Enforcing Policy on Azure OpenAI On Your Data Requests That Pull Customer Records
Azure OpenAI On Your Data can retrieve customer records from sources such as Azure AI Search and return them in a grounded response. Native Azure controls decide who and what can reach the data, but not whether a specific request may surface specific fields. To close that gap, place a policy enforcement point in the request and response path. It evaluates each interaction and applies one outcome: allow, redact, block, rate limit or escalate for human approval. Each decision should produce a durable record.
What each enforced decision should record
- Authenticated user identity and role at the time of the request
- Data source or index addressed and the query parameters used for retrieval
- Citations or documents returned with the response
- Customer-record fields detected in the request and in the response
- Policy identifier and version evaluated
- Action taken (allow, redact, block, rate limit, escalate), with timestamp and approver where applicable
What native Azure controls cover and where gaps remain
Entra ID and RBAC
Covers who can call the service and manage resources. Gap: does not decide whether a given request may retrieve given customer fields.
Index and document-level access
Covers which documents a caller can see in Azure AI Search. Gap: a permitted document may still contain fields that should be redacted in a particular response.
Managed identities
Covers how the service authenticates to the data source. Gap: says nothing about the content of an individual request or response.
Network isolation
Covers where traffic can flow. Gap: does not evaluate what is inside permitted traffic.
Content filtering
Covers content safety categories. Gap: does not apply your policy on customer-record fields, volume or escalation.
Standard application logs
Record that a call occurred. Gap: they are not a decision record of which policy was applied and what action followed.
Where customer records can enter prompts and responses
On Your Data grounds model output in enterprise data sources such as Azure AI Search indexes and storage. A single request therefore follows four stages: the user prompt, retrieval from the connected data source, grounded generation, and the response returned to the caller. Customer record fields can surface at more than one of these stages.
The prompt itself may contain customer identifiers or ask for a specific person's data. Retrieval can return documents or chunks that contain account, contact or other record fields. Those chunks become context for generation, so the model can quote, summarize or recombine them. The response, including any citations, is the final point where the fields leave the system boundary and reach a user or downstream application.
The practical consequence is that a request which is fully authorized at the identity and index level can still return more customer data than the interaction warrants. Whether that is acceptable depends on the user, the purpose, the fields involved and the volume, which are per-interaction attributes.
Placing enforcement in the request and response path
NIST SP 800-53 Rev. 5 describes information flow enforcement as comparing the security and privacy attributes of information, and of its source and destination, against a flow policy, then responding when a flow is not explicitly allowed. It notes that explicit attributes can be used to control the release of certain types of information. That is a useful model here, and it is a different control from access authorization.
In practice, put an enforcement point in the execution path between the calling application and Azure OpenAI. On the request side, it evaluates the caller's identity, the data source being addressed and the content of the prompt before the call proceeds. On the response side, it inspects the generated answer and any returned citations for customer-record fields, then applies the policy outcome before anything reaches the user.
This placement means policy is enforced at runtime, not reviewed after the fact. Reviewing logs later can show that a disclosure happened. An in-path decision prevents it, or shapes it, before the response is delivered.
Define policy outcomes and the attributes that trigger them
Decide the outcomes first, then the attributes that select each one. Thresholds are an organizational decision and should follow your risk assessment. The following mapping is a reasonable starting structure.
- Allow: The user, data source and detected fields match an explicit policy. Pass the response through and record the decision.
- Redact: The request is permitted but the response contains fields the user should not see, such as direct identifiers. Mask those fields and deliver the remainder.
- Block: The flow is not allowed by policy, for example a request for another customer's full record from a role without that purpose. Return a refusal.
- Rate limit: Individually permitted requests show a pattern of bulk retrieval. Throttle by user, source or time window.
- Escalate: The request is sensitive but potentially legitimate. Hold it for human approval and define whether it is released, redacted or denied on timeout.
Retention, governance and where Trussed fits
Decision records are evidence, so retention needs a deliberate setting. 16 CFR Part 314, which requires a written information security program based on a risk assessment, also calls for periodic review of the data retention policy to minimize unnecessary retention. As of November 2026, confirm with counsel whether the rule applies to your organization. Audit needs and minimization pull in opposite directions, so record the decision metadata you need and avoid storing full customer content in the evidence store unless required.
Align runtime controls with what you tell customers. The FTC has stated that it takes enforcement action when companies fail to keep privacy and security commitments, including for AI services. Also review the data-handling terms with your AI service provider, since data sent through an API may be governed by provider policies and business agreements.
Trussed is one option for the enforcement layer. It enforces policy in the execution path on individual interactions and tool calls, with allow, redact, block, rate limit and escalate for human approval. Enforcement produces audit-ready evidence mapped to frameworks including HIPAA, NYDFS and NIST. It can be deployed self-managed in the customer cloud, with configurable retention and data residency. It complements native Azure access controls rather than replacing them.
Enforcement points in the On Your Data flow
Label: User prompt
Description: Evaluate identity, intent and request attributes before retrieval.
Label: Retrieval
Description: Native index and document-level access limits what can be returned.
Label: Grounded generation
Description: Retrieved fields may enter the model context.
Label: Response
Description: Detect customer-record fields and citations, then allow, redact or block.
Label: Decision record
Description: Capture the policy applied and the action taken.
Review your On Your Data enforcement design
Walk through your request flow, policy outcomes and decision record fields with a Trussed specialist.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.