AI Spend Visibility for Academic Medical Centers: Separating Clinical, Research, and Education Use
AI spend visibility for academic medical centers depends on attributing each AI interaction to a clinical, research, or education context when it happens. Invoices and vendor accounts cannot do this because the three mission areas share models, tools, and accounts. The reliable approach is to capture identity, context, and data-classification signals in the execution path, apply policy per use category, and record the decision as audit-ready evidence.
Why billing data cannot separate mission areas
An academic medical center runs AI for three purposes at once: clinical care involving PHI, research governed by IRB approvals and data use agreements, and education involving students and trainees. These workloads frequently call the same models, use the same tools, and sit under the same vendor accounts. An invoice shows volume per account or per key. It does not show whether a given interaction was a clinician drafting a note, a researcher analyzing a dataset, or a trainee studying for an exam.
When attribution relies on manual tagging or after-the-fact reconstruction, finance, compliance, and architecture teams each build a different answer, and auditors cannot readily verify any of them. Treat visibility as an evidentiary problem. The question is not only how much AI was used, but which policy applied to each interaction and whether PHI-bearing use stayed within its permitted boundary. Spend by mission area then becomes an aggregation of verified interaction records rather than an estimate.
Signals that can be captured at the interaction and tool-call level
Use category has to be derived from signals present when the interaction occurs. Several are usually available at the point where an application or agent calls a model, tool, or MCP server: the authenticated user or service identity and its role, the application or workspace making the call, a declared purpose or protocol identifier, the data classification of the content (for example, PHI present or absent), the model requested, and the tool or MCP server invoked.
No single signal is sufficient. A physician who also teaches and conducts research holds one identity across all three contexts, so role alone is ambiguous. Application or workspace often disambiguates, and a protocol identifier or course context can confirm it. The practical design is to decide which system is authoritative for each signal, such as identity provider groups, a protocol registry, or course enrollment, and to deliver that context to the enforcement point automatically rather than asking users to tag requests.
- Authenticated identity and role
- Application or workspace
- Declared purpose or protocol identifier
- Data classification, such as PHI present or absent
- Model requested
- Tool or MCP server invoked
Enforcing policy per use category in the execution path
Attribution and enforcement work best when one control point produces both. If policy is enforced in the execution path, between the calling application or agent and the model, tool, or MCP server, the same component that decides also records what it decided and why. A separate observability or reporting layer has to infer after the fact what the policy would have been.
Policy should differ by use category. A clinical context might allow PHI to reach an approved model, redact it when a request goes to a tool outside that boundary, and block it for an unapproved model. A research context might permit only specific models and tools tied to a protocol, and block PHI that falls outside its approved data use. An education context might block PHI entirely. The available outcomes are allow, redact, block, and escalation for human approval where a decision needs a reviewer. Each outcome should be stored with the policy version that produced it.
Scope matters here. This pattern governs individual interactions and tool calls. It does not by itself establish transaction-level governance, model or dataset provenance, or end-to-end tracing of an agentic workflow, and requirements should not assume it does.
Evidence that supports HIPAA audit-control expectations
The HIPAA Security Rule technical safeguards at 45 CFR 164.312 include audit controls at 164.312(b), and the section directs covered entities and business associates to comply in accordance with 164.306. HHS also published a proposed rule in the Federal Register on January 6, 2025, titled HIPAA Security Rule To Strengthen the Cybersecurity of Electronic Protected Health Information. As of November 2026, architects should confirm the current text of 164.312(b) with compliance counsel before fixing requirements, rather than designing to the proposal.
Whatever the final wording, an auditor evaluating AI usage will need records that are specific and tamper-resistant enough to answer the same few questions. Who or what made the request, in which use category, under which policy, with what decision, and which tool or server was invoked? A useful evidence record carries a timestamp, identity, use category, policy applied, decision, and the model, tool, or MCP server involved. It should be exportable in a form compliance teams, IRB administrators, and internal audit can read without engineering help.
Because the record is produced by enforcement rather than assembled later, it reflects what the system actually did, not what a team recalls or a spreadsheet claims.
Implementation decisions to settle first
Three decisions shape the design more than product selection does. The first is the authoritative source of use-category context and how it reaches the enforcement point. The second is what happens when no category can be determined. Reasonable options are default deny, default to the most restrictive policy, or quarantine for review. Whichever is chosen, the handling itself should appear in the evidence so it can be shown to auditors. The third is the evidence record schema and its export path.
There are tradeoffs. Default deny protects PHI but interrupts legitimate work while identity and protocol mappings are incomplete. Applying the most restrictive policy keeps work moving but may under-serve research use. Introducing human approval for ambiguous cases adds reviewer load and latency. Start with the highest-risk paths, those where PHI can reach external models or tools, and widen coverage as context signals mature.
Evaluating a runtime approach
A runtime approach fits when attribution must hold up to audit and when the same models serve several mission areas. Evaluate it on five points: which signals it can read at the enforcement point, whether policy can differ by use category, whether decisions are recorded with the policy applied, whether coverage extends to tool calls and MCP servers as well as model calls, and what it explicitly leaves out of scope. Deployment fit matters equally. Confirm that it can run self-managed in your own cloud, with retention and data residency settings that meet institutional requirements, and establish who can access the stored records.
Trussed enforces policy in the execution path across models, agents, tools, and MCP servers, with allow, redact, block, rate limit, and escalate for human approval decisions at the level of individual interactions and tool calls. Enforcement generates audit-ready evidence mapped to HIPAA, among other frameworks. It deploys self-managed in the customer cloud with configurable retention and data residency, and it runs in production with healthcare organizations. It does not provide transaction-level governance, model or dataset provenance, or end-to-end agentic workflow tracing.
Attribution at the enforcement point
Clinical
PHI-bearing interactions evaluated under clinical policy.
Research
Interactions tied to a protocol or approved data use context.
Education
Student and trainee interactions evaluated under education policy.
Evidence record
Identity, use category, policy applied, decision, and tool invoked.
Test your attribution requirements against a runtime design
Walk through your clinical, research, and education workloads with an expert and see how interaction-level evidence is produced.
Talk to an Expert