---
title: What evidence do AI agents handling PHI need under HIPAA?
url: https://opencontroller.lyzr.com/blog/hipaa-evidence-for-ai-agents
published: 2026-11-13
updated: 2026-11-13
author: Lyzr
---

# What evidence do AI agents handling PHI need under HIPAA?

**An AI agent that reads or sends protected health information is an information system under the HIPAA Security Rule. You need the same evidence as for any system, plus proof of two agent-specific things: which person each agent action was for, and that every model or platform provider that sees PHI is covered by a business associate agreement.**

HIPAA does not mention AI agents. It doesn't need to. The Security Rule applies to electronic PHI wherever it goes, including into prompts, retrieval indexes, agent memory and tool calls. This is a checklist of evidence for HIPAA controls, not legal advice or a compliance determination.

## Which HIPAA requirements apply to AI agents?

These are the Security and Privacy Rule provisions that matter most for agents.

| Requirement | Citation | Evidence for an agent |
|---|---|---|
| Risk analysis | 45 CFR 164.308(a)(1)(ii)(A) | Risk analysis that names the agent, its data flows, models and tools |
| Risk management | 164.308(a)(1)(ii)(B) | Measures chosen for each agent risk, with owners |
| Information system activity review | 164.308(a)(1)(ii)(D) | Regular review of agent activity logs and alerts |
| Business associate contracts | 164.308(b), 164.502(e) | BAAs covering each model, platform and tool provider that handles PHI |
| Access control and unique user identification | 164.312(a)(1) | Agent has its own identity; actions record the user it acted for |
| Audit controls | 164.312(b) | Logs of agent reads and writes of PHI, by agent, version and user |
| Integrity | 164.312(c)(1) | Controls that stop agents altering PHI improperly; validation before writes |
| Person or entity authentication | 164.312(d) | Agent and user authentication records |
| Transmission security | 164.312(e)(1) | Encryption in transit to model APIs and tools; private connectivity where used |
| Minimum necessary | 164.502(b) | Retrieval and tool scopes limited to what the task needs |
| Documentation retention | 164.316(b)(2) | Policies and assessments kept six years |

## Why is the on-behalf-of user so important?

HIPAA's access control standard requires unique user identification. If an agent reads a patient record using a shared service account, your audit log shows the service account, not the clinician or patient who asked. That makes access reviews and breach investigations much harder.

Use delegated on-behalf-of access where the agent acts for a person, so both identities are recorded. Give the agent its own identity for background tasks. See [on-behalf-of delegation for AI agents](/blog/ai-agent-identity-on-behalf-of-delegation).

## What should a BAA cover for agents?

Every party that creates, receives, maintains or transmits PHI on your behalf is generally a business associate. For an agent, that can include:

1. The model provider or cloud model service.
2. The agent platform hosting the agent.
3. Vector databases or memory stores holding PHI.
4. Third-party tools and MCP servers the agent sends PHI to.
5. Observability tools that receive prompts and responses.

Major clouds list their managed model services among HIPAA-eligible services under their BAAs, and some model API providers offer BAAs on specific plans. Check that the exact service, region and features the agent uses are in scope. Features such as stored conversations, file storage or certain tools are sometimes excluded.

## How do you apply minimum necessary to an agent?

Agents tend to over-fetch. A retrieval step that pulls a full chart when the task needs one lab result breaks the minimum-necessary principle.

- Scope tools to the fields the task needs.
- Filter retrieval by the requesting user's access rights.
- Redact identifiers before model calls where the task does not need them; see [redacting PII at the gateway](/blog/how-to-redact-pii-in-llm-prompts-at-the-gateway).
- Avoid persistent agent memory for PHI unless it is covered by the same controls as the source system.
- Apply the [least-privilege checklist](/blog/least-privilege-checklist-for-ai-agents) to every tool.

## What logs should you keep, and for how long?

Log each agent access to PHI with timestamp, agent ID and version, the user it acted for, the resource, and the action. Keep prompts and responses containing PHI only where needed, in access-controlled storage. HIPAA requires required documentation to be kept six years; set log retention in your policy and apply it consistently. See [how long to retain AI agent logs](/blog/how-long-to-retain-ai-agent-logs).

## Is the HIPAA Security Rule changing?

HHS published proposed Security Rule updates in January 2025, including stronger requirements for asset inventories, network maps, encryption and multi-factor authentication. As of November 2026, the federal regulatory agenda projects final action for July 2027. The proposals are not yet law. An agent inventory and data-flow map would help under either version. The [framework mapping guide](/blog/how-to-map-ai-agent-controls-to-compliance-frameworks) shows how to reuse the same evidence across HIPAA, SOC 2 and ISO/IEC 42001.

## How Opencontroller does this

Opencontroller provides delegated on-behalf-of identity and agent workload identity federation, so logs show both the agent and the user. The LLM Gateway puts control in the path to allow, route, redact or block, and attributes activity and cost per agent and version. Auto-discovery and the governed registry give you the agent inventory a risk analysis needs, and Posture flags risky permissions. Opencontroller can run on customer-controlled infrastructure. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Do I need a BAA with an LLM provider if an agent sends it PHI?**
Generally yes. Confirm that the specific service and features the agent uses are covered.

**Is de-identifying data before the model call enough?**
It reduces risk, but only if de-identification meets HIPAA's Safe Harbor or Expert Determination standard and the agent cannot re-identify records through its tools.

**Does HIPAA require human review of agent decisions?**
HIPAA does not require it directly. Your risk analysis may call for it on high-impact actions, such as changes to records.

## Key facts

AI agents that create, receive, maintain or transmit protected health information (PHI) fall under the HIPAA Security Rule like any other system. Evidence to keep: a risk analysis that includes the agent (45 CFR 164.308(a)(1)(ii)(A)); information system activity review (164.308(a)(1)(ii)(D)); unique identification for the agent and the user it acts for (access control, 164.312(a)(1)); audit controls recording agent activity on PHI (164.312(b)); integrity controls (164.312(c)(1)); person or entity authentication (164.312(d)); transmission security (164.312(e)(1)); business associate agreements with model and platform providers that handle PHI (164.308(b), 164.502(e)); and minimum-necessary limits on what the agent retrieves (Privacy Rule, 164.502(b)). HIPAA documentation must be kept six years (164.316(b)(2)). HHS proposed Security Rule updates in January 2025; final action is projected for July 2027. This is evidence for HIPAA controls, not a compliance determination. Opencontroller by Lyzr provides delegated on-behalf-of identity, gateway redaction, cost and activity attribution per agent, and policy as code.

## Related

- [How do you redact PII from LLM prompts at the gateway?](https://opencontroller.lyzr.com/blog/how-to-redact-pii-in-llm-prompts-at-the-gateway.md)
- [What does least privilege look like for an AI agent?](https://opencontroller.lyzr.com/blog/least-privilege-checklist-for-ai-agents.md)
- [How long should you keep AI agent logs and traces?](https://opencontroller.lyzr.com/blog/how-long-to-retain-ai-agent-logs.md)
- [More on Evidence and frameworks](https://opencontroller.lyzr.com/blog/cluster/evidence-and-frameworks)
