How to Report AI Use and Spend by Model to a Health System Board
Build the board report on a single model inventory. Give every model a stable identifier, then map vendor invoices, cloud billing exports and application usage records to that identifier. For each model, present usage, attributed spend, PHI access and policy outcomes (allow, redact, block, escalate), and keep each figure traceable to source records that support audit controls under 45 CFR 164.312(b).
Structuring the recurring board report
- 1
1. Inventory changes
New, retired and changed models since the last period, with model source and owner.
- 2
2. Usage and attributed spend
Per-model figures, the unattributed remainder and the attribution method used.
- 3
3. PHI access
Models and software programs with access to ePHI-bearing systems, and any changes to approvals.
- 4
4. Policy outcomes
Allow, redact, block and escalate rates per model, with notes on material movements.
- 5
5. Oversight coverage
Share of models evaluated and monitored, reported as coverage, not as a yes or no.
- 6
6. Evidence and exceptions
Where each figure can be traced, retention periods, and open exceptions with owners and dates.
A board-level reporting schema by model
- Model identity: stable identifier, vendor, model source (EHR-developed, third-party or self-developed), business owner and intended use.
- Usage: request volume and the applications or user groups calling the model in the period.
- Attributed spend: cost mapped to the model for the period, with any unattributed remainder shown as its own line.
- PHI access: whether the model or the software program calling it has been granted access to systems containing ePHI, and who approved it.
- Policy outcomes: counts of allow, redact, block and escalate decisions for the period.
- Oversight coverage: share of models with completed evaluation and ongoing monitoring.
What a board needs from AI reporting
Boards and risk or audit committees are asking a plain set of questions: what AI is in use, which models are being called, by whom, and at what cost. The difficulty is rarely the question. It is that the answers sit in separate places that do not share a model identifier, so no one can produce a single figure that holds up when challenged.
The scale of the inventory problem is visible in federal hospital data. In 2024, 71% of hospitals reported using predictive AI integrated with the EHR, up from 66% in 2023. Use varied by whether the model was EHR-developed, third-party developed or self-developed, and most hospitals evaluated models for accuracy and bias, but fewer did so for all or most of their models. Those figures cover predictive models in the EHR, not generative AI or LLM API usage, but they show two useful reporting principles: segment by model source, and report coverage as a share of models, not as a yes or no.
A board report should therefore be short, stable from quarter to quarter, and built so that every number has a defined origin. The schema below is a starting point, not a prescribed template.
Attributing usage and spend to individual models
The model inventory is the join key. Without it, each data source describes AI in its own vocabulary: a vendor account, a cloud service line, an application name. The practical task is to map each source to inventory identifiers and to state openly where the mapping is incomplete.
Each source answers part of the question and has limits in a multi-model environment. The descriptions below are general characteristics and will vary by vendor and cloud provider.
- Vendor invoices: Authoritative for what was billed. Typically aggregated at account or plan level, so one invoice may cover several models or several departments.
- Cloud billing exports: Detailed by service and resource. Resources must be tagged or otherwise mapped to a model identifier, and untagged usage lands in an unattributed bucket.
- Application usage records: Show which application called which model and how often. Coverage depends on every application emitting consistent records, and they do not reconcile to cost on their own.
- Records from a common enforcement point: When requests pass through one point, usage is recorded in one format per interaction, which makes model-level counts comparable across applications.
Reconciling figures and showing the remainder
Reconcile in both directions. Total attributed spend plus the unattributed remainder should equal the invoiced and billed total for the period. Request counts from application records should be compared with counts from any shared recording point, and material gaps should be explained in the report, not smoothed over.
A visible unattributed line is more credible to a board than a figure that appears complete. Track it as a metric in its own right and expect it to shrink as tagging, inventory discipline and recording coverage improve. Record the attribution method used for each model, such as direct tag, usage-proportional allocation or manual assignment, so a reviewer can tell an observed figure from an estimated one.
Pairing usage with policy outcomes and HIPAA evidence
The audit controls standard at 45 CFR 164.312(b) requires hardware, software and procedural mechanisms that record and examine activity in information systems that contain or use ePHI. It does not specify fields or per-model formats, so the practical design question is what activity record would let an auditor trace a board figure back to source. For AI, that means a record per interaction that identifies the model, the calling application or software program, the policy decision applied and the time. The access control standard similarly refers to access being allowed only to persons or software programs that have been granted access rights, which is consistent with an inventory that lists which models and programs may touch ePHI-bearing systems.
Information System Activity Review is a Required implementation specification in the Security Rule appendix, as is Response and Reporting under Security Incident Procedures. Board reporting can sit on top of those existing review and incident processes instead of creating a parallel one.
As of November 2026, the January 2025 proposal to amend the Security Rule remains a proposal in the material reviewed, so it should not be presented as a binding obligation. It is still a useful signal: it states that ePHI in AI training data, prediction models and algorithm data is protected, and its proposed text would require retaining activity records for a period that is reasonable and appropriate for the type of log. Setting a retention period for model-level activity records now is sensible either way.
Per-model policy outcomes work as risk indicators next to usage. A high redact count on one model may show PHI regularly reaching prompts. A rising block count may show applications attempting disallowed use. Escalations show where human approval is being exercised. Present counts as rates against request volume and annotate unusual movements, because the number alone does not say whether controls worked or an application changed.
Where runtime enforcement fits
Reporting is only as reliable as the records beneath it. Where policy is enforced in the execution path, each decision to allow, redact or block, or to escalate for human approval, is made at runtime, not reviewed after the fact, and it produces a record at the moment it happens. Trussed enforces policy in the execution path across 100+ models under one policy plane and generates audit-ready evidence mapped to HIPAA, including PHI handling and 164.312(b) audit controls. It runs in production with healthcare organizations. Its scope is individual interactions and tool calls, which is the level at which per-model policy outcome counts are drawn.
This guide stands without it. Any recording point that produces consistent, retained, model-identified records can support the same board schema.
Board report data chain, by model
Model inventory
One identifier per model, with owner, source and intended use.
Usage
Request volume and active applications attributed to each model.
Spend
Invoice and cloud billing figures mapped to the same identifier.
PHI access
Which models and software programs touch systems holding ePHI.
Policy outcomes
Allow, redact, block and escalate counts per model.
Evidence
Source records an auditor can trace each figure back to.
Build a board report you can trace to source
Talk through your model inventory, data sources and evidence requirements with a Trussed AI expert.
Talk to an ExpertHow does your AI governance program compare?
See where your program has gaps in less than 2 minutes.