Four AI BAA Clauses That Raise Red Flags for Healthcare Buyers
Last week, a client asked me to review an AI vendor's BAA before signing the contract.
At first glance, the agreement referenced HIPAA Compliance and included standard security language.
Then I reviewed the actual terms.
The security commitments were too vague.
Several provisions gave the vendor more control over patient data than I would advise a healthcare organization to accept.
The BAA did not create enough confidence for me to recommend signing it.
I spend a lot of time reviewing AI vendors on behalf of healthcare organizations.
Healthcare runs on trust.
Patients trust their providers.
Providers trust the technology they use.
Healthcare organizations must trust the vendors they allow to handle patient data.
That trust has to be supported by clear contractual commitments.
Deals stall when a vendor’s BAA does not meet the buyer’s security and compliance requirements.
Broad data rights and vague commitments signal that the vendor’s compliance program may not be ready for enterprise healthcare buyers.
As I reviewed this BAA, I flagged four provisions that needed to be amended before I could recommend signing.
I advised my client not to accept the BAA as written.
If the vendor was unwilling to meet those requirements, my recommendation would be to find another vendor.
1. Broad Rights to Use De-Identified Data for AI Training
One provision allowed the vendor to use aggregated and de-identified data for internal business purposes, including artificial intelligence and machine learning training.
This is becoming increasingly common in AI vendor agreements.
I would not advise a healthcare organization to give a vendor unrestricted rights to use de-identified patient data for AI training.
For an AI vendor, that language is too broad without clear limitations on model training and development.
Before approving the BAA, I would require it to:
Define how the data will be de-identified
Prohibit attempts to re-identify the data
Limit the approved uses of the data
Healthcare organizations need to know where their patient data goes, how it is transformed, and how it may be used in the future.
The agreement did not provide enough information to establish that trust.
2. Unclear Boundaries Around AI Models
The BAA allowed the vendor to use data to improve and enhance its services.
I flagged that language immediately.
“Improve the service” could include model training, tuning, validation, or the development of future AI products. The BAA did not set clear limits.
When AI is involved, enterprise buyers want explicit boundaries around what data can and cannot be used for.
If those boundaries aren't clear, expect follow-up questions.
3. Limited Visibility Into the AI Supply Chain
Modern AI products rarely operate alone.
Most depend on cloud providers, AI services, APIs, etc.
Before recommending an AI vendor, I want to understand who touches the data and how those relationships are governed.
Healthcare organizations are responsible for understanding the full chain of vendors handling their patient data.
An AI vendor should disclose its relevant subcontractors, confirm that required BAAs are in place, and remain accountable for the providers it selects.
If PHI is shared with third-party AI providers, buyers want to know who those providers are, whether BAAs are in place, and whether patient data can be used for model training or improvement.
The more visibility you provide into your AI supply chain, the easier it is for buyers to trust how patient data is being handled.
If you want to know what else your buyers are scoring in vendor agreements, the Security Review Playbook breaks it down. Request it here.
4. Security Commitments That Need Specificity
The BAA referenced HIPAA Compliance and security safeguards.
A general statement of HIPAA compliance does not provide enough protection for the buyer.
The BAA should include clear commitments covering:
Encryption at rest and in transit
Multifactor authentication
Access controls and access reviews
Audit logging and monitoring
Vulnerability management
Penetration testing
Endpoint protection
Security awareness training
Incident response testing
Disaster recovery
Independent security assessments
Vendors that make specific security commitments demonstrate that they understand the responsibilities that come with handling PHI.
Relying on vague references to “appropriate safeguards” is too risky.
Why This Matters for Health Tech
I've reviewed enough AI vendors to notice a pattern.
The vendors that create the least friction are the ones whose BAAs already meet the security and compliance expectations of enterprise healthcare buyers.
If your BAA gives you broad rights to customer data, hides the AI supply chain, or makes vague security commitments, expect redlines, delays, and possible rejection.
I review these agreements for healthcare buyers.
When the terms create too much risk, I advise my clients not to sign.
Your next buyer’s security and compliance team may reach the same conclusion.
👉 If you want to know where your AI governance strategy stands before your next deal, take the AI Readiness Self-Assessment. Request it here.
Larry | Founder & Principal CISO, Inherent Security