AI Use Visibility for Health System Shared Services: What Can Be Enforced and Evidenced at Runtime
AI use visibility in revenue cycle, HR, and supply chain means being able to attribute each AI interaction and tool call to a model, a user or service identity, a function, and the policy decision applied. Records produced where policy is enforced, in the execution path, are more direct evidence for HIPAA 164.312(b) than records reconstructed from vendor, cloud, and business-unit reports. Interaction-level and tool-call-level evidence has limits, which are described below.
Visibility is an enforcement and evidence problem
Shared services teams adopt AI through three routes: features embedded in vendor products, copilots used by staff, and direct calls to model APIs from internal applications. Each route may sit in a different vendor contract, cloud account, or business unit. A usage dashboard assembled from those sources can show that AI is in use. It usually cannot show which data a given request carried or what policy applied when it ran.
For a compliance leader, the useful definition is narrower than a usage report. Visibility means that for each interaction or tool call you can state the model or tool, the caller, the function, the data class involved, and the policy decision taken. Spend attribution follows from the same records, because usage volume tied to a model and a function can be joined to invoices and cloud billing data. Spend is a by-product of attribution here, not the objective.
The attribution fields above are a design framing, not terms defined in HIPAA. They are a practical way to answer the question an auditor is likely to ask: what did your AI do with this data, and who allowed it?
Function-specific use patterns and the data they expose
The data exposure depends on what a workflow sends to a model or tool, not on the function's name. The patterns below are common examples. Confirm which apply in your organization through your own inventory and risk analysis.
- Revenue cycle: Coding assistance, denial analysis, prior authorization support, and patient billing communications. Where these workflows send claim, account, or clinical detail to a model, PHI is exposed to that model and to any tool it can call. CMS describes prior authorization as a large administrative burden for providers, so automation in this area may grow.
- HR: Recruiting, benefits questions, and employee support. Prompts and attachments can contain employee records, compensation, and health-related benefits information. HIPAA sources do not govern this data in general, so its controls need their own policy basis.
- Supply chain: Demand forecasting, contract review, and vendor communications. The exposure is typically vendor terms, pricing, and contract data. Supply chain workflows are less likely to touch ePHI, but that conclusion should come from your risk analysis, not an assumption.
What 164.312(b) requires and what evidence to capture
As of November 2026, 45 CFR 164.312(b) reads: implement hardware, software, and/or procedural mechanisms that record and examine activity in information systems that contain or use electronic protected health information. The standard is technology-neutral and does not prescribe log fields. In its 2003 preamble, HHS described the requirement as the capability to record and examine system activity. Appendix A to Subpart C also lists Information System Activity Review as Required, which pairs recording with periodic examination.
In January 2025, HHS published a proposed rule that would modify the standard to require technology assets or technical controls that record and identify activity in relevant electronic information systems. HHS cited the need for more specificity. It is a proposal, so plan against the current text and track the rulemaking. The proposal also says each entity should rely on its risk analysis to determine reasonable and appropriate audit controls.
Whether an AI tool or model endpoint is an information system that contains or uses ePHI depends on whether PHI enters it. That is an inference to confirm in your risk analysis. Access control under 164.312(a)(1) refers to persons or software programs granted access rights, which supports treating API integrations as identities that need defined rights.
No regulation lists AI-specific evidence fields. The following are recommendations for what an auditor could reasonably expect to examine, not HIPAA requirements.
- Timestamp, and the user or service identity that made the call
- Function or business unit and the application or entry point used
- Model or tool invoked, including tool calls made on the model's behalf
- Data classification detected, such as PHI or employee data
- Policy rule evaluated and the decision taken: allow, redact, block, rate limit, or escalate
- Approver identity and outcome where a human approval was required
Four ways to attribute usage, and where each falls short
No single data source covers every route. The table describes general coverage characteristics. Embedded vendor features are the hardest case, because the model call happens inside the vendor's service and may never cross your own infrastructure.
- Enforcement in the execution path: Sees each request and tool call that passes through it, so it can attribute and apply policy before the action completes. Gap: traffic that does not pass through it, including many embedded vendor features.
- Network-level inspection: Can reveal destinations and volumes across managed networks. Gap: limited insight into encrypted content, user intent, or tool-level actions, and it cannot see traffic off the managed network.
- Vendor-reported usage: Is the main view into embedded features. Gap: report fields and retention are set by the vendor, and records may not carry your policy decision or data class.
- Cloud billing data: Ties consumption to accounts and models for direct API usage. Gap: shows volume and charges, not who sent which data or which policy applied.
Implementation sequence
Start with scope, then place enforcement where it can see the most traffic, and treat what remains as an explicit exception list. Allow, redact, block, rate limit, and escalate for human approval are design options to match to data class and function. Only the access-rights language in 164.312(a)(1) gives a direct regulatory hook for allow and block decisions.
Assign a named owner for periodic review of AI activity records. Recording without examination does not fully serve the audit-controls purpose, and the owner should be documented alongside the mechanisms you chose and your rationale in the risk analysis.
- Inventory AI use cases across the three functions, including embedded features, copilots, and direct API calls.
- Determine through your risk analysis which workflows contain or use ePHI.
- Route direct API and copilot traffic through a point that enforces policy at runtime.
- Map data classes to decisions, for example redact PHI in billing prompts or escalate bulk exports for approval.
- Request usage and retention terms from vendors for embedded features outside your path.
- Set retention, evidence export, and a review cadence with a named owner.
Limits of this approach and how to evaluate it
Interaction-level and tool-call-level evidence records individual events. It does not establish business transaction identity across tasks and agents. It does not provide cross-agent workflow tracing. It does not provide model or dataset lineage. If your audit questions depend on those, they need separate mechanisms. Visibility is also limited to traffic that passes through enforcement, so the exception list for embedded vendor features is part of the evidence you maintain.
When evaluating a runtime enforcement product, ask where enforcement sits relative to the model call and what falls outside it. Ask what is recorded for each interaction and tool call, and whether the policy decision is shown. Ask where evidence is stored, who sets retention, how redacted PHI and prompts are handled in stored records, and whether evidence can be exported for auditors and mapped to HIPAA and NIST AI RMF if you use it. Deployment model and data residency belong on the list, because evidence containing sensitive prompts should stay under your control.
Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites, and coding assistants, scoped to individual interactions and tool calls. It generates audit-ready evidence from that enforcement, mapped to HIPAA, and supports self-managed deployment in the customer cloud with configurable retention and data residency. It runs in production with healthcare organizations and applies one policy plane across 100+ models.
What each AI event should be attributable to
Model or tool
Which model, embedded feature, or tool call handled the request.
User or service identity
The person or software program that initiated it.
Shared services function
Revenue cycle, HR, or supply chain.
Policy decision
Allow, redact, block, rate limit, or escalate, with the rule applied.
Review your shared services AI evidence
Walk through how policy enforced at runtime produces attributable, audit-ready evidence for revenue cycle, HR, and supply chain.
Request a DemoHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.