An AI agent can book a patient appointment, check insurance benefits, or prepare a follow-up message. But can it handle protected health information without putting your organization at risk?

HIPAA-compliant AI agents require a properly configured deployment and signed business associate agreements (BAAs) with the vendors handling patient data. The agent's features alone don't establish compliance.

This guide explains the controls you need, which platforms offer BAAs, and what to verify before allowing an agent to access patient records.

What makes an AI agent HIPAA compliant?

A HIPAA-compliant AI agent needs appropriate contracts, technical safeguards, and organizational controls covering every system that handles protected health information (PHI).

Consider an agent that answers a patient's call, retrieves their insurance information, and updates a record. That single conversation could involve a voice provider, language model, cloud infrastructure, and electronic health record (EHR).

Each system may process PHI, and each needs appropriate protection.

The U.S. Department of Health and Human Services (HHS) doesn’t certify HIPAA compliance or recognize private HIPAA certifications, and that includes AI software. Vendors can advertise HIPAA support, but your organization remains responsible for its own compliance.

An AI agent needs five controls to be HIPAA compliant: a signed BAA, encryption, access controls, audit logs, and data retention and training controls. Here’s what each one requires:

1. A signed business associate agreement

A BAA is the first requirement to check when an AI vendor handles PHI on your behalf. It defines how the vendor may use, disclose, and protect that information.

For example, your scheduling agent might send patient details to a language model and save appointments in another system.

You need appropriate BAA coverage throughout that arrangement. Your organization generally signs agreements with its business associates, while those business associates must establish agreements with subcontractors that handle PHI.

HHS guidance also explains that cloud providers can qualify as business associates even when they store only encrypted PHI.

Check the agreement’s actual scope. A vendor may cover its core API but exclude browser automation, third-party connectors, or certain models.

2. Encryption at rest and in transit

Encryption protects patient information while it's stored and transferred. Ask the vendor how it protects the full flow, including:

  • Patient information retrieved from the EHR
  • Prompts and model responses
  • Call recordings and transcripts
  • Connected application data
  • Temporary files and logs

TLS 1.2 or newer for connections and AES-256 for stored data are common technical choices.

HIPAA treats encryption as an addressable Security Rule specification. That requires organizations to assess whether it’s reasonable and appropriate. If it is, they implement it. If not, they document why and use an equivalent alternative where reasonable.

Also check who manages encryption keys and whether the agent's subcontractors provide equivalent protection.

3. Access controls enforced outside the model

An AI agent should only access the patient records needed to complete its assigned task.

For example, a scheduling assistant may need appointment details and patient contact information, but there's no reason to give it unrestricted access to clinical notes. Similarly, an insurance verification agent shouldn't be able to modify medication records.

These restrictions need to be enforced by the connected systems. Even if you instruct an agent to retrieve scheduling information only, a poorly configured API could still expose sensitive records.

Use role-based permissions, scoped API credentials, and record-level restrictions to control what the agent can access or change. The agent should have no more access than the employee or process it's authorized to support.

Before deployment, test these boundaries by asking the agent to retrieve a record outside its permitted scope. The underlying system should block the request, regardless of the agent's instructions or response.

4. Audit logs for data access and actions

Audit logs help your team establish who accessed PHI, what they did, and when.

For AI agents, useful records include:

  • The user or agent that initiated the request
  • The system and records accessed
  • Tool calls and attempted changes
  • Approval decisions
  • Errors and unexpected activity

Some investigations may also require reviewing prompts or outputs. However, those records can contain PHI and need appropriate protection.

Avoid collecting full transcripts or patient details in logs without a clear purpose.

HIPAA requires certain compliance documentation to be retained for six years. It does not automatically impose a six-year retention period on every raw agent log.

Set log retention according to applicable legal requirements, your risk assessment, and the organization's documentation policies.

5. Data retention and model training controls

Find out where patient information goes after an agent finishes its task.

A vendor might retain prompts for a limited period, store transcripts, or send data to a subprocessing service. Ask whether patient information can be used for model training and which retention settings apply to your account.

Review the following:

  • Training: Whether PHI is excluded from model training and other secondary uses.
  • Retention: How long prompts, recordings, outputs, and backups remain stored.
  • Deletion: How the vendor removes data after the agreed period.
  • Subprocessors: Which additional companies receive or process PHI.
  • Configuration: Whether the required settings are available on your selected plan.

Zero data retention can be useful, but it isn't a universal HIPAA requirement. Some BAA-covered services use limited retention under defined conditions.

The important part is confirming that the actual data handling matches your agreement and compliance obligations.

Different kinds of agents touch different data, so our guide to the types of AI agents can help you scope access.

8 questions to ask before signing an AI agent BAA

A vendor's security page can help you build a shortlist. The contract and technical documentation tell you whether the product can handle your organization's PHI.

Bring these eight questions to your vendor review:

  1. Will you sign a BAA for our intended use? Confirm the agreement is available before sending patient information.
  2. Which products and features does it cover? Ask specifically about the models, connectors, tools, and automation features you intend to use.
  3. Which subprocessors handle PHI? Request the list and confirm appropriate downstream agreements.
  4. Where is patient information stored and processed? Ask about hosting regions, encryption, and access to stored data.
  5. Will our data be used for model training? Get the training and retention terms in writing.
  6. How does the agent enforce permissions? Request details about identity controls, access restrictions, and audit logs.
  7. How can we delete or export patient information? Include transcripts, backups, logs, and data held by subprocessors.
  8. How are security incidents reported? Confirm the vendor's notification obligations and the timelines in its contract.

A vendor may agree to sign a BAA while excluding the browser tool your agent needs to retrieve patient records. Even with other platform features covered, the planned workflow may not be approved for processing PHI.

Before testing with real patient data, request a list of features covered by the BAA. Review the vendor's SOC 2 Type II report, security assessment, and subprocessor list as supporting evidence. These documents don't replace the agreement or your organization's risk assessment.

Our comparison of agentic AI tools can help you narrow the field, but each platform still needs a separate compliance review before handling patient information.

AI agent platforms that offer HIPAA BAA options

Several AI platforms support HIPAA-regulated workloads, but BAA coverage depends on the product, account, and configuration.

Some offer covered APIs for building custom agents. Others provide ready-made voice workflows or the integration infrastructure those agents depend on.

General AI agent platforms

These platforms provide models, tools, or agent-building environments. Their BAAs typically cover specific services and configurations.

PlatformDocumented BAA coverageWhat to check
OpenAIChatGPT for Healthcare, ChatGPT for Clinicians, eligible regulated ChatGPT workspaces, and API services with Modified RetentionExcluded features, connected tools, and workspace settings
AnthropicHIPAA-ready Claude Enterprise and eligible API servicesHIPAA activation, approved features, and third-party connections
MicrosoftCopilot Studio and eligible Microsoft cloud services under Microsoft’s BAAService scope, connectors, and tenant configuration
Google CloudEligible Gemini Enterprise Agent Platform (formerly Vertex AI), Gemini, and Workspace services under Google’s BAASupported service list and feature-specific restrictions
AWSAmazon Bedrock and other HIPAA-eligible AWS services under its BAAModel and service eligibility, access controls, and logging

These platforms are useful for building AI agents that work with your existing applications, patient records, and access controls.

Check the BAA’s scope before connecting other tools. A platform’s agreement may not cover third-party services, and sending patient information through an uncovered connection could create a compliance risk.

Healthcare-specific AI agents

Healthcare-focused vendors offer agents designed for patient communication and administrative work.

Two examples are:

  1. Hippocratic AI: Builds patient-facing agents for outreach, follow-ups, and other healthcare workflows.
  2. Infinitus: Automates healthcare phone conversations, including administrative communication with payers.

Infinitus says on its security page that it signs a BAA when it processes a customer’s PHI. Hippocratic AI’s public materials reviewed for this guide didn’t spell out its BAA terms.

Request each vendor’s BAA documents directly before treating either product as approved for your deployment.

Also confirm which tasks require human review and how each vendor prevents an agent from performing clinical actions outside its intended role.

Voice AI agent platforms

Voice agents introduce additional services because a single call may involve telephony, transcription, a language model, and speech generation.

Three providers publish BAA options.

PlatformBAA availabilityImportant condition
Retell AIBAA available through its online agreement portalSign before sending PHI; configure recording and retention controls
SynthflowBAA executed for customers using its HIPAA SettingHIPAA Setting and agreement must be in place
VapiBAA through its paid HIPAA compliance optionEnable HIPAA mode and use compatible model, voice, and transcription providers

Check the complete call path before deployment. For instance, an appointment agent may use one vendor to answer the phone, another to transcribe speech, and a third to generate responses.

The primary voice platform's BAA doesn't automatically establish coverage for services you connect separately.

Integration platforms

Integration software connects an AI agent to EHRs, CRMs, calendars, and other business systems. Those connections can transmit PHI and account credentials, so the integration layer belongs in the security review.

Nango is one example. Its security documentation lists a BAA available with the Growth add-on.

Before using it, confirm that the selected deployment and connectors are covered. Also check whether the integration service stores copies of patient records, how long those copies remain available, and which credentials the agent can use.

Explore integrations

ChatGPT and Claude plans that support HIPAA-regulated use

Consumer ChatGPT and Claude subscriptions should not be treated as BAA-covered healthcare workspaces. Both companies offer specific business products with additional contractual and technical requirements.

ChatGPT and OpenAI API products eligible for a BAA

OpenAI lists several HIPAA-eligible offerings, including:

  • ChatGPT for Healthcare
  • ChatGPT Enterprise with Regulated Workspace
  • ChatGPT for Clinicians
  • Eligible API services with Modified Retention

These products require the applicable BAA and configuration.

ChatGPT Free, Go, Plus, Pro, and Business do not provide the same BAA coverage. Even within an eligible workspace, some features may remain outside the agreement.

Review OpenAI’s official eligibility list to confirm which agents, connectors, and services can handle PHI before enabling them.

Claude and Anthropic API products eligible for a BAA

Anthropic offers BAAs for approved API services and HIPAA-ready Claude Enterprise configurations.

However, standard Claude Enterprise access alone doesn’t establish BAA coverage. The organization’s Primary Owner must activate the HIPAA configuration and accept the agreement. Claude Free, Pro, Max, and Ultra aren’t covered.

Anthropic also documents what its BAA covers and excludes:

  • Claude Tag: Not available for HIPAA-ready organizations.
  • Claude Cowork in the cloud: Not covered.
  • Connectors and third-party data flows: Outside Anthropic’s BAA coverage. For Claude Code and Cowork, that includes every connector, MCP server, and plugin.
  • Claude Code and Cowork in local mode: Eligible under specified Enterprise configurations.

BAA coverage varies by feature, especially when agents use external tools or run tasks in the background. Check Anthropic’s BAA coverage table to confirm which features are approved for your workflow.

Make sure employees know which AI tools and accounts they can use with patient information. Personal AI accounts without the required BAA coverage should never receive PHI.

Healthcare workflows to consider for a first AI agent

Start with routine administrative tasks that staff can easily review. Limit the agent's access, require approval for sensitive actions, and apply HIPAA safeguards when handling PHI.

Here are four workflows to consider for your first AI agent:

Appointment scheduling

An agent checks appointment availability, confirms permitted patient details, and books or reschedules a visit. Limit its access to the information needed for scheduling and apply patient identity checks before disclosing appointment information.

Patient intake

An agent collects intake details and prepares a record for staff review. Use a controlled connection to write the information into the appropriate system. Check how sensitive answers are stored and whether anyone must review the record before submission.

Prior authorization support

An agent gathers required documents, checks payer requirements, and prepares a submission for an authorized employee. Keep clinical judgment and final approval with qualified staff. Log the records accessed and any changes made.

Patient follow-ups

An agent sends approved reminders or asks patients to complete follow-up tasks. Define clear escalation rules for messages involving symptoms, medical concerns, or requests that need a clinician.

Teams exploring AI employees can start with one supervised administrative responsibility before expanding access to other systems.

How to roll out a HIPAA-compliant AI agent

Before connecting a live patient account, test the entire workflow with synthetic data.

A practical AI agent rollout follows five steps:

  1. Choose one task. Define what the agent can read, change, and send.
  2. Complete the risk assessment. Map where PHI will move and identify every vendor involved.
  3. Confirm BAA coverage. Get the required agreements signed and configure approved services.
  4. Test the controls. Try unauthorized record requests, incorrect patient details, failed connections, and actions requiring approval.
  5. Monitor the pilot. Review logs, errors, staff interventions, and access permissions before expanding the deployment.

For example, a scheduling pilot could begin with one location and a limited set of appointment types.

Staff can review completed bookings, investigate mistakes, and confirm the agent never retrieves records outside its assigned permissions.

Once the workflow passes review, expand to additional locations or tasks.

For teams evaluating an AI teammate, Lindy Enterprise lists HIPAA support with a signed BAA, along with SSO, SCIM, and audit logs. Confirm the agreement's scope and connected services before using it with PHI.

The important result is a documented, tested workflow your compliance and IT teams can approve.

Frequently asked questions

1. Are any AI agents HIPAA compliant?

Yes, AI agents can support HIPAA-compliant use when the full deployment meets the applicable requirements. This includes appropriate BAA coverage, security controls, access restrictions, and organizational policies. HHS doesn’t certify HIPAA compliance for any product, including AI.

2. How do you make an AI agent HIPAA-compliant?

To make an AI agent HIPAA-compliant, sign the required BAAs, protect PHI, restrict data access, maintain appropriate audit controls, and complete a risk assessment. Test the agent's connected tools and approval rules before processing real patient information.

3. Does an AI agent need a BAA?

Yes, an AI agent vendor generally needs a BAA when it handles PHI on behalf of a HIPAA-covered entity or business associate. Downstream vendors that handle PHI also need appropriate business associate arrangements.

4. Is there a HIPAA-compliant version of ChatGPT?

Yes, OpenAI offers HIPAA-eligible ChatGPT products, including ChatGPT for Healthcare and eligible regulated Enterprise workspaces. These require the applicable BAA and configuration. Consumer ChatGPT plans and ChatGPT Business aren’t covered.

5. Can Claude process protected health information?

Yes, approved Claude API services and HIPAA-ready Claude Enterprise configurations can process PHI under Anthropic’s BAA. Consumer plans aren’t covered. Claude Tag isn’t available, and cloud-based Cowork isn’t covered.