45-minute AI video interviews, technical or coding, with anti-cheat

Most AI video interview software is asynchronous: the candidate is sent a list of questions, records themselves answering into a webcam with nobody on the other end, and uploads the result. A recruiter watches it later at 1.5× speed.

That format is popular because it is easy to build and easy to scale. It is also the reason candidates dislike video interviews. There is no one to ask a clarifying question, no chance to recover from a bad opening, and no way to demonstrate anything the fixed question list did not anticipate. Strong candidates with options routinely refuse to do them.

A QuillHire video interview is a live conversation. The AI is present, it listens to the answer, and the next question depends on what the candidate just said. When somebody gives a rehearsed answer, it asks the follow-up that a recording never can. That follow-up is where most of the actual signal lives.

It also changes what anti-cheat can see. Reviewing a finished recording tells you what was submitted. Monitoring a live session tells you what happened while it was being produced: the pause before an answer, the second voice, the eyes tracking a second screen, each with a timestamp your team can jump straight to.

There is a single 45-minute format, and it changes shape depending on who is taking it. Pick a developer role and the session includes a live coding round. Pick anything else and the same 45 minutes is a technical conversation with no code in it at all.

The choice is made during setup, or the platform makes it from the role you selected. You can override it either way: a coding round for a data analyst who will be writing SQL every day, or a discussion-only round for a staff engineer you want assessed on architecture rather than syntax.

This replaces what used to be two separate products. Running two formats meant every hiring manager had to decide which one a role needed before they could book anything, and that decision was wrong often enough to be a real cost. One format with a module removes the question.

What does not change between the two shapes: the length, the price, the anti-cheat coverage, and the evidence you get back.

“Technical” here means depth in the craft of the role, not proximity to code. A marketing analyst has a technical craft. So does an operations lead. The question is whether a candidate can reason inside their domain under pressure, and that is testable without asking anyone to write a loop.

What that looks like in practice, by role:

Product manager. Given a roadmap and a hard constraint (a quarter shortened, an engineer reassigned), what gets cut, and what is the reasoning? Strong candidates name the trade-off and who it hurts. Weaker ones reach for “we would prioritise ruthlessly” and stop there.

Marketing or growth analyst. A metric moved the wrong way last week. Walk through the debug. Strong candidates ask what else changed before proposing a cause. Weaker ones jump to a channel and defend it.

Operations or finance. A process breaks at three times current volume. Where does it break first, and what would you instrument to see it coming? Strong candidates think about the failure before the fix.

Customer success. An account is quiet and the renewal is ninety days out. What do you do, in what order, and what would tell you it is already lost?

The AI probes the answer rather than accepting it. A candidate who says they would "align stakeholders" is asked which stakeholder, and what they would say when that person disagrees. That follow-up is the entire test. It is where a rehearsed answer runs out and an actual one continues. Across 45 minutes, that produces a scorecard about judgement rather than about vocabulary.

On a developer role the session includes a real problem, calibrated to your stack and seniority band, worked live while the AI watches how the solution is built rather than only what arrives at the end.

It follows where the candidate starts, which cases they handle first, and what they reach for when stuck. A working answer that appears with no visible development path is itself a signal. It interrupts to ask why a data structure was chosen, what breaks at scale, and how the candidate would test it, which is the set of questions that separates understanding from recall.

Anti-cheat here covers the failure mode that matters most in 2026: a candidate quietly piping the problem through an AI assistant. Alongside the standard signals it checks whether the submitted code actually matches the assigned problem, and whether response timing is consistent with someone genuinely working it out.

Nothing is auto-rejected on the strength of that, which is deliberate. Fast, clean work looks different from pasted work, and both end up in front of a human with the full session captured.