollout, test access boundaries and the removal of stored material where supported. Record unresolved constraints as product decisions: change the workflow, choose a different deployment or defer that use. The useful outcome is a service whose actual data behavior fits the task.