Azure OpenAI Allocation Across Health System Departments
Treat the allocation as an enforceable policy, not a budget line. Define each department's share, approved models, data handling rules and limits, then enforce them on every interaction and tool call in the execution path. Retain per-interaction evidence showing which department used which model under which policy, so the record supports HIPAA 164.312(b) audit controls review.
Audit-ready evidence to retain per department
- Department and the user or application identity that made the request
- Model and deployment identifier, including region
- Policy version in force at the time of the interaction
- Action taken: allow, redact, block, rate limit or escalate
- Approver identity and decision when an interaction was escalated
- Timestamp, plus tool calls attempted and their outcomes
Why platform capacity constructs do not divide a commitment by department
Azure OpenAI capacity is organized around model deployments, quota and, where purchased, provisioned throughput, all placed in specific regions. These constructs describe capacity for a deployment. They do not describe a hospital department. A caller with valid credentials and access to a deployment consumes whatever capacity that deployment has, regardless of whether the caller works in revenue cycle or clinical documentation.
This shapes the available designs. One shared deployment per model is simple to operate, but nothing stops one department's burst from exhausting capacity another department depends on. Separate deployments per department give isolation, but they fragment the commitment, multiply configuration to manage, and leave unused capacity stranded in one place while another is constrained. Regional placement adds a third variable, because where a deployment runs is also a data residency decision for PHI.
The practical conclusion is that the platform sets the outer boundary and a separate layer has to express the internal division. Without that layer, allocations live in a spreadsheet and are honored only by convention.
Define the department allocation model
Start with the attributes that change how a department should be treated. The most useful are PHI exposure, clinical versus administrative use, research versus operational use, and workload pattern, such as interactive requests versus scheduled batch work. These are design judgments, and they should be recorded in the organization's risk analysis. HHS guidance says audit controls should be scaled to the entity's risk analysis and technical infrastructure, so the department tiers you define become the reference for how much control and evidence each one needs.
Express each allocation as more than a share of capacity. A workable definition includes a baseline limit, a burst rule, a priority relative to other departments, and a defined behavior when the limit is reached. Clinical documentation may justify higher priority during clinic hours. A research group working with de-identified data may tolerate lower priority but need access to a different set of models.
Identity must be part of the model. The access control standard at 164.312(a)(1) allows access only to persons or software programs that have been granted access rights. Applications and agents that call models are software programs in that sense. Give each department and application its own identity, such as a service principal or equivalent, and treat that identity as the holder of access rights. Maintain an inventory of LLM use cases by department so each identity maps to a documented purpose.
Map each department to models, data rules and limits
Write the mapping as a policy table that a reviewer can read and a runtime system can enforce. The entries below are illustrative starting points. Set the actual values from your risk analysis and clinical governance.
- Clinical documentation: High PHI exposure and clinical risk. Approved models limited to deployments in the approved region. PHI permitted to approved deployments only. Highest priority during clinical hours. Clinician review remains required for output used in the record.
- Revenue cycle: PHI present in claims and coding work. Narrow set of approved models and tools. Redact identifiers not needed for the task. Rate limits sized for batch and interactive work separately.
- Research: Separate identity and deployments from operational use. Block PHI unless an approved protocol applies. Escalate for human approval before any new model or tool is used.
- Patient access: Patient-facing interactions. Redact PHI not required for the request. Block tool calls that write to clinical systems. Rate limit per session to contain misuse.
- IT: Lower PHI exposure but broad tool access. Restrict tool calls and code-related actions. Allow wider model choice in non-production deployments only.
Enforce at runtime: subscription, application or execution path
There are three places to enforce a department allocation, and they differ in what they can see. Subscription-level settings are coarse. They govern deployments and quota but cannot distinguish departments sharing a deployment, and they cannot act on the content of a request. Application-level enforcement can apply department rules, but every application has to implement them consistently, and a new or unreviewed application can bypass them. Enforcement in the execution path applies one policy to every interaction and tool call regardless of which application originated it.
The tradeoff is operational. A policy layer in the path becomes a dependency for every LLM call, so it needs availability targets, a defined failure mode and a change process for policy updates. Decide in advance whether a policy outage fails closed or fails open for each department, and write down the reasoning.
At this layer, each interaction receives one of five outcomes: allow, redact, block, rate limit, or escalate for human approval. Tool calls are evaluated the same way, so a department approved for a model is not automatically approved for every tool that model can invoke. Trussed enforces policy in the execution path in this manner, at the level of individual interactions and tool calls, and can be self-managed in the customer's cloud with configurable retention and data residency.
Plan for exhaustion and urgency. The access control standard includes a Required emergency access procedure. For LLM-dependent clinical workflows, define what happens when a department reaches its limit or a policy blocks a clinically urgent request. An escalate-for-approval path with a named approver is one option, and the approval itself becomes part of the record.
A practical rollout sequence
Begin with the inventory and the risk tiers, because every later decision depends on them. Next, give each department and application a distinct identity and route its calls through the enforcement layer, initially in a mode that applies policy to a single low-risk department. Review the evidence it produces with your compliance team before extending to departments with high PHI exposure.
Expect to revise limits after observing real workload patterns. Because each record carries a policy version, changes to a department's allocation are themselves traceable, and a reviewer can see which rules applied before and after the change. That traceability is what turns an allocation from a spreadsheet into something an auditor can examine.
From shared commitment to department-level evidence
Allocation model
Department shares, priorities and burst rules defined from risk tier and workload pattern.
Policy map
Approved models, tools, data handling rules and limits per department.
Runtime enforcement
Allow, redact, block, rate limit or escalate each interaction and tool call.
Evidence
Department-attributed records tied to a policy version for audit review.
Review your department allocation model
Talk through department policy mapping, runtime enforcement and the evidence your HIPAA review needs.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.