How to Track AI Use by Affiliated Physician Groups on a Health System's AI Contracts
To track AI use by affiliated physician groups, record each interaction with a unique user ID, an affiliation resolved from an authoritative record, the AI tool, the interaction type, the PHI handling action, and the policy decision. Map each contract and BAA term to a rule that can be enforced, and generate the record at the point of enforcement rather than reconstructing it from vendor reports later.
A repeatable process for audit-ready evidence
- 1
1. Establish identity
Provision a unique ID for each affiliated-group user and link it to the affiliation record.
- 2
2. Enforce policy
Apply the mapped rules at runtime so each interaction yields a decision record.
- 3
3. Capture evidence
Store the six data elements per interaction, with retention set by policy and counsel.
- 4
4. Review regularly
Examine audit logs and access reports on a set cadence for unusual access patterns.
- 5
5. Report across groups
Produce a report by group, tool, policy, and PHI handling action for a given period.
- 6
6. Resolve and update
Route exceptions to group compliance contacts and revise the mapping register.
The usage data needed to attribute AI activity to a group
Why attribution breaks down across affiliated groups
A health system typically signs enterprise AI contracts for ambient documentation, clinical decision support, and generative assistants. Affiliated physician groups are separately organized, yet they often reach those tools through the system's tenant, licenses, or identity services. The contract sits with one legal entity while the usage comes from several.
That gap creates three problems. Usage is hard to attribute to a specific group, user, and purpose. Permitted-use limits and PHI handling rules in the agreements may be applied differently from group to group. And when an auditor asks which groups used which tools, under what policy, and with what handling of PHI, the answer is assembled by hand from fragments.
Scale makes this more pressing. ASTP data show 71% of hospitals reported using predictive AI integrated with the EHR in 2024, up from 66% in 2023, and system-affiliated hospitals reported higher use than independent hospitals (86% vs. 37% in 2024). Those figures cover hospitals, not physician groups, but they show how much AI use sits inside system structures.
Begin with the structures, not the tooling. List each enterprise AI contract, the BAA that covers it, and the participation or affiliation terms that extend access to each group. For each group, decide whether it is a covered entity, a business associate, or neither in relation to the health system, because that answer shapes who is responsible for the records. Counsel should confirm these determinations.
Mapping contract terms to enforceable policy
Take each term from the enterprise agreement, the BAA, and the participation terms and write it as a rule with a named owner. A permitted-use limit such as restricting a tool to clinical documentation becomes a rule on which interaction types are allowed for which groups. A PHI handling restriction becomes a rule on what is redacted or blocked before data reaches a tool. A term that requires human sign-off becomes an escalation path.
Some terms will not translate. Where a term cannot be enforced or evidenced technically, record it as a gap with an owner and a compensating control, such as a periodic attestation from the group's compliance contact. Gaps documented this way are easier to defend than terms assumed to be followed.
Keep a single mapping register that links each clause to its rule, the groups it covers, and the evidence it produces. When a contract is renegotiated or a group joins or leaves, the register shows which rules change.
- Inventory: Contracts, BAAs, and participation terms, with the groups that can reach each tool.
- Translate: Convert each clause into a rule: allow, redact, block, or escalate.
- Flag gaps: Record terms that cannot be enforced technically, with compensating controls.
- Assign owners: Name an owner at the system and a compliance contact at each group.
Vendor reports versus evidence generated at enforcement
Vendor usage reports and billing data are useful for capacity and invoicing, but they are a weak basis for compliance evidence. They reflect what the vendor chose to record, typically at the account or tenant level. They may not carry your affiliation attribute, may not show the PHI handling action, and cannot show which policy applied, because the vendor does not know your policy. Reconstructing group-level use from them after the fact means joining data from systems that were not designed to agree.
Evidence produced in the execution path differs in kind. When policy is enforced at runtime, each decision to allow, redact, block, or escalate is recorded as it happens, with the identity and context available at that moment. The record shows what the control did, not an inference about what occurred. It is also durable, because it does not depend on a vendor's retention choices or on a later review being completed.
Trussed enforces policy in the execution path across models, agents, tools, and MCP servers, and generates continuous audit-ready evidence from that enforcement, mapped to HIPAA among other frameworks. Its scope is individual interactions and tool calls. It can be deployed self-managed in the customer's cloud, with configurable retention and data residency. Whatever tool you evaluate, test whether group attribution comes from your affiliation record and whether the evidence format matches what your auditors request.
Governance decisions and evaluation criteria
Ownership matters as much as tooling. The health system's compliance team should own the cross-group report and the mapping register, while each group's compliance contact owns user provisioning, attestations, and exception follow-up. Write this split into the participation terms or an operating procedure so it survives staff changes.
Retention and export need explicit decisions. Set retention periods with counsel against current regulatory text, and confirm that evidence can be exported in a form an auditor can read without access to the live system.
Finally, keep the regulatory picture current. HHS published a request for information on accelerating AI adoption in December 2025, and a January 2025 Federal Register document addresses Security Rule cybersecurity. As of November 2026, check the status of both before citing new obligations.
When evaluating any approach, ask five questions. Can each interaction be attributed to a unique user and a group? Which contract terms are enforced in the execution path, and which are only documented? What evidence does each allow, redact, block, or escalate decision produce? How long is it retained? Can you produce a cross-group report of tools, policy, and PHI handling for a period?
From contract to evidence
Contract and affiliation
Enterprise agreements, BAAs, and participation terms define who may use which tool.
Identity and group
A unique user ID linked to an authoritative affiliation record.
Policy at runtime
Each term becomes a rule: allow, redact, block, or escalate.
Cross-group evidence
Attributable records support review and reporting by group, tool, and period.
See what attributable AI evidence looks like
Walk through how policy enforced in the execution path produces audit-ready evidence for affiliated groups on shared AI contracts.
Request a Demo