Automated Candidate Screening: Build vs Buy in 2026

A CTO at an Indian fintech told us last quarter that her engineering team had quoted six months and two engineers to “build the AI screening thing internally.” She came into the conversation expecting us to argue against it. Instead we ran the actual scope past her: speech-to-text, text-to-speech, the LLM stack, WebRTC plumbing, telephony for landline candidates, scoring rubric, recording storage, candidate-facing scheduling, the admin dashboard, the audit trail her legal team would eventually ask about. By minute fifteen she had stopped writing.

The build-versus-buy question for AI screening is really two questions in a trenchcoat. The first is what “automated screening” means inside your stack: a feature in the ATS you already pay for, a separate platform, or a layer that sits between the two. The second is the classic engineering question of whether your team can ship and own this thing without it becoming the project that swallows H2.

What you actually have today

Most TA teams already pay for an ATS that lists “AI screening” on its feature page. Open the screen and what you will usually find is a resume-ranking model that scores incoming applications against a job description. That is a useful feature. It is not screening. Screening means asking the candidate a question, listening to how they answer, and producing a signal good enough to decide whether to invest a hiring manager’s hour in them. Resume-rank does not do that; it just sorts the queue. If your ATS pitch sounds like screening, ask the vendor for a sample 12-minute conversation transcript with a candidate’s voice in it. Most cannot show you one because the product does not have one.

This is the gap the build-versus-buy debate is actually about. Resume-rank is solved. Conversational screening, the first 15 to 20 minute structured chat, is not.

What “build” actually entails

Suppose you decide to build. Sketch the stack honestly.

A real-time voice agent needs speech-to-text that works on Indian-English accents (Deepgram is the current default; competitors exist but lag). It needs text-to-speech that does not sound robotic enough to make a candidate hang up (Cartesia leads on naturalness right now). It needs an LLM that can handle conversational follow-up without losing the rubric (we run Azure OpenAI’s GPT-4o-mini for cost reasons; you will want your own deployment so candidate audio does not cross a third-party boundary). It needs WebRTC infrastructure to actually carry the audio (LiveKit is the open-source option). It needs SIP bridging if you want to call candidates whose only contact is a phone number. It needs recording storage on something like Azure Blob, an admin dashboard for the recruiter, and a scoring pipeline that produces audit-friendly outputs.

For reference, our production stack at HireQwik runs as a 12-service Docker Compose deployment on an Azure VM. That is not because we love operating infrastructure. It is because the minimum viable surface area of “AI voice screening” is genuinely that wide. The team underestimates this consistently. Six months and two engineers gets you a demo, not a product that survives a 3,000-candidate drive.

What “buy” actually looks like

The market has split into two shapes. There are platforms (HireVue, Eklavvya, Alex) that want to be the place your candidates live during screening; they bring their own funnel, their own dashboard, their own opinion about how the conversation goes. Then there is an emerging category of plug-in screening layers that sit between the candidate and your existing ATS, doing the conversation and pushing structured outputs back into whatever pipeline tool you already use.

We are biased toward the second shape because that is where we built. The reason is not ideological. No TA team we have talked to wants a second pipeline tracker. They want their existing ATS, plus a screening layer that does not make them re-key candidates. The platforms charge for the ATS-shaped surface area you did not ask for. Plug-in layers charge for the conversation, which is the part that actually moves.

On price, the spread is wide. Phone-screen outsourcing costs ₹85 to ₹150 per candidate. Video interview platforms run ₹100 to ₹300 per screen with the additional headache of a roughly 50% drop-off rate, because asking a fresher to record themselves on video introduces friction that does not exist in a phone call. AI voice screening platforms vary, but for context our own pricing sits at ₹59 per interview, substantially cheaper than the platforms it competes with. Cost should not drive the decision alone, but it does shape the build math.

A decision framework that actually fits

Three questions will resolve build-versus-buy faster than a vendor demo.

First: Is conversational screening a wedge in your product, or a back-office function for your team? If you are a hiring marketplace selling screening to your customers, build it. If you are an enterprise running campus drives twice a year, you are not in the screening business; buying is almost certainly correct.

Second: How many candidates flow through your funnel per quarter? At low volumes, say a few hundred candidates per quarter, the build math collapses immediately, because the per-candidate engineering cost of a custom platform is enormous. At very high volumes (tens of thousands per quarter and a domain-specific scoring rubric no off-the-shelf vendor can replicate), build starts to make sense. Most enterprise hiring teams sit between those poles and should buy.

Third: Does your legal or compliance org have an audit-trail standard for AI-driven hiring decisions yet? If yes, ask vendors for it explicitly. If no, your engineering team will eventually have to retrofit one anyway, which is one of the largest hidden costs of build. As we argued earlier, the ROI of AI screening is rejection accuracy, not speed, and rejection accuracy lives or dies on the audit trail.

The honest answer

Build if screening is your product. Buy a platform if you want a self-contained second pipeline. Buy a plug-in layer if you want your existing ATS to keep being the system of record and just need the first conversation handled. The wrong move is treating “AI screening” as an ATS feature checkbox; that is the option that produces the worst possible version of all three.

If you would like to talk through where your team falls on those three questions, drop us a note. Happy to compare math without a sales pitch attached.