Most hiring teams shortlist an AI interview platform on three things: supported languages, price per interview and how good the demo looks. All three matter, but none of them says whether the platform’s scores are trustworthy.
That is the buying decision. An AI interviewer that cannot show why a candidate got 7 out of 10 hands your hiring manager a number to argue about and a rejected candidate nothing to dispute. This guide gives you eight questions for every vendor, a weighted scorecard to compare them and a two week pilot to test the answers before you sign.
It is written for hiring teams in India, so it covers mobile first candidates, regional languages and the Digital Personal Data Protection (DPDP) Act, and flags international rules where they apply. Sources were checked in October 2026.
What counts as an AI interview platform
Three different products are sold under this label. TheHireHub’s 2026 guide separates them, and the split saves time:
- AI screening tools rank or filter applicants from resumes or short assessments.
- Interview intelligence tools record, transcribe and summarise interviews that humans conduct.
- AI interviewer agents run the interview themselves, by text, voice or video.
This guide is mainly about the third category, but the eight questions apply to the other two wherever software scores people. Pick the category that matches your bottleneck before you compare vendors:
| Your bottleneck | Category to shortlist | First thing to test |
|---|---|---|
| Hundreds of applicants, few qualified | AI screening | Wrongly rejected applicants: re read a sample of rejections |
| Engineers lose hours on first round calls | AI interviewer | Whether each score traces back to the candidate’s answers |
| Interviewers’ notes are thin or inconsistent | Interview intelligence | Transcript accuracy on Indian accented English and code mixed speech |
| Candidates drop out before the interview | Any, with a candidate experience check | Completion rate on a mid range phone over 4G |
Eight questions to ask every vendor
1. Can you trace every score to what the candidate actually said?
Ask for a finished scorecard from a real interview, anonymised, and click from one finding to its evidence. In a strong product, each dimension score links to specific candidate answers or actions, such as a verbatim quote or a line of submitted code, and a hiring manager can verify the finding in a minute. In a weak one you get a number, a confidence percentage and a paragraph of summary.
A model written summary is itself model output, so the only way to know it is accurate is to check it against the transcript. If the vendor makes that check hard, your team will either trust the score blindly or ignore it.
Also ask:
- Can I export the full transcript and scorecard?
- What does “confidence” measure, and how was it validated?
- When the system is unsure, does it flag the uncertainty or guess?
2. What does it score, and from which signals?
Separate content (what the candidate said, how they reasoned, what their code did) from delivery and appearance (facial expression, tone, “confidence”). Ask the vendor to list every input the model sees and mark which are scored and which are only recorded.
HireVue, a large US video interview vendor, dropped facial analysis from its assessments. Fortune reported in January 2021 that the company’s chief data scientist said nonverbal data added about 0.25% to a model’s predictive power in most cases and about 4% for customer facing roles, with the content of answers the strongest predictor. That is the vendor describing its own models, but it matches most buyer guidance: score what candidates say and do, not how they look. TheHireHub’s guide calls facial and tone analysis “contested” for the same reason.
Language claims need the same scrutiny, because “50+ languages” says little. Ask for a demo interview in the language your candidates actually use, and have a native speaker on your team judge whether the questions and the scoring make sense.
3. Who makes the decision?
Check that the AI output is a recommendation, that a named person confirms it, and that every override is logged. Be wary of any product that rejects candidates at the interview stage with no human looking.
India has no standalone AI law. MeitY’s AI Governance Guidelines, issued in November 2025, are non-binding and set out principles that include putting people first, accountability for developers and deployers, and systems that are understandable by design. In the EU, AI used in employment is classed as high risk, and the date those obligations apply moved from 2 August 2026 to 2 December 2027 once the Digital Omnibus on AI came into force on 27 July 2026. That matters if you hire EU residents or your vendor sells into Europe. Human oversight is good practice either way.
4. What evidence shows it works for roles like yours?
A vendor case study is not validation. The most credible public evidence on AI led interviews is a field experiment by Brian Jabarian and Luca Henkel, “Voice AI in Firms” (SSRN paper 5395709). It randomised about 67,000 applications for entry level customer service jobs in the Philippines. Applicants were interviewed by a human recruiter, by an AI voice agent, or by whichever of the two they chose, and human recruiters made every hiring decision. Compared with human led interviews, AI led interviews were associated with 12% more job offers, 18% more job starts and 17% higher one month retention, all in relative terms.
Read the limits before you quote it. It covers one recruitment firm in one country, in high turnover call centre roles, with retention measured at one month. It says nothing about Indian engineering hiring or about any particular product. Treat it as evidence that well run AI led interviews can work, not that a given tool does.
Ask vendors for these instead:
- Validation on your role family, with the method and sample size.
- Adverse impact checks across gender, language background and region.
- Customers hiring for similar roles whom you can call.
5. How does it behave for your candidates, on their devices?
Test the candidate flow yourself on a mid range Android phone over 4G, as TheHireHub suggests. Check whether candidates must install an app or create an account, how microphone permission is requested, what happens if the connection drops mid interview, and how long the whole thing takes.
The Philippines experiment shows that even a successful rollout has friction. About 5% of applicants ended the AI interview because they did not want to speak to an AI, and the voice agent had technical difficulties in 7% of cases. Budget for a human interview fallback, and ask each vendor for completion and drop off rates on a comparable candidate population rather than an average across all customers.
Tell candidates up front that an AI is conducting the interview, what it will assess, and that a person reviews the result.
6. How does it handle candidate data under the DPDP Act?
Interview recordings, transcripts and video are personal data. The DPDP Rules were notified on 13 November 2025, and under the phased schedule the main duties for data fiduciaries (notice, consent, security safeguards, breach reporting and data principal rights) start around 13 May 2027. MeitY floated cutting that period to 12 months in January 2026. A September 2026 timeline review found no amending notification, but check the Gazette for the current position before you rely on a date.
Put these questions to the vendor and get the answers into the contract:
- Where is candidate data hosted, and which sub processors, including AI model providers, receive it?
- Is personal information stripped or minimised before it reaches a model?
- How long are recordings and transcripts kept, and is deletion automatic?
- How can a candidate request their data or its deletion, and how fast do you respond?
- Is candidate data ever used to train models?
- What is the breach notification process?
This is a buyer’s checklist, not legal advice. Involve your counsel before you finalise data processing terms.
7. Does it fit the hiring process you already run?
A scorecard that lives in a separate dashboard gets forgotten. Find out where results land: in your applicant tracking system beside the résumé and assessment results, or in a tool your hiring managers have to remember to open. Check scheduling, calendar and assessment links, and ask for a clean export in case you leave.
Every extra tool is another place candidate data sits and another data processing agreement to manage. If a vendor lists an integration as available, ask to see it working with your stack, not on a slide.
8. What will it cost at your volume?
Pricing models vary: per interview, per assessment, per seat or a flat plan, as TheHireHub notes. Compare total monthly cost at the number of interviews you expect to run, not the headline unit price. Ask what counts as a billable interview (abandoned, rescheduled and retaken sessions), which volume tiers and overages apply, whether voice, regional languages and recording storage cost extra, and what the minimum commitment is.
Then compare the total with what the interview replaces. If an engineer spends an hour on every first round call today, count that time too.
Compare vendors with a weighted scorecard
Score each vendor from 0 to 5 on every criterion, multiply each score by its weight divided by 5, and add the results. The maximum is 100.
| Criterion | Weight | A 5 looks like | Evidence to request |
|---|---|---|---|
| Traceable scoring | 20 | Every finding links to candidate quotes or actions | Live click through of a real scorecard |
| Human decision and override | 15 | Recommendation only, named approver, override logged | Demo of the approval and override flow |
| Role relevant validation | 15 | Method and results for your role family | Study, sample size, reference customers |
| Candidate experience | 15 | High completion on mid range phones, clear AI disclosure | Completion data; your own test on a real phone |
| Data protection (DPDP) | 15 | Clear hosting, retention, deletion and sub processor terms | Draft data processing agreement |
| Process fit | 10 | Scorecards land in your ATS; clean export | Working demo with your stack |
| Cost at your volume | 10 | Predictable total at your expected interviews | Written quote with billing definitions |
A vendor scoring 4 on traceable scoring earns 4 ÷ 5 × 20 = 16 points.
The weights are a starting point for engineering led hiring. Change them to suit your context, but fix them before you see any demo so a polished presentation cannot move them. Treat the first two criteria as gates: a vendor scoring below 3 on traceable scoring or on human decision should be dropped whatever its total.
Comparing vendors right now? Book a 15 minute demo and bring this list. We will walk through a real scorecard against it.
Run a two week pilot before you sign
A demo shows the best case. A pilot shows yours.
- Set pass criteria first. Write down the agreement rate, completion rate and failure rate you will accept before the pilot starts.
- Run it alongside your normal first round for one open role. Humans make every decision during the pilot.
- Review blind. Have two of your own interviewers read a sample of transcripts and scorecards without seeing the AI recommendation, then compare their view with it.
- Re read the rejections. Pull a sample of candidates the AI marked Hold or Reject and check them by hand. Wrongly rejected candidates are the costliest errors, and nobody notices them unless someone looks.
- Measure the candidate’s side. Track completion, drop off, technical failures and time to scorecard, and ask each candidate two questions about the experience.
- Compare groups. Where you can lawfully collect the information, look for large gaps in outcomes by language, region or gender.
- Score the vendor on the table above, using what you saw rather than what you were told.
Red flags
- “Bias free” or “removes bias” with no description of how it was audited. Bias can enter through training data, prompts and your own rubric.
- Scores built on facial expression, tone or “confidence”.
- One overall number with no breakdown and no evidence.
- Automatic rejection at the interview stage with no human step.
- No transcript export and no deletion process.
- Language claims with no sample interview in the language you need.
- Case studies with no method, sample or numbers you can check.
Where Kestrel fits
Kestrel is an AI-native hiring platform for engineering teams at startups and scale-ups. Its AI interviews are adaptive, question based conversations in text, with an optional voice mode that recruiters can switch on per job where it is set up. A separate Smart Canvas format covers system design. Each interview produces a scorecard with a recommended outcome (Advance, Hold or Reject), and findings cite the candidate’s own words with a timestamp that links to that point in the transcript. The full transcript downloads as a PDF.
The AI recommends and a person decides. After an interview, the candidate moves to the next human review stage, where a recruiter or hiring manager makes the hire or reject call. The AI interview never rejects a candidate or makes an offer on its own. Candidates join from a link with no account.
Recording is configurable per workspace. Where it is on, the webcam and screen are recorded with the candidate’s consent. Video does not feed the interview score, and any integrity flags are reviewed by a person.
The interview sits in the same pipeline as résumé screening, MCQ and coding rounds and the applicant tracking system, so the scorecard sits beside the rest of a candidate’s record. Kestrel is designed for engineering-led hiring. If your main need is very high-volume non-technical hiring, check the fit by booking a demo.
Frequently asked questions
Are AI interviews legal in India?
India has no standalone AI law, and the AI governance guidelines issued in November 2025 are non binding. In practice your main obligations come from the DPDP Act and Rules for candidate data and from your general duty to treat candidates fairly. This is not legal advice, so ask your counsel to review your use case.
Can an AI interviewer replace the first human round?
It can take over a structured first screen, and the Philippines field experiment found better outcomes under its conditions with humans making every decision. Keep a person accountable for each hire or no hire call, and keep a human option for candidates who do not want an AI interview.
Will candidates accept an AI interview?
In the same experiment, 78% of applicants offered a choice picked the AI interviewer. That was one population of entry level applicants, so test it with yours. Disclosure, a clear explanation of how results are used and a working fallback all make acceptance more likely.