How Many Candidates Can Your AI Screen at 9 AM? The Concurrency Question Vendors Dodge

Read enough AI-screening vendor pages and you’ll see the same promise: unlimited concurrent interviews, infinite scale, screen everyone at once. It’s a comfortable claim because it’s hard to disprove on a sales call. It’s also the wrong way to think about your hiring drive — and the gap between “unlimited” and “what actually happens at 9 AM on launch day” is where campaigns go sideways.

Here’s the question to ask any vendor: if you email a self-schedule link to 500 campus candidates and 200 of them click “join” in the same fifteen-minute window, what happens? The honest answer is never “all 200 interview simultaneously.” The honest answer involves a number — a real ceiling — and a plan for what happens when you hit it.

“Unlimited” is a media-server claim, not a screening claim

A voice AI interview isn’t a web page load. Each live interview holds a real-time audio session: speech-to-text running continuously, an LLM generating responses turn by turn, text-to-speech synthesizing the reply, all inside a latency budget tight enough to feel like a conversation. That’s a stateful, compute-bound workload. You can scale it, but every concurrent interview consumes a fixed slice of capacity for its full 15–20 minute duration.

“Unlimited concurrency” quietly assumes someone else’s infrastructure is infinite and free. In practice every deployment has a real ceiling — set by GPU/STT throughput, model rate limits, and the box the agent runs on. A vendor who won’t name that number is either hiding it or hasn’t load-tested their own product. Both are bad signs going into a 3,000-candidate drive.

What real throughput looks like

Throughput on a screening campaign isn’t “max simultaneous interviews.” It’s a function of two things: how many interviews can run at the same time, and how long each one takes. Our default ceiling is 20 candidates per 15-minute slot — chosen because it’s a load the stack handles cleanly, not a number that looks good on a slide. With a 15–20 minute interview, that pacing screens hundreds of candidates in an afternoon without ever pretending that 500 people are talking to the agent at the same instant.

That’s how we screened 3,000 candidates in two hours in a single evening — not by claiming infinite concurrency, but by pacing admissions so the system stayed inside its real capacity the entire time. The candidate experience is identical whether they’re the first or the four-hundredth: they pick a slot, they join, the agent talks to them. What changes behind the scenes is admission control deciding when each interview starts.

Why pacing beats “unlimited”

When everyone hits a link at once with no admission control, one of two things happens, and both are worse than a short wait:

• Degraded interviews. The system accepts more sessions than it can serve, latency climbs, the agent starts talking over candidates or pausing awkwardly, and your screening quality collapses for everyone in the window. A bad interview is worse than a delayed one — you’ve now got a candidate with a poor impression of your company and an unreliable score.
• Silent failures. Sessions that can’t get a slice of capacity drop, candidates stare at a frozen screen, and you find out from angry emails instead of from a dashboard.

A real concurrency model refuses the overload instead of degrading. When a 15-minute window is full, the next candidate is offered another slot rather than admitted into a session the system can’t serve well. That’s the difference between a campaign that holds quality under load and one that quietly falls apart at peak — and peak, on a campus drive, is always 9 AM on day one.

Questions to ask before you sign

If you’re evaluating AI screening for a high-volume drive, push past the marketing:

1. What is your real concurrent-interview ceiling on the infrastructure I’ll be on? A vendor who can answer has load-tested. One who says “unlimited” hasn’t.
2. What happens when more candidates try to join than the ceiling allows? Listen for “we offer them another slot” — that’s admission control. Silence or “it just scales” means no plan.
3. Does interview quality hold at peak, or only in the demo? Ask for behavior under load, not a one-candidate demo.
4. Can I pace a campaign, or does everyone get the link at the same moment? Self-scheduling with slot capacity is what keeps the load survivable.

The takeaway

“Unlimited concurrent interviews” is a slogan, not an architecture. Every real voice-screening system has a ceiling, and the good ones are honest about it — they pace admissions, offer alternate slots when a window is full, and protect interview quality under load instead of accepting more sessions than they can serve. The right question for your next hiring drive isn’t “can it scale infinitely?” It’s “what’s the real number, and what happens when I hit it?” A vendor who answers that plainly is one you can run a campus drive with. One who dodges it will find the ceiling for you — on launch morning, in front of 500 candidates.

We built HireQwik’s concurrency model around honest admission control because that’s what survives a real Indian campus drive. If you want to pressure-test the throughput math against your own candidate volumes, talk to us.