A 90-Day Plan for AI Usage and Spend Visibility in Health Systems
A health system CIO can establish AI usage and spend visibility in 90 days by inventorying and baselining AI use in days 1-30, enforcing policy in the execution path and attributing every interaction to a team and model in days 31-60, and moving to reporting, review cadence and evidence handoff to compliance and finance in days 61-90. Attribution captured at the point of enforcement is consistent and tied to the policy decision. Attribution rebuilt later from billing exports and log review is not.
What visibility requires and the attribution schema to define first
Health systems now use AI in clinical, revenue cycle, operations, research and IT teams, through direct model access, vendor-embedded features, coding assistants and browser-based AI sites. The CIO question is specific: which teams use which models, what data reaches them, and how consumption is distributed. Vendor and cloud billing data is generally organized around accounts and subscriptions, so it rarely answers that question at the level of team, model or use.
The compliance question is different but depends on the same data. Under 45 CFR 164.312(b), a regulated entity must implement hardware, software, and/or procedural mechanisms that record and examine activity in information systems that contain or use electronic protected health information. The text is technology-neutral. HHS describes the choice of audit controls as driven by the entity's risk analysis and its technical infrastructure. A runtime enforcement layer is therefore one design option, not a stated requirement.
Define the attribution schema before collecting anything, so that every phase produces comparable records. A workable starting set is team, model, application, data classification (PHI or not), tool or MCP server, user role, policy action (allow, redact, block, rate limit, escalate) and timestamp. Treat service accounts, agents and API keys as actors alongside people. The access control standard in 164.312 refers to software programs as well as persons, so each of these identities needs an owner and an authorized purpose.
Assign ownership at the start. The CIO office owns the plan. Security owns discovery and enforcement. Compliance owns the mapping to audit controls. Finance owns the consumption baseline. A multidisciplinary governance group owns the inventory and review dates.
Days 1-30: inventory and baseline
The goal of the first phase is an honest inventory, not a clean one. Start with the inventory fields in ASTP/ONC's SAFER guidance for clinically deployed AI-enabled systems: deployment date, current version, responsible personnel, last reviewed date, authorized users, authorized purpose, source of data, and external validation. The guide scopes these fields to clinical systems, so extend them to revenue cycle, research and IT use as a local decision. Add a model-source tag (EHR-developed, third-party or self-developed), since hospital evaluation and monitoring practices vary by model source.
Discovery should combine several views because no single one is complete. Useful inputs include egress and proxy data for known AI destinations, single sign-on and OAuth application grants, endpoint and browser extension inventories, procurement and contract review for AI features embedded in existing vendor products, cloud accounts and secret stores for departmental API keys, and structured interviews with department leaders. Coding assistants and browser-based AI sites are the usual blind spots, so check for them explicitly.
In parallel, document the risk analysis that sets the scope of audit controls, and record which entries touch PHI or could.
Exit criteria, set by the CIO and not by any external threshold: every discovered AI use has a named responsible owner and a model-source tag, PHI exposure is classified for each entry, the attribution schema is approved by security, compliance and finance, and a consumption baseline exists, even where it can only be attributed to the account level for now.
Days 31-60: enforce policy and attribute every interaction
The second phase moves from observing to enforcing in the execution path, so that the record of each interaction is produced at the moment the policy decision is made. Capturing team, model, application, data classification, tool and user role at that point avoids the reconciliation step that makes after-the-fact attribution incomplete.
Roll out in stages to avoid disrupting clinical and operational workflows. Begin in monitor-only mode, where decisions are recorded but not applied, and review what the policy would have done with each owning team. Then enable actions in order of lowest workflow impact: redact PHI in prompts bound for approved models, block PHI to destinations that are not approved, rate limit high-volume callers, and escalate for human approval where a use needs a named reviewer. Each step should have a rollback path and a named approver in the affected department.
Cover the full surface found in phase one: models, agents, tools, MCP servers, AI sites and coding assistants. Agents and service credentials need the same owner and authorized-purpose fields as people.
Exit criteria, again CIO-defined: the share of interactions attributed to a team and model meets the target set at the start of the phase, every enforced policy has an owner and a documented exception path, no unattributed traffic to unapproved destinations remains in monitor-only review, and compliance has confirmed the recorded fields match the audit-control scope from the risk analysis.
Days 61-90: reporting, review cadence and evidence handoff
The third phase turns attributed records into recurring outputs for two audiences. Finance needs consumption by team and model against the baseline from phase one, so budget owners have a defensible starting position. This is a visibility objective. It establishes who uses what, and it does not by itself change spend. Compliance needs evidence that usage stayed within policy: policy decisions by action, PHI-related interactions and their outcomes, and the record of exceptions and approvals, mapped to 164.312(b).
Set a standing review cadence with the multidisciplinary governance group. Monthly is a reasonable default. Each review should cover inventory completeness, new or changed AI uses, policy exceptions, escalations and their resolution, and monitoring coverage across all models, not only the ones that are easiest to measure.
Keep evidence design adaptable. HHS issued a proposed rule in December 2024 to strengthen Security Rule protections for ePHI, including more specificity on audit controls. As of November 2026 it is a proposal and not a final rule, so retention periods and evidence formats should be configurable and not hard-coded.
Exit criteria: a first finance report and a first compliance evidence package have been delivered and accepted by their owners, the review calendar and attendees are fixed, and the inventory has a named steward who keeps it current after day 90.
Billing exports, log review and runtime enforcement compared
The three approaches answer different questions. The comparison below is based on how each method works, and most health systems will use more than one.
- Billing exports: Reflect charges by account, subscription or project. Useful for total consumption. Attribution to team, model and use depends on tagging discipline, and there is no record of data classification or policy decisions.
- After-the-fact log review: Can show what happened once logs are collected and joined. Completeness depends on every system emitting usable logs. Linking activity to a policy decision is a manual step.
- Runtime enforcement: Records team, model, application, data classification and policy action at the moment of each interaction. Covers only traffic that passes through the enforcement path, so discovery in phase one still matters.
- What to evaluate: Attribution completeness, linkage to allow, redact, block, rate limit and escalate decisions, treatment of agents and service accounts as actors, export formats for compliance, and retention and residency options.
Where a runtime enforcement layer fits
Trussed enforces policy in the execution path across models, agents, tools, MCP servers, AI sites and coding assistants under one policy plane. Actions are allow, redact, block, rate limit and escalate for human approval, applied to individual interactions and tool calls. The enforcement itself generates continuous, audit-ready evidence mapped to HIPAA, including 164.312(b) audit controls, so attribution and evidence come from the same event.
Trussed deploys self-managed in the customer cloud, with configurable retention and data residency, which matters when evidence about PHI-related interactions must stay inside the health system's own environment. It runs in production with healthcare and pharma organizations.
Trussed does not replace the phase one inventory, and it does not provide model lineage, transaction-level governance or end-to-end tracing of agentic workflows. Its contribution to this plan is narrower: consistent attribution and policy-linked evidence for the interactions it enforces.
90-day visibility plan
Days 1-30: Inventory and baseline
Discover AI use, models, teams and data exposure. Name an owner for every system.
Days 31-60: Enforce and attribute
Apply policy in the execution path. Attribute each interaction to a team and model.
Days 61-90: Report and hand off
Set the review cadence and deliver evidence to compliance and finance.
Plan the first 30 days with your security and compliance leads
Review your attribution schema, discovery scope and evidence requirements with a Trussed expert.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.