To answer a hospital's AI security questionnaire with proof, stop sending a folder of attestation documents and start sending verifiable evidence. Build one reusable evidence pack that shows, item by item, that a given AI output came from a stated model version on intact data, who accessed which record and why, and how the reviewer can check each claim without taking your word for it. A questionnaire is a request for assurance. Most vendors answer it with promises. The vendors that close fast answer it with things the hospital can verify, then reuse that pack for the next hospital instead of starting over. Healthcare is the most expensive industry in the world to get security wrong: the average healthcare data breach reached 7.42 million dollars in 2025, the highest of any sector for the fourteenth year running.[1] That number is why a small health-tech vendor gets handed a review that can run to hundreds of questions.
This guide covers what those questionnaires are really testing, why a stack of certification PDFs answers the wrong question, and how to build a reusable evidence pack that turns a stalled review into a signature. RankShieldMD attests to the provenance of clinical-AI decisions and access, and never renders the decision itself. It is non-device, PHI-free by construction, and supports your obligations without being a clearance. See how clinical AI provenance works and how it fits verifiable AI versus AI governance.
What a hospital AI security questionnaire actually asks
A hospital AI security questionnaire is a structured request for evidence that your product will not become the hospital's next breach, audit finding, or patient-safety event.
It bundles standard vendor-security questions, encryption, access control, incident response, and business associate agreements, with newer AI-specific ones: which model version runs, how outputs are logged, whether a human reviews them, and how the hospital can reconstruct what your AI did if a decision is later questioned. The AI-specific half is where most vendors get stuck. A hospital's reviewers are not asking whether you have good intentions. They are asking whether, months from now, they can answer a regulator, a payer, or a plaintiff who asks what your model did on a specific patient's data. In practice the reviewer is testing three things at once: your program (do you have controls), your data handling (does protected health information stay protected), and your evidence (can anyone verify what actually happened). The first two are familiar. The third is the one no certification checkbox fully covers, and the one where clinicians using AI are most exposed. 81 percent of physicians now use AI professionally, up from 38 percent in 2023,[2] and roughly two thirds of US hospitals on major EHRs already run ambient AI,[3] so the hospital sending you a review has likely already been burned by a tool that could not show its work.
What hospitals are really trying to reduce
Hospitals are trying to reduce unverifiable trust. Every question on the review exists to shrink the set of claims they have to accept on faith.
A SOC 2 report and a HITRUST certification reduce that set: they show an independent auditor examined your controls over a period. They are worth having. But they attest to your program, not to any single AI decision, so they leave the highest-risk question open. That gap matters because the Office for Civil Rights ties much of its HIPAA enforcement back to risk-analysis failures,[4] and the proposed update to the HIPAA Security Rule would require covered entities to maintain a written asset inventory and network map of everything that touches electronic protected health information.[5] A hospital inherits your risk. If your AI touches data that feeds a clinical workflow, the hospital's own risk analysis now has to account for your product, and "the vendor is SOC 2 certified" is a thin answer when the reviewer's real fear is a decision no one can reconstruct. Here is the distinction most vendor content misses. SOC 2 and HITRUST answer whether this company runs a credible security program. The hospital's hardest question is different: if this specific AI output is challenged, can anyone prove it came from the model you claim, on data that was not altered. A certification cannot answer that. Verifiable evidence can. RankShieldMD attests to that evidence and never renders the clinical decision itself, which is what keeps it non-device by design.
Building the evidence layer a certification cannot supply?
Request early access →Build a reusable evidence pack
An evidence pack is a single, reusable set of artifacts you assemble once and send to every hospital, structured so each questionnaire item maps to something the reviewer can check.
It pairs your certifications with verifiable evidence: a per-decision attestation, a PHI-free access record, and a plain verification recipe. Built once, it turns each new review from a blank page into a copy, paste, and personalize job. Assemble it in five parts. First, the program layer: your SOC 2 report, HITRUST status, and business associate agreement template. Second, the decision layer: a sample attestation showing that a given output came from a stated model version on inputs whose integrity can be checked. Third, the access layer: a sample access record proving who touched which record and why, using one-way digests rather than identifiers, so the log itself holds no protected health information. Fourth, the verification recipe: a short, literal set of steps a hospital reviewer can run to confirm any single claim independently. Fifth, an honest scope statement: what your product attests to, and what it explicitly does not do. Keep every artifact PHI-free by construction, so you can share the pack without a new data agreement for every prospect. If you are small, start with the decision and access layers, because they are the two a certification cannot supply and the two that close the trust gap fastest.
Attestation and a SOC 2 PDF are not the same as proof
A certification proves your program was sound over a past window. A per-decision attestation proves what a specific AI output was, at the moment it happened.
Hospitals need both, but they are different instruments answering different questions, and treating a SOC 2 PDF as if it settles the model-trust question is the most common way a vendor's review stalls. The table below is the artifact to put directly in your evidence pack. It gives a reviewer a clean way to see why you are sending two kinds of assurance, not one. Read the rows as complements, not competitors. The left column is table stakes, so ship it. The right column is what turns "trust us" into "check us," and it is the column almost no vendor supplies. That is the information gain here: not a better certification, but a second, verifiable layer underneath it.
| Attribute | Certification (SOC 2 / HITRUST) | Verifiable per-decision evidence |
|---|---|---|
| What it proves | A credible security program existed over a period | A specific output came from a stated model on intact data |
| When it is created | Annually, by an external auditor, after the fact | The moment the decision or access happens |
| Independently checkable | By trusting the auditor's report | By the hospital, using a published verification recipe |
| Exposes PHI | Not applicable (program level) | No, PHI-free by construction (one-way digests only) |
| Answers "did this output come from this model" | No | Yes |
| Reusable across hospitals | Yes | Yes |
When to lead with verifiable evidence
Lead with verifiable evidence whenever a review stalls on trust rather than paperwork. Match the artifact to the fear.
If a hospital's security team keeps circling "how do we know your model produced this," send the per-decision attestation first, because it answers the exact question a certification cannot. If the sticking point is data exposure, lead with the PHI-free access record, because it proves accountability without widening the protected-health-information footprint. If procurement is comparing you against a competitor who only offers certifications, lead with the verification recipe, because letting the reviewer check a claim themselves is a difference they can feel. And if the reviewer is short on time, lead with the comparison table, because it reframes the entire conversation in one screen. A security review is a series of specific worries wearing the costume of a checklist, and the vendor who answers the worry underneath each question, with something checkable, is the vendor who gets the countersignature. One caution, stated plainly. Verifiable evidence shortens reviews and narrows disputes. It does not certify a hospital as compliant, and it is not a clearance. Position it as proof that supports the hospital's own obligations, offered so their reviewers have less to take on faith.
