How Do You Assess AI Vendor Exposure in Healthcare?

Most healthcare AI purchases treat vendor risk as a signature: get the business associate agreement executed, file it, move on. That covers the legal minimum but misses the actual exposure, which is what the model does with protected health information once it is ingested, whether the vendor's own subcontractors or sub-processors also touch that data, and whether anyone outside the vendor's sales team can describe how the model was trained or validated. A signed BAA answers whether the parties agreed to the rules. It does not answer whether the buyer actually knows what the system does with the data it receives, and those are two different questions that too many procurement processes collapse into one.

The stakes for getting this wrong are rising, not falling. Business associates were involved in an average of 34 percent of healthcare data breaches from 2018 through 2026, and that rate jumped to 43 percent of breaches reported in just the first six months of 2026, per HIPAA Journal's analysis of HHS Office for Civil Rights breach data. The share of individuals affected by business-associate-linked breaches rose even more sharply, climbing from 5 percent of affected individuals in 2015 to 65 percent in 2025.

Two of the largest healthcare breaches on record illustrate why this distinction matters in practice, not just in theory. The 2024 Change Healthcare attack and the 2025 Conduent Business Services incident both originated at a vendor rather than at the covered entity itself, and together the two breaches affected close to 255 million individuals. Neither organization was undone by a missing signature on a BAA. Both were exposed by a vendor relationship that nobody had mapped for what data actually flowed where, and by the time the incident was discovered the exposure had already happened.

The Coalition for Health AI, a nonprofit standards body known as CHAI, has published a draft framework for certifying independent assurance labs that test health AI models before they ever reach a health system's shortlist. The framework was built in partnership with the ANSI National Accreditation Board and is aligned to ISO 17025, the standard already used worldwide for testing and calibration laboratories generally, which gives it a recognizable accreditation path rather than a bespoke healthcare-only process.

A specific and useful detail in CHAI's draft program is that it requires assurance labs to disclose any conflict of interest with the model developer being evaluated, and to protect the vendor's underlying data and intellectual property during testing, per CHAI's own program announcement. That dual requirement matters for a buyer because it addresses the two most common objections vendors raise against independent testing: that it exposes trade secrets, and that the tester might be compromised by a commercial relationship with a competitor. CHAI's framework tries to answer both at once rather than trading one risk for the other.

The practical output for a buyer is the CHAI model card, a standardized document covering intended use, target patient population, maintenance requirements, and known risks or biases, functioning as something close to a nutrition label for a health AI product. CHAI assurance lab certification itself is not yet a legal requirement anywhere, but the model card's structure is a usable evaluation checklist today regardless of certification status: any vendor claiming to be enterprise-ready in healthcare should be able to answer the same questions the model card asks, whether or not a CHAI-certified lab has reviewed the answers yet.

Kriv AI's healthcare vendor exposure reviews combine the three pieces buyers typically do separately: a data-flow map of exactly what protected health information reaches the vendor and any subcontractors, a documented request for the vendor's HTI-1 source attributes evaluated against the FAVES criteria, and a CHAI-style model-card checklist covering intended use, validation population, and known limitations. Findings are mapped to HIPAA Security Rule controls and the NIST AI Risk Management Framework, in a form a compliance team can actually defend to an auditor or a board, not a generic vendor risk score with no supporting documentation behind it.

We do not have a named healthcare vendor-exposure client case study to publish yet, and we are not going to invent one to make this page look more finished than it is. What we do have is the same model risk and governance discipline applied across our regulated-industry practice, adapted specifically to the vendor-selection stage rather than only to models a health system builds internally.

This work sits inside Kriv AI's regulated-industry practice and is priced at our healthcare and enterprise floor of $200 per hour, with most vendor exposure engagements structured as fixed-scope reviews starting at an $8,000 floor depending on the number of vendors and systems in scope. See our full engagement pricing for the complete rate breakdown across service types.