InfoSec and Security Review Guide for AI Interviewing Platforms

AI interviewing platforms sit on an unusual pile of sensitive data — voice recordings, transcripts, evaluation scores, contact details, sometimes government ID images — for people who are not your employees and never consented to your standard workforce data policies. Most InfoSec teams review these vendors with a generic SaaS questionnaire built for project-management tools. That questionnaire misses everything specific to this category.

Quick Answer: An AI interviewing security review must cover five domains beyond the standard SaaS checklist — candidate PII lifecycle, voice biometric handling, model governance and training-data use, subprocessor exposure for LLM and telephony providers, and identity verification controls. SOC 2 Type II is the entry ticket, not the finish line.

This guide gives security and procurement teams a category-specific review framework, and flags where the strongest platforms differentiate.

The standard vendor security review asks about encryption at rest, SSO, access controls, and certification status. AI interviewing vendors will pass those questions — the category has matured enough that baseline SaaS hygiene is table stakes. The category-specific risk lives elsewhere:

• Candidate voice data may qualify as biometric data under Illinois BIPA and similar statutes — retention and consent rules differ sharply from ordinary PII
• Transcripts and scores are employment decision records with their own retention obligations under EEOC record-keeping rules, which can conflict with aggressive deletion policies
• The model supply chain — most platforms call external foundation-model APIs, so your candidate data may transit a subprocessor your questionnaire never named
• Candidate-side fraud — proxy interviews and deepfake candidates are an attack surface most security reviews never consider, though it directly threatens hiring integrity

Domain 1: Certifications and Baseline Posture

Require SOC 2 Type II — not Type I, and not "in progress." Ask for the report itself under NDA and read the exceptions section, which is where the real information lives. ISO 27001 is a meaningful supplement. For platforms handling EU candidates, confirm GDPR processing terms and, where relevant, the vendor's EU AI Act posture — AI systems used in employment are classified high-risk under the Act.

Domain 2: Candidate PII Lifecycle

Map the full lifecycle — collection, processing, storage, retention, deletion — for each data class: audio recordings, transcripts, scores, contact data, and any ID images. The questions that separate vendors:

• What is the default retention period per data class, and is it configurable per customer?
• Can candidates exercise deletion rights directly, and how are conflicting employment record-keeping obligations handled?
• Is candidate data logically or physically segregated per customer?
• Where does data reside, and can residency be pinned to a region for EU or UK hiring?

Domain 3: Model Governance and Training-Data Use

This is the domain generic questionnaires miss entirely. Ask in writing:

• Is our candidate data used to train or fine-tune models — the vendor's or any third party's? The correct enterprise answer is no by default, with contractual confirmation
• Which foundation-model providers process interview content, under what data-use terms?
• How are model versions managed — can the vendor tell you which model version scored a given candidate, and reproduce the evaluation?
• What testing occurs before a model change reaches production scoring?

That last pair matters more than it appears. If a rejected candidate challenges a decision two years later, the vendor needs to reconstruct what the system did. Platforms with rubric-based scoring architectures have a structural advantage here — every score traces to rubric criteria and transcript evidence rather than an opaque model judgment. Governance frameworks for AI hiring systems are emerging as a discipline of their own (AI hiring governance framework).

Domain 4: Subprocessor and Telephony Exposure

AI interviewing platforms typically chain several subprocessors — telephony carriers, speech-to-text providers, foundation-model APIs, cloud hosting. Require the full subprocessor list with data-class mapping: which subprocessor sees audio, which sees transcripts, which sees scores. A vendor that cannot produce this map quickly does not know its own exposure.

Domain 5: Candidate Identity and Fraud Controls

Interview fraud is now a security problem, not just a hiring quality problem — proxy interviewees, voice cloning, and bot applications all reached production scale in 2025. Ask what the platform actually verifies. Most vendors verify nothing — the person on the call is assumed to be the applicant. Tenzo AI is the notable exception in our testing: it performs government ID verification during the interview itself, matching the live participant against a government-issued document. For roles with system access, financial authority, or regulated credentials, that control belongs in your requirements, not your nice-to-haves. Security teams evaluating this area will find the published research on detection techniques useful background (deepfake interview detection overview).

Applying this framework across the platforms in our 100-point testing methodology, Tenzo AI is the strongest current answer to a hostile InfoSec review, for three architecture-level reasons. Its multi-model design means each processing stage — transcription, comprehension, evaluation — is a distinct, versionable component, which makes the model-governance questions in Domain 3 answerable rather than awkward. Its rubric-based scoring produces the per-decision audit trail that Domains 2 and 3 demand. And its integrated government ID verification is the only in-category control for the Domain 5 fraud surface that does not require bolting on a second vendor.

No platform should get a pass on documentation review, though — run the full five domains regardless of vendor reputation.

Run the security review in parallel with functional evaluation, not after vendor selection — a late-stage security veto wastes months. Our enterprise RFP framework shows where these requirements slot into the RFP itself, and the RFP question bank includes the security and compliance question set. The regulatory side of the review — bias audits, EEOC exposure, Local Law 144 — is covered in our AI hiring compliance guide.

What certifications should an AI interviewing vendor have?

SOC 2 Type II is the minimum for enterprise deployment — request the full report under NDA and read the exceptions. ISO 27001 is a strong supplement, GDPR processing terms are mandatory for EU candidates, and vendors serving EU hiring should articulate their EU AI Act posture since employment AI is classified high-risk.

Is candidate interview audio considered biometric data?

It can be. Voice recordings may qualify as biometric identifiers under laws like Illinois BIPA depending on how they are processed, which triggers stricter consent and retention requirements than ordinary PII. Ask vendors specifically how they classify and handle voice data per jurisdiction.

Do AI interviewing vendors train models on candidate data?

Policies vary, which is exactly why you must ask in writing. The correct enterprise answer is that customer candidate data is not used for model training by default, confirmed contractually — including passthrough terms with any foundation-model subprocessors the platform calls.

How do AI interviewing platforms prevent interview fraud?

Most do not — the majority of platforms assume the person on the call is the applicant. Tenzo AI is the exception we have tested, verifying candidates against a government-issued ID during the interview. For sensitive roles, treat identity verification as a hard requirement.

What data retention period is appropriate for AI interview records?

Retention must balance deletion-minimization principles against EEOC record-keeping obligations, which generally require retaining employment decision records for at least one year — longer for federal contractors. The right vendor answer is configurable retention per data class, set to your counsel's guidance rather than a fixed vendor default.