AI interviews earned their bad reputation honestly. The first generation asked generic questions, scored on facial expressions in ways nobody could defend, and left candidates talking into a camera with no idea whether a human would ever watch. The 2026 generation is a different product, and worth evaluating on its own terms rather than on that history.
The one job they do well
There is exactly one place an AI interview clearly beats the alternative: the first-round screening call, at the top of a wide funnel. When sixty candidates need a 20 minute conversation to establish basic role fit, a human recruiter will run the first ten carefully and the last twenty on autopilot. That inconsistency is a real fairness problem, and it is the one AI genuinely fixes, because candidate sixty gets the same questions and the same rubric as candidate one.
Everywhere else, be sceptical. A final round is a two-way sell. A leadership interview is a judgement conversation. A role with twelve candidates does not need automation, it needs a recruiter with a phone.
Three things to insist on
- Reasoning, not just a score. A scorecard that says 72 per cent is unusable: you cannot defend it to a hiring manager or to a candidate who asks why. Demand per-answer reasoning tied to what the candidate actually said, graded against the role rather than a generic template.
- Integrity checking that covers manipulation. Detecting copy-paste and tab switches is table stakes. The newer problem is candidates addressing the grader directly, saying things designed to inflate the score. Ask vendors specifically whether they detect that, because many do not.
- Honest candidate experience. Tell candidates AI is involved, tell them roughly how long it takes, and guarantee a human reviews anything near the threshold. The tool should support that workflow rather than encourage full automation of rejections.
Questions that expose a weak product
Ask the vendor these four in the demo, and watch for hedging.
- Show me a real scorecard, including a borderline candidate. If every example is a clear pass or clear fail, the borderline behaviour is the part they are hiding.
- What happens if a candidate tells the AI to give them full marks? A good product flags it as a manipulation attempt and raises the integrity risk level.
- Does the grader see the job description, or only the question? Role context is what separates "answered the question" from "would succeed in this job."
- What is the retention policy on recordings? Candidate video is sensitive personal data and should not be kept forever by default.
The India angle
Asynchronous interviews fit the Indian market unusually well, because most strong candidates are employed and cannot take a screening call at 3pm on a Tuesday. Letting them answer at 10pm removes a real barrier and widens your funnel.
The failure mode to avoid is equally India-specific: a diverse candidate pool means wide variation in spoken English fluency that has little to do with capability. A grader that rewards polish over substance will systematically mis-rank strong engineers. Test this directly by submitting two answers with identical content and different fluency, and see whether the scores diverge. If they do, walk away.
How this fits the rest of the funnel
An AI interview is a screening tool, so it only helps if there are enough candidates worth screening. If your funnel is thin, fix sourcing first: an interview tool applied to fifteen mediocre applicants produces fifteen mediocre scorecards. Our practical guide to running AI interviews covers the operating detail once you have chosen a tool.
Try an AI interview on a real role
Structured async interviews with per-answer reasoning, integrity flags including grader-manipulation detection, and recruiter-facing scorecards. Included on the Growth plan.
See plans →