An AI recruitment assistant that makes the first screening call and hands the recruiter what the candidate actually said — quotes, not a score. Problem framing, prioritization, an AI feature design with guardrails, and a prototype you can use.
And only about four of those hours go to the part that actually needs a human.
Recruiters spend most of the week on repetitive work — screening resumes, making first screening calls, coordinating interviews, chasing status. That caps how many roles one recruiter can carry. Roles stay open around 45 days, cost per hire climbs, and most applicants never hear anything at all.
Who's affected: recruiters and HR as primary users; candidates and panel interviewers as secondary. The buyer is the Head of TA, with CHRO, IT & Security, Legal, Finance and the incumbent ATS vendor all in the room.
A sellable enterprise product, not an internal tool. A mid-to-large in-house TA team — roughly 1,200 hires a year across 30 recruiters. The customer already runs an ATS, so we connect by API rather than replace it. Hiring in India only. And they'll connect Email, Calendar and the WhatsApp Business API.
Shadow three recruiters for a day and count how many calls ended in "not a fit" inside five minutes. Then pull the ATS and check what share of applicants ever reached a final status. Both are cheap, and both can kill the premise.
Two things done properly beats six done badly.
Vision: every candidate judged on the same evidence, and every decision made in days.
Goal, 12 months: cut application-to-interview from 14 days to under 48 hours, give every applicant a decision, and hold shortlist quality flat.
| The problem | Cost | What we build |
|---|---|---|
| Interview scheduling | 17 hrs/wk | Self-scheduling link — the candidate books their own slot from real panel availability |
| First screening calls | 11 hrs/wk | AI voice agent — makes the first call, hands back what the candidate actually said |
The call that shaped the product: voice, not a text chat.
Voice costs roughly 15× more per screening than text, and speech-to-text is measurably less accurate on regional Indian accents. That second one isn't a rounding error on a hiring product — so accent fairness is a launch gate, not a later fix.
The recruiter sets up two of them once. The other two then run on their own.
| Feature | What it does | Priority |
|---|---|---|
| Integration | ATS by API or webhook, plus Email, Calendar and WhatsApp — nothing works without it | FOUNDATION |
| AI voice agent | One agent per role type: instructions, 5–7 questions, call length. Test on your own phone, then publish | P0 |
| Create opening | Pick source, agent, workflow and scheduling link. Candidate list with stage, status and "new today" | P0 |
| Calendar & scheduling link | Self-scheduling built from real panel availability; every booked interview on one calendar | P0 |
| Workflow | Email and WhatsApp template per stage, sent automatically on status change | P1 |
Four principles the whole product is built on:
Sourcing. AI rejecting anyone. Fake-candidate detection. Offer and background checks. Predicting quality of hire. Replacing the ATS. Each one is a real product — none of them move the 28 hours, and the last one turns a six-week integration into a two-year migration.
Written from the sentence a recruiter would actually say out loud.
I want the AI to run the first screening call and hand me what the candidate actually said, so I stop spending my day discovering people aren't a fit.
Key criterion: a candidate can never reach Rejected without me clicking it.I want messages to send themselves when a candidate's status changes, so I stop writing the same note 180 times and nobody gets forgotten.
Key criterion: a rejection message must name the reason area — no generic template allowed.I want the scheduling link to go out the moment a candidate is approved, so they book themselves and I never chase a slot again.
Key criterion: only genuinely free panel slots are ever shown.I want one screen showing where every candidate is and what's waiting on me today, so nobody falls through while I'm busy with another role.
Key criterion: anyone waiting on me over 48 hours gets flagged.I want to know where I stand within 48 hours, and to be told why if it's a no, so I can get on with my life instead of guessing.
Key criterion: every message carries a route to ask for a human review.I want to reuse my agent, workflow and scheduling setup on a new opening, so launching a role takes ten minutes instead of half a day.
Key criterion: the same role next quarter asks the same questions, so rounds are comparable.The problem isn't reading resumes faster. It's that the resume no longer tells us who is worth calling — so we create new evidence.
Score only what the candidate said — never accent, tone or confidence. Name, photo, gender, age and college are never used. No quote means not assessed, never guessed. Consent taken before recording; data stays in India. If the model is down, candidates go to a manual queue — nobody is stranded.
If the recruiter override rate drops below 5%, that's an alert, not a win. It means recruiters have stopped reading the evidence and started rubber-stamping the machine — which is exactly the failure this design exists to prevent.
The recruiter console, the evidence card, and the signed links the hiring manager and candidate actually see.
Opens in a new tab. Every state is real — not reachable, no phone number, withdrawn and noisy line are all in there, because those are the cases a demo usually hides.
Open the Backend Engineer opening, select the new candidates and send them to the AI call.
Watch a call run, then open the evidence card — every rating carries the quote behind it.
Try to reject someone without a reason. You can't. Then override a rating and see it logged.
Advance a candidate, approve as the hiring manager, and book a slot as the candidate.
Targets I'd hold the launch to — not results.
| Before | After | |
|---|---|---|
| Recruiter | 28 of 45 hours on calls and coordination | ~10 hours back weekly, spent on judgment |
| Candidate | Applied, heard nothing for weeks | Knows in 48 hours, books their own slot |
| Company | 45 days to fill a role | ~30 days — 15 days recovered, 1,200 times a year |
Roughly ₹55 lakh of recruiter capacity returned each year across 30 recruiters, against ₹9–12 lakh of AI calling cost. Built on 1,200 hires, 5 screening calls per hire, 40 minutes per call and ₹520 per recruiter hour — every one of which I'd replace with the customer's real numbers in week one.
Where I'd stop and rethink:
Open questions still on the wall:
Want to go deeper, or talk through the calls?