AI Business Associate Agreements: What to Ask Vendors First
A health-related company should check how an AI vendor handles patient data, its sub-processors and the signed agreements before any agent touches records.

An AI business associate agreement is the written contract that makes a vendor accountable for health information when its tools create, receive, store or transmit that data for your company. Before an AI agent touches patient records, ask the vendor for the signed agreement, the current sub-processor list, a data flow map covering memory, logs and model calls, and the controls that limit what the agent can do. Whether a formal agreement is required depends on your jurisdiction and your role, so confirm the answer with counsel.
Key takeaways
- Ask whether the exact product and tier you will use is covered by the agreement, not only whether the vendor company signs such agreements.
- Get a named, current list of sub-processors and the terms that flow down to each one.
- Confirm what the agent stores as memory, prompts, transcripts and logs, and how long each category is kept.
- Require written approval steps before the agent sends anything to a patient or changes a clinical or billing record.
- Treat the contract as one control among several, backed by scoped access, redaction and logs you have tested.
What an AI business associate agreement actually covers
In the United States, a business associate agreement (BAA) is the contract HIPAA requires when a vendor creates, receives, maintains or transmits protected health information on behalf of a covered entity or another business associate. The term is often used loosely for AI products, but the duties behind it are what matter. The vendor must safeguard the data, limit its uses, report security incidents and make sure its own subcontractors accept the same restrictions. An agent does not change that logic. If it reads a patient message, stores a summary in memory or drafts a reminder that mentions a diagnosis, the vendor and every component it uses are handling health information.
Brazilian companies face a parallel question under the LGPD, which treats health data as sensitive personal data and expects written terms with any processor that handles it. The instrument is different, and a data processing agreement often covers the same ground. Which document you need, whether one is enough, and how it interacts with professional secrecy rules are legal questions for your counsel. A vendor's marketing page cannot settle them.
Map the data before you read the contract
Contracts describe what a vendor promises. A data map describes what will actually happen once an agent is running. Before negotiating, write down every place patient information could travel: the input channel, the model call, any retrieval store, the agent memory, the logs, backups, support tooling and any human review queue. Most surprises sit at the edges. A debugging tool that copies prompts, a transcription service added for voice calls, or an analytics plugin can each receive data that nobody listed in the original scope.
The NIST AI Risk Management Framework is a useful structure for this exercise because it treats AI risk as something to govern, map, measure and manage across the system's life, not only at purchase. Keep the map to one page and update it whenever a channel, tool or vendor changes.
- Input: where requests, messages and documents enter the agent.
- Processing: which model, retrieval source and tools see each request.
- Memory: what is remembered, for how long, and who can view or edit it.
- Records: which logs, transcripts and exports are created, and where they are stored.
- People: which staff, vendor employees or contractors can view any of the above.
Questions to ask about data handling
Once the map exists, ask the vendor to answer each point in writing. A vague answer is useful information in itself, so note it and keep asking.
- Which categories of patient information will the agent see, and which fields can be removed or masked before a model call?
- Is customer data used to train or improve models, and can that be disabled in writing for our account and product tier?
- Where is memory stored, who can read it, and how is one person's data found and deleted when they ask?
- How long are prompts, outputs, tool calls and logs retained, and can retention differ by data type?
- What happens to our data at the end of the contract, and how does the vendor confirm deletion, including backups?
- Are logs redacted so that the audit trail does not repeat the sensitive content it was meant to record?
Retention deserves its own decision. Audit records are needed to explain what an agent did, but a complete transcript of every patient conversation may be more than you need. The question of how long to keep agent logs and transcripts should be answered by your policy and counsel, then written into the vendor terms.
Questions to ask about sub-processors and where data lives
A sub-processor is any company that handles data on the vendor's behalf: a cloud host, a model provider, a speech recognition service, an email or SMS sender, a monitoring tool or a support platform. Each one widens the set of organizations that could be involved in an incident, so the list should be specific. A sub-processor review is one of the few checks that can be done entirely from documents before any patient data moves.
- Request the named list of sub-processors, with the function each performs and the country or region where it processes data.
- Ask how you will be notified before a new sub-processor receives your data, and whether you can object or leave without penalty.
- Confirm that each sub-processor is bound by terms at least as protective as the vendor's own, and ask to see the clause.
- Ask whether the model provider is a direct sub-processor, and whether its terms for the specific product you will use cover health information.
- Ask whether vendor support staff can view prompts, memory or records, and under which approval and logging rules.
Residency questions often surface here. If your policy or local rules limit where health information may be stored or processed, ask for the regions in writing, including for backups and logs, not only for the primary database. Treat a refusal to share the sub-processor list as a blocker, not a negotiating point.
A pre-launch sequence for patient information
Use the following sequence before an agent is allowed to read real records. If a step does not produce written evidence, pause at that step.
- Step 1: Name the single task the agent will perform and the smallest set of patient fields it needs to perform it.
- Step 2: Collect the signed agreement, the sub-processor list and the data flow map from the vendor, and have counsel review them.
- Step 3: Confirm in writing how memory, prompts, transcripts and logs are retained, deleted and redacted.
- Step 4: Give the agent scoped credentials that cannot export records in bulk or reach systems outside the task.
- Step 5: Define which actions need a person to approve, such as sending a message to a patient or changing a record, and log each approval.
- Step 6: Test with synthetic records, then check that the logs do not expose the test content beyond what the audit needs.
- Step 7: Name an incident owner and document how to pause the agent and notify the people your contracts require.
- Step 8: Schedule a recurring review of sub-processors, permissions and retention settings.
Illustrative scenario: a small clinic group wants an agent to draft appointment reminders from its scheduling system. The team requests the vendor's signed agreement, the sub-processor list for the messaging provider, and written confirmation that message bodies are not kept in agent memory. The agent may read only the appointment date and the first name, and any reminder that mentions a procedure goes to a staff member for approval before it is sent. This scenario is invented to show the sequence and does not describe a real client.
A claim that a vendor is compliant is not the same as evidence. Ask what documents, controls and test results support the claim, and keep copies.
How Be Human handles this in a managed setup, and where to start
Be Human designs, implements and operates personalized AI infrastructure in which each agent keeps its identity, memory, channels, tools, permissions and operating context under human control. A first engagement defines one useful role, the memory and tools it needs, the approval boundaries and an observable success criterion, so the contract and data questions above are answered against a specific data flow rather than a general product. Managed operations can then include monitoring, recurring evaluations, integration maintenance, permission review and incident response, and each of these leaves records that your counsel and compliance team can inspect. Be Human does not replace your legal advice or the obligations your vendors accept in their own agreements.
If you are not sure which task is safe to hand to an agent first, take the free Be Human AI diagnostic to find the first task worth handing to an agent.
Frequently asked questions
Is a business associate agreement the same as a data processing agreement? They overlap but are not identical. A BAA is the specific contract HIPAA requires in the United States when a vendor handles protected health information for a covered entity or another business associate. A data processing agreement is the common term in data protection regimes such as the LGPD, and it sets the processor's duties for personal data. Your counsel should decide which instrument applies to your situation.
Does an AI vendor that says it is HIPAA compliant need a BAA with us? A statement on a website does not create a contractual duty between you and the vendor. Ask for a signed agreement that names the product, the tier and the services you intend to use. If the vendor will not sign one for the relevant service, do not send patient information to it until counsel has reviewed the options.
Can we use a general AI chat tool with patient information if we turn off training? Turning off training is one setting, and it does not answer the other questions about retention, sub-processors, access by vendor staff, residency and deletion. Check the agreement and the data flow map first, and confirm that your internal policy allows the use. For most teams, a tool that has not been reviewed should not receive identifiable patient information.
What should we do if a vendor refuses to share its sub-processor list? Treat the refusal as a finding in its own right. You cannot assess risk for organizations you have never been told about, and an incident response plan cannot name parties you cannot identify. Pause the rollout, record the refusal, and either find a vendor that provides the list or limit the agent to data that contains no patient information until the question is resolved.
Your next step
See where an AI agent fits in your company first.
Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.
Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.


