New — The State of AI Hiring 2026. Every number sourced→
AI & Technology

AI Vendor Security Questionnaire for Hiring: 30 Questions to Ask Any AI Interview Vendor

A copyable AI vendor security questionnaire for hiring: 30 questions on data, retention, access, subprocessors and AI risk, plus red flags and documents.

The AI vendor security questionnaire. 30 questions on candidate data, access, AI model providers and incidents, plus the standards to start from.

TL;DR: A vendor security questionnaire is a list of questions you send a supplier to learn how it protects your data. For an AI interview vendor, cover what candidate data it collects, where it lives, who can see it, who else processes it, and whether it trains models on it. Use the 30 questions below, and get every answer in writing.

Updated 18 November 2026. Part of the AI interviewer RFP template kit.

What is a vendor security questionnaire?

It's a set of questions, usually yes/no plus evidence, that a buyer sends a supplier before signing. It shows how the vendor handles your data, who else sees it, and what happens when something goes wrong. For AI interviews the data is video, audio, transcripts and resumes, so the stakes are high.

When you need this

Use it after you shortlist, before you sign. For law-by-law questions, pair it with AI hiring compliance questions. Agencies: see the staffing-agency questions.

Start from a standard

You don't have to write a security review from scratch. Buyers, auditors and vendors already share a handful of standard questionnaires, frameworks and reports, and most mature vendors have completed at least one. Asking which ones they've done is a fast filter. A vendor with a finished questionnaire and a current report can answer quickly, while one with nothing to show may be working it out as you ask.

Ask vendors which of these they've completed.

FrameworkWhat buyers use it forSource
SIG and SIG LiteThird-party risk questionnaire; Lite is the shorter version for lower-risk suppliers.Shared Assessments
CSA CAIQYes/no cloud questionnaire from the Cloud Controls Matrix, 17 control domains (read 18 Nov 2026).CSA
NIST AI RMF 1.0 and the Generative AI Profile (AI 600-1)Voluntary AI risk framework, released 26 Jan 2023; the profile (26 Jul 2024) covers generative AI (read 18 Nov 2026).NIST
ISO/IEC 27001Information security management system standard.ISO
ISO/IEC 42001AI management system standard.ISO
SOC 2 Type I vs Type IIAuditor's report on controls. Type I is commonly described as controls at one date; Type II covers a period, often 3 to 12 months.AICPA

For EU or UK candidates, also ask for the processor contract terms in GDPR Article 28 (read 18 Nov 2026) and the transfer mechanism, such as the EU-US Data Privacy Framework.

The questionnaire

Copy these. Ask for written answers with evidence.

1. Data collected

An AI interview tool collects far more than a resume. It can hold video, audio, transcripts, scores and any demographic fields you switch on. What usually goes wrong is quiet extras: inferred traits, voiceprints or optional fields that were never meant to feed a score. A good vendor hands over a plain data inventory, marks what's optional, and says clearly what does and doesn't influence a candidate's result.

QuestionWhy it mattersA strong answer includes
Which candidate data do you collect: video, audio, transcripts, resumes, scores, demographic fields?You can't protect what you haven't listed.A data inventory, with optional fields marked.
Do you infer emotion or personality from face, voice or words?Inference can feed scores and raises legal risk.Yes or no, and what feeds a score.
Do you process biometric identifiers, such as face geometry or voiceprints?Biometric rules differ by state and country.What's processed, and how consent is captured.

2. Storage and residency

Residency means where candidate data actually sits: the main database, backups and logs. It matters because client contracts, works councils and some laws limit where recordings and transcripts can go. Things usually go wrong when a vendor names one region on a website but runs backups or support tooling somewhere else. A good answer lists regions by data type, names who can move data between them, and puts the region commitment into the contract instead of a brochure.

QuestionWhy it mattersA strong answer includes
Where is candidate data stored and processed, including backups and logs?Rules and client contracts may limit location.Regions listed by data type.
Can we choose a region, and is it fixed in the contract?A marketing page isn't a commitment.A contract clause, not a brochure line.

3. Retention and deletion

Retention covers how long video, audio, transcripts and scores stay on file, and who decides. Deletion usually fails at the edges, where a request clears the main record but leaves backups, subprocessor copies or AI-generated summaries behind. A good vendor gives defaults per data type, lets you change them, and commits to a deletion time limit with a named process and proof.

QuestionWhy it mattersA strong answer includes
What is the default retention for video, audio, transcripts and scores? Can we set it?Old recordings are a liability.Defaults per data type; customer setting.
Can you delete one candidate's data on request, including at subprocessors?Some laws require it.A time limit and named process.
Does deletion reach backups and AI summaries?Copies often survive.Backup deletion timing, and proof.

4. Access control

Access control is about who can open candidate recordings and scores, inside your team and inside the vendor. Interview data is sensitive, and a shared login or an over-broad role can expose a whole pipeline. What usually goes wrong is weak sign-in, roles that are too coarse for agencies with several clients, and vendor staff who can browse recordings unchecked. The CSA CAIQ treats identity and logging as separate control domains (read 18 Nov 2026). A good answer names the roles, shows an editable permission matrix, and offers exportable access logs.

QuestionWhy it mattersA strong answer includes
Do you support single sign-on, and can MFA be enforced for all accounts?Shared passwords are a common breach path.Supported methods; enforceable MFA.
What roles exist, and can we edit permissions?Roles need different views.A permission matrix.
Which vendor staff can see candidate recordings, and how is that approved?Insider access is a risk.Named roles, approval, logs.
Are access logs kept for candidate data, and can we export them?You need to see who opened what, and when.Log scope and retention, and an export you can request.

5. Encryption

Encryption protects video, audio and transcripts if a storage bucket or backup is exposed, and again while files travel between candidate, vendor and your systems. Problems appear when a vendor says "we encrypt everything" without saying where, or who holds the keys. The CSA CAIQ has a dedicated cryptography and key management domain (read 18 Nov 2026). A good answer covers data in transit and at rest, names who controls the keys, and avoids vague slogans.

QuestionWhy it mattersA strong answer includes
How is data encrypted in transit and at rest, and who holds the keys?It limits damage if storage is exposed.Both states, and key ownership.

6. Subprocessors and AI model providers

A subprocessor is any outside company that touches your candidates' data on the vendor's behalf: cloud hosting, speech-to-text, email, and the AI model providers that read transcripts. Under GDPR Article 28, the processor contract must cover subprocessors (read 18 Nov 2026). What usually goes wrong is a list that skips AI providers, or a promise not to train on your data that lives on a web page, not in the contract. A good vendor shares a current list, gives notice of changes, and lets you object.

QuestionWhy it mattersA strong answer includes
List every subprocessor touching candidate data, with function and location.Your risk includes theirs.A current list, with AI model providers.
Does any of our data train or improve your models or a third party's?Candidate data shouldn't become someone's training set.A contract ban, not a policy page.
Do you notify us before adding a subprocessor, and can we object?Surprises break your own commitments.A notice period and an objection right.
What covers transfers outside the EU or UK?Bad transfers fall on you too.The mechanism, with dates.

7. Certifications and reports

Certifications and audit reports are how a vendor proves its controls to someone who can't inspect the systems. They matter because they're independent evidence. They go wrong when a certificate covers a different product, an old period, or a narrow scope that leaves out the interview system. A good answer comes with the actual report or certificate, the period covered, the scope statement, any exceptions, and a straight "no" where the vendor doesn't hold something.

QuestionWhy it mattersA strong answer includes
Provide your SOC 2 report. Is it Type I or II, what period, and does the scope cover our product?Another system's report proves nothing.Dates, scope, exceptions, subservice organizations.
Provide your ISO/IEC 27001 certificate and scope statement.Scope can exclude the product you're buying.Certifying body, scope, expiry.
Do you hold ISO/IEC 42001 or run a documented AI management system?AI governance is separate from security.The scope, or an honest "no".

8. Incident response

Incident response is what a vendor does when something goes wrong: how it detects a problem, tells you, contains it and learns from it. What usually goes wrong is vague promises to notify "promptly" and no history to examine. The CSA CAIQ has an incident management domain (read 18 Nov 2026). A good vendor puts a notice window in the contract and describes past incidents honestly.

QuestionWhy it mattersA strong answer includes
How fast do you notify us after confirming an incident affecting candidate data?Your own notice duties may be short.Hours or days, in the contract.
Have you had a security incident affecting candidate or customer data in the last three years?Past handling predicts future handling.Scope, notice given and what changed.

9. Testing

Testing means independent people trying to break the vendor's systems before an attacker does, usually through penetration tests and routine scanning for known flaws. It matters because the product changes with every release. Watch for testing that was a one-off done years ago, or findings that are logged and never fixed. The CSA CAIQ includes application security and vulnerability management domains (read 18 Nov 2026). A good answer gives frequency, the date of the last test, scope, and how fast serious findings get fixed.

QuestionWhy it mattersA strong answer includes
How often do independent penetration tests run, and when was the last?Untested systems drift.Frequency, scope, date, fix timeline.

This area covers what candidates are told before an AI interview starts, and what they agree to. It matters because you're recording faces and voices, and consent and notice rules differ by place. What usually goes wrong is buried wording, no record of who agreed to what, and no alternative for a candidate who needs an accommodation or simply declines. A good vendor shows you the exact notice text, stores a time-stamped record of each consent, and offers a clearly logged route to a human process.

QuestionWhy it mattersA strong answer includes
What exact notice and consent do candidates see, and is it time-stamped?Consent rules vary (see recording consent).The text, and a stored record.
Is there an alternative path for candidates who can't or won't use the AI interview?Accommodations and opt-outs matter.A human route, logged.

11. AI-specific risks

AI-specific risk is what a normal security review misses: models that change, scores nobody can explain, and text that tries to steer the model. It matters because a quiet model update can change who ranks first. The NIST AI Risk Management Framework is a voluntary reference here (read 18 Nov 2026). Vendors often can't show evidence behind a score or test against hidden instructions in resumes. A good answer names its controls, versions its models and shows per-question reasoning.

QuestionWhy it mattersA strong answer includes
How do you tell us about model or scoring changes, and can we pin a version?A silent change can alter who's ranked first.Advance notice and version records.
How do you defend against prompt injection in resumes and typed answers, and how is it tested?Candidates can hide instructions in text.A named control and test method.
Can you show the evidence behind each score?You must explain decisions (explainable AI hiring).Per-question reasoning, transcript.

12. Continuity, offboarding and export

Continuity is whether the tool keeps working when something fails, and offboarding is how you leave. If you can't export your records, you're locked in with data you can't move or delete. The CSA CAIQ covers business continuity as its own domain (read 18 Nov 2026). Gaps usually show up as untested recovery plans and vague exit terms. A good answer gives recovery targets, a recent test date, export formats and a deletion date.

QuestionWhy it mattersA strong answer includes
What are your recovery time and recovery point targets, and when did you last test recovery?Downtime stalls hiring.Targets, test date, backup site.
At contract end, can we export all records in machine-readable form, and when is our data deleted?Lock-in is a security issue too.Export formats, deletion date.

Red flags in vendor answers

  • "We're compliant" with nothing attached.
  • A certificate with no scope or dates.
  • A SOC 2 with an old period or the wrong scope.
  • "We don't train on your data" absent from the contract.
  • A subprocessor list missing AI model providers.
  • Deletion with no time limit.
  • Vague answers on prompt injection.

Documents to ask for

  • Current SOC 2 report (Type II preferred), under NDA
  • ISO/IEC 27001 certificate with scope statement
  • Completed SIG Lite or CAIQ
  • Data processing agreement, with subprocessor list
  • Retention and deletion policy
  • Incident response summary and notice terms
  • Penetration test summary
  • Candidate notice and consent text

How ZeaHire answers these

ZeaHire is our product. Ask us for ZeaHire's answers to this list in your demo.

  • Access is role-based, with an editable permission matrix, and Sign in with Microsoft is available.
  • Candidates verify with a code sent to their invite email.
  • ZeaHire connects with nearly a dozen ATS and HR systems, including Workday, SAP SuccessFactors, Greenhouse, JazzHR and JobDiva. If yours isn't on the list, we'll build the connection. Ask to see it with your own system in a demo.

FAQ

What is a vendor security questionnaire?

Questions a buyer sends a supplier to assess how it protects data, with evidence attached. Often built from SIG or CAIQ.

What is the difference between SOC 2 Type I and Type II?

Type I is commonly described as a report on controls at one date; Type II covers a period, often 3 to 12 months. Ask for Type II where you can, and check that the scope covers your product.

Do AI interview vendors train models on candidate data?

Ask, and get the answer in the contract or data processing agreement, not a web page. Ask whether their AI model providers can use your data too.

What should I ask about a SOC 2 report?

Check the period, Type I or II, and whether the system description covers your product. Read the exceptions.

Which security framework should I use?

Use your security team's. Without one, SIG Lite or CAIQ is a common start, and NIST AI RMF helps for AI risk. See also enterprise HR data security and the AI note taker and GDPR.

How long should a vendor security review take?

It varies. Vendors with a completed standard questionnaire and a current report answer faster, so ask for those first.

Book a demo

Book a demo.

AI interviewsBuying guides
Was this useful?

Comments

No comments yet — start the conversation.

Leave a comment

Your email is never published. Comments appear once a moderator approves them.

Replying to
Plain text only. Be kind and stay on topic.

By commenting you agree to our privacy policy.