AI interview software conducts a structured, competency-based interview with every applicant and returns a scored, ranked shortlist to the recruiter. It is not the same as CV screening, which filters on what a candidate has already done; a structured interview surfaces what they can actually do. It is also narrower than the broad AI recruiting platforms category: the job is the screening interview itself, not sourcing or scheduling.

The distinction that matters most sits inside the AI. The strongest AI tools use two separate layers: a conversational layer that runs the interview in natural language so candidates feel heard, and an assessment layer that scores the answers. How that second layer works; whether it is deterministic or a probabilistic large language model (LLM), is the single biggest factor in whether the result is fair and defensible. Hold that thought as we run through the criterion below.

Use these as your evaluation checklist. Each one is a question to put to the vendor, why it matters, and what a good answer looks like.

1. Bias mitigation: does every candidate get the same interview?

The most reliable way to reduce bias is structural, not aspirational. Ask whether every applicant receives the same competency-based interview, assessed against the same criteria, regardless of background or CV polish. Then ask the harder question: does the vendor run bias checks across protected groups during model development, and monitor for adverse impact after deployment? "Fair" is now industry-common language; the proof is whether the process is genuinely identical for everyone and audited to confirm it. A fair process is also a better process: a tool that filters on demographic signals instead of competency is prioritizing irrelevant data and handing you a weaker shortlist.

2. Explainability: is the explanation the real scoring logic?

Any vendor will say their tool is explainable. But press for the specifics. Can the software show why a given candidate scored the way they did, tied to their actual responses? And critically: is that explanation the genuine scoring logic, or a plausible-sounding narrative generated after the score was assigned? Large language models (LLMs) used for scoring tend to do the latter; they produce a reasonable explanation after the fact that may not reflect how the score was reached. EU AI Act Article 13 requires high-risk systems to be transparent enough for a human to interpret their output. You cannot stand behind a decision you cannot reconstruct.

3. Consistency: same answer, same score, every time?

Ask the vendor to confirm that identical answers always produce an identical score. This sounds obvious; it is not guaranteed. Research has documented that large language models can rank the same candidate inputs differently across runs (Redstone, 2025; Seshadri et al., 2025; Varshney & Ganuthula, 2025). A candidate's career should not depend on what time of day the model was queried. Deterministic assessment models, where the same input always yields the same output and models are versioned and locked for the duration of a hiring cycle, remove that variance by design. Every candidate in a cohort is then assessed under identical conditions.

4. Candidate experience: will people actually finish?

High-volume screening only works if candidates complete the interview, and the experience reflects on your employer brand whether you intend it to or not. Ask for real completion rates, not satisfaction scores in isolation. Ask how many languages the interview supports, whether it works on mobile, and how quickly a candidate can start. A warm, conversational experience is what drives completion; a cold, form-like one drives drop-off and damages your brand at scale.

5. ATS integration: scored shortlists where recruiters already work

Screening output is only useful if it lands inside your existing workflow. Ask how deep the ATS integration goes: does the tool push scored, ranked, auditable shortlists directly into the candidate record, or does it bolt on as a separate system recruiters have to check? Confirm the specific ATS platforms supported and what data flows back. An integration that forces recruiters to live in a second tool will quietly fail to be adopted.

6. Legal defensibility and human oversight: who is accountable?

The EU AI Act requires effective human oversight for high-risk systems, and there is a deeper reason behind the rule: only a human can be morally and legally accountable for a hiring decision. Confirm that the software never auto-rejects candidates, that the final hire decision always rests with a recruiter, and that there is a full audit trail of scoring, overrides, and recruiter actions. Legally defensible means a decision can be explained from first principles in an audit or a tribunal.

7. Measurable screening efficiency: prove the time and cost savings

This is the criterion most buyers lead with, and it is fine to, as long as you demand evidence. Ask for named-customer outcomes on time-to-hire, screening time reduction, and cost per hire, not vendor averages with no source. Efficiency without the six criteria above is a high-speed way to hire the wrong people; efficiency with them is the whole point.

8. Data security and sovereignty: where does candidate data live?

Recruitment data is sensitive, and candidates are increasingly wary of how it is used. Confirm GDPR compliance, where candidate data is processed and stored, and whether candidate data is ever used to train third-party models. Data minimization and EU data residency are reasonable baseline expectations for enterprise hiring.

• The vendor cannot tell you whether scoring is deterministic or probabilistic. If they cannot answer, assume probabilistic, and assume variance.
• "Explainability" turns out to mean a generated summary rather than the actual scoring logic.
• The tool can auto-reject candidates without human review.
• Completion rates and candidate satisfaction are quoted with no source or named deployment.
• The "integration" is a CSV export rather than scored shortlists inside your ATS.
• Compliance is described as a feature that was added later, rather than a design principle.

Hubert was built around exactly these criteria, years before the generative AI wave, on the science of structured interviewing (Schmidt & Hunter, 1998). Every candidate completes the same structured, competency-based interview in any of 30+ languages, assessed by deterministic AI models: same input, same output, full explainability. The conversation feels human; the scoring is auditable. You do not have to choose between the two.

The result for recruiters is a scored, auditable shortlist delivered directly into the ATS across 30+ integrations, with the final decision always staying with your team. Hubert predicts hiring success with 5x greater accuracy than traditional methods, delivers up to 80% faster time-to-hire, and earns a 9/10 average candidate experience score.

This combination, a warm conversational interview paired with a deterministic, faithfully explainable assessment layer, is what makes the shortlist legally defensible by design rather than retrofitted with compliance language.

What is the difference between AI interview software and AI recruiting platforms? AI recruiting platforms is a broad category covering sourcing, matching, scheduling, and screening. AI interview software is the specific part that replaces the screening interview: it conducts a structured, competency-based interview with every applicant and returns a scored shortlist.

Does AI interview software reduce bias in hiring? It can, but only structurally. The mechanism is giving every candidate the same structured interview, scoring it consistently, and auditing for adverse impact across protected groups. Tools that score with probabilistic models, where the same answer can produce different results, undercut that consistency.

Is AI interview software allowed under the EU AI Act? AI used in recruitment is classified as high-risk under the EU AI Act. That does not make it prohibited; it means the software must meet requirements for transparency, accuracy, and human oversight, and the employer must retain accountability for decisions. Choose software built on those principles rather than retrofitted to them.

What is the most important thing to evaluate? How the assessment layer scores answers. Deterministic scoring, where the same input always produces the same output, is what makes results consistent, explainable, and defensible. Ask the question directly; if a vendor cannot answer it, that is your answer.

If fair, fast, and defensible screening at high volume is the brief, see how structured AI interviews work end to end with Hubert. Book a demo and we will run your real roles through it.