Can AI review a vendor security questionnaire? Yes, it can triage one. In about three minutes, a general-purpose AI assistant can flag evasive answers, rate each risk, spot contradictions across sections, and draft follow-up questions. What it can't do is tell you whether the vendor is telling the truth.
That's the short version of a live experiment Trustero CEO Phillip Liu ran at Onspring GRC Day in Denver. He fed real documents to Claude, an off-the-shelf AI assistant, and showed the unedited results. The companies and documents were fictional. The AI outputs were not.
Why vendor questionnaires eat your week
Third-party risk management is mostly reading. In the Onspring 2026 GRC Benchmarking Report, 25.9% of GRC teams named evidence collection and documentation as their most time-consuming work, and another 16.5% named third-party reviews. Security questionnaires sit right in the middle of both.
The stakes are going up. Verizon's 2026 Data Breach Investigations Report found third-party involvement in 48% of breaches, up from 30% the year before.
The problem isn't length. It's that vendor answers are written to sound like answers. Common patterns include:
- The "yes, but." MFA is enforced for everyone, except the admin console.
- The reassuring qualifier. Engineers have admin access "so we can respond quickly."
- The vendor's own definition. "No material incidents," where "material" is their word, not yours.
- The contradiction three sections apart. An answer in section A quietly conflicts with one in section G.
Someone has to find those, rate them, and write the follow-up email. That someone usually has twenty more vendors in the queue. And as we've written before, TPRM is a connect-the-dots problem, not a questionnaire problem: the questionnaire is one input, not the decision.
What happened when we asked AI to triage one
The scenario: Redrock Outdoor Co., a fictional 1,250-employee retailer, is onboarding Halcyon Ledger, a fictional payroll SaaS vendor. Halcyon will hold SSNs, bank accounts and salaries for every employee. Halcyon's Director of IT answered all 30 questions on Redrock's security questionnaire. Eleven answers were some version of "yes, but."
The prompt asked Claude to flag red flags and dodges, rate each high, medium or low, draft follow-ups, and list which answers looked fine to skip.
The result, in about three minutes:
- 25 rated flags, including 11 High and 9 Medium, plus contradictions across sections (A3 vs. B4 vs. G3).
- Standing production admin access for engineers (High). The vendor pointed to logging. Claude noted that logging is detection, not prevention.
- "No material incidents" (High). Flagged as a classic dodge because the vendor defines "material."
- TLS 1.0 on legacy bank SFTP links (High). Flagged as technically inconsistent, since SFTP runs over SSH. Which is it?
- A skip list. AES-256 encryption via KMS, quarterly access reviews and SAML SSO looked fine.
- 24 follow-up questions, grouped by section and ready to send.
The skip list matters as much as the flags. Knowing what not to chase is how one part-time analyst gets through a queue.
Where AI stops: it can spot a dodge, not a lie
The more useful half of the experiment was watching where the AI hit its limits.
Truth is out of scope. AI flags inconsistency, not veracity. "No material incidents" is exactly as unverifiable after the review as before. In Claude's own words, "only the vendor, or public breach databases, can answer that."
It doesn't know your risk appetite. Every High was rated against generic expectations for a payroll vendor, not against Redrock's policy. A $1M cyber insurance limit might be fine for you. Claude said as much, unprompted.
The vendor has AI too. More and more questionnaire answers are drafted by a model. A model grading a model is a fast way for two systems to agree with each other. Someone still has to pick up the phone.
A questionnaire is a snapshot. Even a perfect review tells you what the vendor said on one day. The SecurityScorecard 2026 Supply Chain Cybersecurity Trends Report found 67% of organizations still rely on static, point-in-time audits to assess vendors.
How to use AI for vendor questionnaire review
- Give it your standard, not just the document. Ask it to compare answers against your requirements and data sensitivity, not generic best practice.
- Start with the evidence you already have. Review the SOC 2 and other documents first, then send questions only about the gaps. We walk through that approach in How to Build an Evidence-First, Gap-Driven TPRM Workflow.
- Ask for a skip list. Clean answers you don't have to reread are real time saved.
- Always ask what it couldn't determine. One sentence at the end of the prompt turns confident output into honest output.
- Keep a human between output and signature. Findings go to a person before they go to the risk register.
- Store the results somewhere that lasts. Flags, follow-ups and vendor responses belong in your GRC platform, not a chat window.
If your team reviews the same kinds of vendor documents every week, that review can become a repeatable workflow. Trustero TI Playbooks let TPRM analysts describe a review in plain language, such as checking a vendor's SOC 2 report, questionnaire and pen test results, and run it on every new vendor.
The mental model from the talk holds up: treat AI like a sharp new intern. It reads everything, remembers nothing, and signs off nothing. Your job is to decide what's acceptable, verify what matters, and own the call.
FAQ
Can AI detect if a vendor is lying on a security questionnaire? No. AI can flag evasive wording and internal contradictions, but it can't verify claims against reality. Verification still takes a follow-up call, evidence requests or outside sources.
How long does AI take to review a security questionnaire? In this test, about three minutes for a 30-question questionnaire, producing 25 rated flags, a skip list and 24 follow-up questions.
Should AI decide whether a vendor is approved? No. AI rates risk against generic expectations unless you give it your policy. Approval, risk acceptance and sign-off stay with a person.
Trustero builds a multi-agent AI system for GRC teams that works alongside the GRC platforms you already use. See how Trustero AI works.