How to Evaluate AI Vendors for Healthcare Compliance
Saying "we're HIPAA compliant" costs a vendor nothing. There's no license, no government registry, no stamp of approval. Any vendor can put it on their website. Many do — with varying degrees of actual compliance behind the claim.
For healthcare organizations evaluating AI vendors, the question is never "are you HIPAA compliant?" The real questions are specific, verifiable, and frankly, most vendors will stumble on at least a few of them.
Here's the full checklist. Save this before your next vendor demo.
1. Business Associate Agreement (BAA): Non-Negotiable Starting Point
If a vendor is processing, storing, or transmitting Protected Health Information (PHI) on your behalf, you need a signed BAA before data touches their system. This is HIPAA law, not a preference.
Questions to ask:
- Do you provide a BAA? (If the answer is no or hesitant, stop here.)
- Is your BAA standard or do you require redlines? Some vendors offer non-negotiable BAA templates. Know this upfront.
- Does the BAA cover all subprocessors who may touch PHI?
- What's your breach notification timeline under the BAA? (HIPAA requires 60 days from discovery; some BAAs are tighter.)
One important alternative: some AI tools are designed from the ground up to operate without PHI. Luma's Safe Harbor approach, for example, generates prior authorization documentation using only a patient's name and clinical data — no dates of birth, no SSNs, no full addresses. Under HIPAA's Safe Harbor de-identification standard, this combination doesn't constitute PHI, which means a BAA isn't required. That's a legitimate architectural choice, not a workaround. But you should understand which approach a vendor is using and verify the logic holds.
2. Data Handling: Where It Goes, How Long It Stays, What It's Used For
This is where well-meaning vendors reveal uncomfortable practices. Ask directly:
- Data retention: How long do you retain patient data processed through your system? 30 days? 1 year? Indefinitely?
- Training data: Is patient data used to train or fine-tune your AI models? Many vendors do this by default. Some allow opt-out. Few advertise it.
- Data residency: Where are servers located? Does data leave the U.S.?
- Data deletion: Can patient data be deleted on request? What's the documented process?
The training data question deserves special attention. Using de-identified patient data for model improvement is a gray area with evolving guidance. Using identifiable PHI for training without explicit authorization is a HIPAA violation. Ask the question explicitly and get the answer in writing.
3. Security Certifications: Verified, Not Self-Assessed
HIPAA has no certification process — you can't be "HIPAA certified." But there are adjacent certifications that provide meaningful third-party verification of security practices.
SOC 2 Type II is the baseline expectation for any cloud vendor handling sensitive data. Type I is a point-in-time assessment. Type II covers a sustained period (typically 6–12 months) and is meaningfully harder to obtain. Ask for the report, not just the badge.
HITRUST CSF is the healthcare-specific framework. It's more rigorous than SOC 2 and designed explicitly for HIPAA-adjacent compliance. A HITRUST r2 certification (formerly called "validated") is the gold standard. Many vendors say they're "HITRUST-aligned" — that phrase has no formal meaning. Ask whether they're certified, and at what level.
ISO 27001 is an international information security standard. More common in enterprise software vendors with global customers. Less healthcare-specific but indicates mature security program management.
Red flag: a vendor who offers to share their self-assessment or internal audit instead of a third-party report. Self-attested security posture is worth almost nothing.
4. Encryption Standards: At Rest and In Transit
Encryption is table stakes, but the details matter:
- In transit: TLS 1.2 minimum, TLS 1.3 preferred. Ask explicitly — some vendors still support older protocols for "compatibility reasons."
- At rest: AES-256 is the standard for stored data. Confirm that encryption applies to backups and not just primary storage.
- Key management: Who controls encryption keys? Are they managed by the vendor, a third-party KMS (like AWS KMS), or can you bring your own keys?
- PHI in logs: Does the system log PHI in plaintext in application or error logs? This is a common compliance failure that's easy to overlook.
5. Breach Notification: Process and Timeline
HIPAA requires notification to covered entities within 60 days of breach discovery. The more important question is what happens before that 60-day window.
Ask vendors:
- What's your internal incident response process?
- How quickly will you notify us if a breach is suspected (not confirmed)?
- Do you have a documented incident response plan? Can we see the relevant sections?
- Have you had any breaches or security incidents in the past 24 months? (They're required to tell you if they have.)
The HHS Breach Portal is public. You can search it. If a vendor has had reportable breaches, they'll appear there. Check before the demo — it changes the conversation.
6. Subprocessor Transparency
Most AI vendors use subprocessors: cloud infrastructure (AWS, GCP, Azure), AI model providers (OpenAI, Anthropic, Google), analytics tools, monitoring services. Each one that touches PHI needs to be covered under your BAA.
Ask for a complete subprocessor list — not a description, an actual list. Questions to ask about each material subprocessor:
- Do you have a BAA with each subprocessor?
- Are subprocessors restricted from using your data for their own model training? (OpenAI's API terms address this; consumer-tier products don't.)
- How are you notified when a subprocessor changes, and what's the process for objecting?
This is one of the most commonly overlooked areas. A vendor's own infrastructure may be locked down — but if they're sending PHI to a large language model provider's API without a proper BAA in place, the compliance posture collapses at that link in the chain.
7. Access Controls and Audit Logging
Who inside the vendor's organization can access your data? What's the process?
- Is access to customer data role-based and least-privilege?
- Is employee access audited and logged?
- Do you have audit logs you can provide to customers in the event of a compliance review?
- What's the background check process for employees with data access?
HIPAA requires covered entities and business associates to implement audit controls — and those controls need to extend through the vendor relationship. If a vendor can't tell you who has access to your data internally, that's a gap.
Red Flags to Watch For in Demos
Sales conversations often reveal more than the security questionnaire. Watch for these signals:
- "We're fully HIPAA compliant" with no supporting documentation offered
- Reluctance to provide a BAA template early in the process ("we can discuss that when you're closer to signing")
- Vague answers about data retention ("we follow industry best practices")
- Security certifications described as "in progress" with no timeline
- No clear answer on model training data usage
- Consumer-tier AI model APIs used in a healthcare product without explicit API agreements
Questions to Ask at the End of Every Demo
Five questions that cut through most vendor ambiguity fast:
- "Can you send me your current SOC 2 Type II report?"
- "Is patient data ever used to train or improve your models?"
- "Who are your material subprocessors, and do you have BAAs with each?"
- "How quickly would you notify us of a suspected breach?"
- "Can I see your standard BAA template today?"
Good vendors answer these without hesitation. The answers themselves matter less than the confidence and specificity behind them.
For more on how HIPAA Safe Harbor approaches can reduce compliance complexity for AI documentation tools specifically, see the related posts on the Luma blog. The architecture choices a vendor makes upfront — PHI vs. non-PHI workflows — affect everything downstream, including your own compliance obligations and audit exposure.
Sources: U.S. Department of Health and Human Services, HIPAA Security Rule (hhs.gov); HHS Office for Civil Rights, Breach Portal (ocrportal.hhs.gov); HITRUST Alliance, CSF Certification Programs (hitrustalliance.net); AICPA, SOC 2 Framework (aicpa.org); National Institute of Standards and Technology, HIPAA Security Rule Crosswalk to NIST Cybersecurity Framework (nist.gov); American Medical Association, "Navigating HIPAA Compliance for AI Vendors" (ama-assn.org).