What 66,000 Resumes Taught Us About Hiring in India

I have spent the last six months watching a number tick up on a dashboard: every resume our system scored, every interview our AI conducted. This week it crossed a threshold that made me stop and pull the full dataset: 66,873 resumes screened, 8,096 voice interviews created, 475 hours of AI-conducted conversation. All of it real hiring, mostly high-volume campus and lateral drives in India, between March and August 2026.

Numbers at this scale start telling you things that single campaigns cannot. Some of what they told us confirmed what every recruiter already suspects. Some of it surprised us enough to change our product. All of it is below, as measured, including the numbers that are not flattering to us. Everything is aggregated across customers; no company or candidate is identifiable.

Half of every applicant pool was never a fit

Of 33,753 resumes that went through relevance scoring, 49.5 percent scored below 30 percent relevance to the role they applied for. Not “weak fit”. Below thirty percent: facility managers applying to customer-success roles, mechanical engineers applying to frontend positions, fresh graduates applying to roles demanding five years of ownership.

At the other end, exactly 1.7 percent scored above 70 percent.

That distribution is the real shape of the application flood. When a recruiter says “we got 3,000 applications”, the honest translation is “we got roughly 1,500 applications that a relevance check could have removed before a human ever looked”. The volume problem in Indian hiring is mostly a relevance problem, and relevance is checkable by machine, with evidence, in seconds. The tragedy of manual screening is not that it is slow. It is that tired human eyes spend half their attention on resumes that arithmetic could have retired.

Fewer than 3 in 100 candidates are a Strong Go

Our screening verdicts land in four calibrated tiers. Across 1,414 fully screened candidates: 54.2 percent No Go, 23.0 percent On Hold, 15.0 percent Go, and 2.7 percent Strong Go.

Fewer than three candidates in a hundred clear the bar that means “move fast on this person”. We recalibrated our scoring three times this year, always in the same direction: recruiters kept telling us the top tier was too generous. The uncomfortable truth of high-volume hiring is that a shortlist worth a hiring manager’s time is a tiny fraction of the pool, and any screening system that says otherwise is flattering its owner.

The useful consequence: when the top of the funnel is that thin, evidence matters more than ranking. A recruiter can disagree with a verdict in seconds when the exact resume line and interview quote sit next to the score. They cannot disagree with a bare percentage.

The most popular interview hour in India is 9 PM

This is the number I quote at dinner. When candidates schedule their own interviews, 39.4 percent choose slots outside 9-to-6, and the single most popular hour across 1,705 booked slots is 9 PM. 8 PM is third.

Think about what that means against the traditional recruiter-coordinated phone screen, which by definition happens during the recruiter’s working hours. Employed candidates, the ones you most want to poach, cannot talk at 3 PM on a Wednesday. Students are in class. For years the industry has effectively insisted that the best candidates take career risks during office hours or drop out of the funnel. Given the choice, four in ten candidates interview at night, on their own terms, and the completion behaviour suggests they are more relaxed doing it.

Self-scheduling is usually pitched as a recruiter convenience. The data says it is a candidate-access feature.

AI cheating is real, rarer than the panic, and visible

Everyone asks about candidates using ChatGPT in interviews. We measure it. Every completed interview runs through a deterministic composite of seven signals: answer latency, tab switches, the absence of self-correction in speech, and a fingerprint we call slow-but-fluent, where a candidate takes fifteen-plus seconds to start answering and then delivers unnaturally polished prose. Humans thinking out loud correct themselves; people reading a generated answer do not.

Across 2,582 assessed interviews: 97.4 percent showed low signals. 1.9 percent flagged elevated or high. Only 2.6 percent of candidates switched tabs at all during a completed interview.

Two readings of that. First, the panic is overblown: at least in live voice interviews, the overwhelming majority of candidates just answer the questions. Second, the format matters enormously. A live conversation that probes follow-ups in real time is hard to outsource to a chatbot at speaking pace. One-way recorded video, where a candidate can retry takes with a script open, is a different risk category entirely. If integrity is the concern, the answer is not more surveillance; it is a format where assistance is structurally visible. And the flags are exactly that, flags for human review, never auto-rejections.

One in five booked interviews never starts

Even with self-chosen slots, calendar invites and a reminder ladder, 21 percent of booked interviews expire without the candidate ever joining. Before the reminder ladder it was worse. No-shows are the quiet tax on every hiring funnel, and no amount of AI removes them; what changed for our customers is the cost per no-show. When an interview slot is an AI agent rather than a recruiter’s blocked half hour, a no-show costs nothing but a row in a dashboard, and the candidate gets an automatic second chance instead of a recruiter’s silent grudge.

Candidates rate the AI interviewer 3.64 out of 5

We ask every candidate to rate the experience after the call. Across 1,443 ratings the average is 3.64 out of 5. I am publishing that number precisely because it is not 4.8. Nobody’s honest measurement of a new interview format is 4.8.

What the accompanying comments say: candidates like judgment-free questioning, scheduling on their own terms, and knowing evaluation criteria are consistent. What they dislike tracks our bug tracker with embarrassing precision: the agent occasionally talking over slow starters, pronunciation of some Indian names, and wanting more feedback about how they did. That last one led us to ship candidate-facing scorecards. The 3.64 is a floor we are climbing, measured, not marketed.

What this means if you hire at volume

Six months of data compresses to four sentences. Half your applicant pool can be retired by relevance checking before a human looks, and your best people are hiding in the 2 or 3 percent the data says deserve speed. Your candidates want to interview at 9 PM, and the ones who cannot interview during office hours might be the ones worth waiting up for. AI-assisted cheating is real but rare, and live conversation is the format that keeps it visible. Measure candidate experience honestly, because a 3.64 you can explain beats a 4.8 you cannot.

We will publish this dataset’s next cut when it crosses 100,000 resumes. If you want to see the screening layer behind these numbers, the fastest way is to experience it: our AI interviewer will screen you, live, in five minutes, and email you the same scorecard a recruiter would see. Take the live demo.

Methodology: aggregates across HireQwik customer hiring pipelines in India, March 9 to August 27, 2026. Resume relevance n=33,753; verdicts n=1,414; booking hours n=1,705; integrity composite n=2,582; candidate ratings n=1,443. No customer or candidate is identifiable in any figure.