Attributing AI Spend to Each Facility in Multi-Hospital Health Systems
Attribute AI spend by capturing facility context at the point of each AI interaction, taken from the authenticated user or service rather than inferred later from a billing account. Use direct metering where platforms allow it, tagging or identity-based allocation on shared platforms, and proportional allocation only for costs that cannot be traced. Keep the record to metadata that holds no PHI, so finance and internal audit can defend it.
The facility-level data to capture at each interaction
- Facility identifier: a code from the facility master, with a crosswalk to CCN or NPI if needed
- Department or cost center: the unit that will carry the allocation
- Authenticated actor: a pseudonymous user identifier, or a named service or agent identity
- Role category: a coarse role class, not detailed job or patient-facing context
- Application, model, and tool: which assistant, model endpoint, or tool call was used
- Timestamp, usage units, and policy outcome: when it happened, how much was consumed, and whether it was allowed, redacted, or blocked
Why account-level billing cannot answer the facility question
Multi-hospital health systems usually run AI models, assistants, and agents on shared platforms, under shared vendor contracts, operated by centralized teams. Usage and billing data tends to arrive at the account or tenant level. When identity, tenancy, and model access are also shared across facilities, an invoice shows what was consumed but not which hospital consumed it.
The vendor mix makes this harder. In 2024, 71% of hospitals reported using predictive AI integrated with the EHR. Of hospitals using predictive AI, 80% used AI from their EHR developer, 52% used third-party AI, and 50% used self-developed AI. Those figures describe predictive AI, but they show that usage commonly spans several vendors and platforms. Any assistants and agents added on top bring further platforms to reconcile into one record set.
It helps to treat this as a data-capture and identity problem at the interaction layer, not a reporting problem. Finance can only allocate what the interaction layer recorded. The goal is accountability: a record that shows which facility's user or service made which AI call, and that an auditor can examine. Note that the HIPAA audit controls standard at 45 CFR 164.312(b) requires mechanisms that record and examine activity in systems that contain or use ePHI. It does not require cost attribution, so attribution records are best framed as evidence under the organization's own policy.
Four attribution methods and their tradeoffs
Most health systems end up combining methods. The useful question for each AI platform is how much facility context it can supply at the time of the call.
- Direct metering: Usage is measured per facility at the source, for example through dedicated deployments or credentials per hospital. Accuracy is highest, but endpoints, contracts, and operations multiply, and it cannot be applied to a model endpoint that is shared by design.
- Tagging: A facility tag is attached to requests or resources. It is simple to start, but coverage depends on consistent application, untagged usage accumulates, and a tag set by the caller is an assertion rather than evidence.
- Identity-based allocation: Facility is derived from the authenticated user or service identity and directory attributes. It ties usage to an accountable actor, but it breaks down for clinicians who work across facilities and for shared groups or service accounts.
- Proportional allocation: Shared costs are split by a driver such as headcount, encounter volume, or bed count. It is easy to run and suits genuinely shared platform costs, but it does not show actual use and is the hardest method to defend in audit.
Resolving shared identity while minimizing PHI
Shared identity providers and central platforms hide facility in predictable ways. A user's home facility may differ from where they are working. A central agent may act for several hospitals. A service account may serve everyone. The pattern that works is to resolve facility at authentication or at the point of the call, using an explicit active-facility context for the session, and to carry that context with the request. For users who work across facilities, the active context, not the directory home, should drive attribution. For central agents and service accounts, the facility of the requesting context should travel with each call. A single owned mapping from identity groups and endpoints to facilities keeps the method consistent.
The tension with HIPAA is that granular attribution wants more context while PHI minimization wants less. Facility codes, role categories, and application names are metadata, but combinations of fields, or fields joined to other data, can become identifying. Run and document a minimization review of each field. Use pseudonymous actor identifiers, keep content out of the attribution record, and store the record apart from clinical data with its own access rules.
An attribution log that records interaction activity may fall within the scope of 164.312(b), depending on the system. Protect it with access control and integrity measures. Encryption is an addressable specification under 164.312, so record the decision either way. The Security Rule amendments proposed in January 2025 should be treated as a proposal, not current law, as of November 2026.
What makes an attribution record defensible, and where runtime enforcement fits
Internal audit, finance, and compliance generally need records that are complete, consistent across platforms, traceable to an authenticated actor, and protected against alteration. They also need to be able to reproduce how an allocation was derived from the records. Records assembled afterward from billing exports and directory lookups struggle on all of these, because facility context has to be inferred.
Runtime policy enforcement addresses this at the source. Trussed enforces policy in the execution path across models, agents, tools, and MCP servers, at the level of individual interactions and tool calls. Because each interaction passes through that path, facility context can be attached from the authenticated identity at the moment of the call, and policy is enforced at runtime, not reviewed after the fact: allow, redact, or block. Redaction keeps PHI out of what is recorded. The same enforcement produces audit-ready evidence mapped to HIPAA, including the 164.312(b) audit controls expectation.
Trussed can be self-managed in the customer's cloud, with configurable retention and data residency, so attribution evidence stays within the health system's own environment. It runs in production with healthcare and pharma organizations. It does not replace the finance tools that perform allocation. It supplies consistent facility-level metadata and evidence for those tools to use.
A practical sequence for designing the approach
Start with the platforms that carry the most usage and work outward. Each step should produce something audit can review.
- Inventory AI platforms and contracts: List EHR-vendor, third-party, and self-developed AI, and note what facility context each can supply at the time of a call.
- Fix the facility master: Establish one facility and cost center reference, with crosswalks to CCN or NPI where useful.
- Assign a method per platform: Use direct metering, tagging, or identity-based allocation where context is available, and proportional allocation only for the residual.
- Define and review the field set: Document each attribution field, its purpose, and the outcome of the PHI minimization review.
- Capture context in the path: Resolve facility from authenticated identity at or before the call, and handle cross-facility users through explicit session context.
- Assign ownership and test with audit: Name owners across IT, finance, compliance, and clinical governance, then have internal audit sample the records against the method.
From AI interaction to defensible facility record
Interaction
A user, agent, or service calls a model or tool through a shared platform.
Facility context
Facility, cost center, role, application, and model are resolved from authenticated identity.
Attribution record
A metadata-only record is written with no prompt or response content.
Allocation and audit
Finance allocates from the records, and internal audit can examine them.
Review how facility-level evidence is produced in the execution path
Walk through how policy enforcement attaches facility context to each interaction and generates audit-ready evidence mapped to HIPAA.
See How Evidence Is ProducedHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.