Controlling outsourced RCM vendor AI agents working inside your EHR
Give each vendor AI agent its own inventoried identity, scope its permitted EHR actions and data to the contracted function, and enforce policy on every model call and tool call in the execution path: allow, redact, block, rate limit, or escalate for human approval. Keep per-action evidence that you can retrieve independently of the vendor. Contracts set expectations. Only runtime enforcement limits what an agent actually does.
Implementation sequence
- 1
1. Inventory agents and identities
List every vendor agent, the credential it uses, the EHR functions it can reach and the vendor contact who owns it. Remove shared accounts.
- 2
2. Scope actions and data
Define permitted actions and data per agent from the contracted function, applying minimum necessary access. Document the access control method chosen and why.
- 3
3. Place enforcement in the execution path
Put policy decisions on model calls and tool calls. For each integration pattern in use, such as FHIR APIs, EHR app frameworks, UI automation or service accounts, confirm where traffic can actually be intercepted.
- 4
4. Apply redaction and escalation
Redact PHI that the task does not need. Route high-risk actions to a human approver before execution.
- 5
5. Fix the evidence fields
Decide what each enforcement decision records, and check the fields against the audit-control requirement to record and examine activity.
- 6
6. Set retention and residency, then test
Configure how long evidence is kept and where. Run out-of-scope requests against each agent and confirm they are blocked and recorded.
Evidence to capture at enforcement time
- Agent identity, vendor and credential used
- Action or tool call requested and the target record or resource
- Data scope requested, data returned, and fields redacted
- Policy decision (allow, redact, block, rate limit, escalate) and the policy version that produced it
- For escalations: approver identity, decision and timestamp
- Timestamp, retention setting and storage region
Which controls belong in the contract and which in the execution path
Satisfactory assurances and BAA terms
Contract. Documents the vendor's obligations and any stricter terms both parties agree to.
Subcontractors such as model providers
Contract. Addressed through the vendor's obligations, since assurances are not required from a subcontractor.
Per-agent identity and credentials
Contract requires it. Verified in the inventory and at the enforcement point.
Limiting access to granted rights
Execution path. Policy decision on each read, write and tool call.
Human approval of high-risk actions
Execution path. The action is held until a named approver acts.
Recording activity
Execution path. Evidence is produced when the decision is made, not reconstructed later.
Define the control surface before choosing controls
An outsourced RCM vendor may run agents for coding, claim preparation, prior authorization, denial management and patient billing. Each agent reads PHI, calls tools and sometimes writes back to the record, usually through vendor-managed credentials. The control surface is every point where that happens: the identity presenting the credential, the data it can reach, and the actions it can take.
The technical safeguards in 45 CFR 164.312 require access control for systems holding ePHI so that access is allowed only to persons or software programs that have been granted access rights. Because the standard names software programs, a vendor agent or its service identity can reasonably be treated as within scope. That is an interpretation, not guidance written for AI agents. The Security Rule preamble published in January 2025 also describes access authorization as covering computer systems as well as users, and says minimum necessary access should follow the function being performed.
The practical consequence is that rights must attach to something specific. A single shared vendor service account used by several agents makes per-agent rights impossible to define and per-agent activity impossible to attribute. Start by requiring a distinct identity for each agent, then record it in an asset inventory. The same preamble says that when more than one party controls a technology asset, each should include it in its inventory and network maps, using a covered entity and its cloud EHR vendor as the example. Applying that logic to vendor agents is a design choice, and a sound one.
Mapping RCM agent actions to enforcement decisions
Risk falls into two kinds: PHI exposure from reads and outbound prompts, and integrity risk from writes and submissions. The mapping below is a starting design for the categories most RCM agents perform, to be tuned against each vendor's contracted function and your risk assessment.
Default-deny for writes is the most useful single rule. An agent that suggests a code and an agent that changes a claim or chart should not hold the same permissions, and policy should distinguish them even when both use the same credential type. Rate limits are the main defense against bulk reads that look legitimate individually, such as an agent walking through many patient records when its queue only calls for a few.
- Chart reads: Allow within the assigned patient or encounter. Redact fields the task does not need. Rate limit bulk access.
- Code suggestion: Allow. Redact direct identifiers from content sent to a model when the task does not require them.
- Claim edits: Allow within approved fields. Block edits outside them. Escalate changes above a threshold you define.
- Record writes: Block by default. Escalate for human approval where a write is part of the contracted function.
- Payer submissions: Escalate for approval where policy requires it. Rate limit to catch runaway loops.
- Patient billing communications: Redact PHI beyond what the message needs. Block content that exceeds the agent's data scope.
Where Trussed fits, and where it stops
Trussed enforces policy in the execution path across models, agents, tools and MCP servers, with the decisions described above: allow, redact, block, rate limit, and escalate for human approval. Enforcement is at runtime, not reviewed after the fact. Each decision produces audit-ready evidence mapped to HIPAA and NIST. Trussed can be deployed self-managed in the customer's cloud, with configurable retention and data residency, which supports holding evidence independently of the vendor. It has run 18+ months in continuous production in a regulated enterprise at 99.95% uptime on live traffic.
The boundary is specific. Policy applies to individual interactions and tool calls. It does not govern a business transaction, such as a single claim as it moves across several agents and steps. If you need that view, treat it as a separate requirement and do not assume per-call enforcement provides it.
Whether any enforcement point can see a vendor agent's actions depends on the integration pattern. Confirm interception for each one before relying on it, and use that finding to decide what the vendor must change in how its agents reach your EHR.
Control surface for a vendor agent in the EHR
Label: Identity
Description: One inventoried identity and credential per agent, not a shared vendor service account.
Label: Scope
Description: Permitted actions and data tied to the contracted RCM function.
Label: Enforcement
Description: Policy decision on each model call and tool call before it executes.
Label: Evidence
Description: A per-action record of the request, the decision and the outcome.
Review enforcement and evidence for your vendor agents
Walk through how policy enforcement in the execution path and audit-ready evidence apply to the RCM agents working in your EHR.
See How Evidence Is Produced