The interview already has an AI in the room. It just isn’t yours.

Earlier this year, Greenhouse surveyed 2,950 active job seekers across five countries, 1,132 of them in the UK. Among the UK respondents, forty-seven per cent had already been interviewed by an AI. Of those, eighty-two per cent were not told beforehand that a machine would be assessing them. Three in ten had walked away from a hiring process because of it.

So the debate about whether AI belongs in interviews is over. It’s in the room. The open questions are which side of the table it sits on, and whether anyone tells you it’s there.

Here’s what tends to go unsaid about how these decisions get made: rarely on the content of the answers alone. They’re made on how the answers arrive. The steadiness. The energy. Whether the person describing the project sounds like they were actually there.

Twenty-five years of hiring for design organisations taught me this the slow way. Interviewers have been reading vocal delivery forever; we just called it gut feel and moved on.

What’s changed is that employers now have software doing that reading at scale, silently, on the other side of the camera.

Candidates, mostly, have nothing.

The tools on offer are the wrong tools

Look at what the market currently sells job seekers. At one end sit the “interview copilots”: tools that whisper answers into your ear during the actual interview, advertised with words like undetectable and stealth mode. Leave the ethics aside; the practicality alone should stop you. Your plan for the biggest conversation of your year is to read from a teleprompter while a stranger studies your face. That isn’t confidence. It’s a dependency you have to hope nobody notices.

At the other end are question banks and one-way video drills. Useful, as far as they go. But rehearsing answers in text, or recording yourself into the void, does nothing for the thing the interviewer is actually reading: how you sound, live, under pressure, when the follow-up arrives.

In the middle, most people just open ChatGPT. It’s free and it’s not useless. It will also cheerfully tell you your answer was great, because it’s built to be agreeable, and it cannot hear you. It has no idea your pace doubled when you reached the difficult part, or that your energy fell off a cliff eight minutes in. The layer of the interview that shapes the decision is invisible to it.

The muscle nobody warns you about

There’s a pattern I keep noticing in people mid-search, and it’s the one that led to this product existing. While you’re employed, you talk about your work constantly without noticing: stand-ups, reviews, pitches, the corridor version of what your team shipped. The muscle that turns what you do into something another person can follow gets exercised every day, for free.

Then you’re job-searching, and it stops. The redundancy that put you on the market is the same event that cancelled your daily reps. Weeks go by in which you write applications, polish documents, think about your experience endlessly, and say almost none of it out loud. By the time the interview arrives, the muscle you need most has been resting precisely as long as you’ve been looking.

This isn’t a metaphor I’m asking you to take on trust. Skill decay is one of the better-measured effects in performance research: a 1998 meta-analysis by Arthur, Bennett and colleagues found that skills decline steadily through periods of nonuse, and that cognitive skills, the explaining kind, decay faster than physical ones. And the reverse effect is measured too. Psychologists call it the production effect: saying something aloud encodes it more deeply than reading it or thinking it, which is why the first spoken telling of a story is clumsy and the fifth is tight. Thinking about your answers is not rehearsal. Your mouth has to do it.

Interviewers meet the consequences of this every day and misread them. A candidate who sounds hesitant about their own career usually isn’t hiding anything; they just haven’t said any of it aloud for four months. The atrophy reads as doubt. That misreading costs people offers.

What we built instead

PrepCall is a voice coach for interviews. You upload the actual job description you’re going for: a Head of Midwifery post, a principal engineer role, a marketing directorship, a teaching job. You have a twenty-minute spoken conversation with a coach that has read it, and your CV, and pushes back the way a good interviewer does. Afterwards you get a written report on both layers of what you did: what you said, and how you sounded saying it. Pace, energy, steadiness. Which answer landed, which one drifted. The next session starts from what the last one found. It is, at its plainest, a place to keep the talking-about-your-work muscle warm for the whole length of a search, so the real interview is your twentieth telling rather than your first.

The analysis under the bonnet is the same category of technology employers are quietly pointing at candidates. That is exactly why we built it, and exactly why we built it the way we did.

The line we drew

From the first architecture decision, PrepCall has had one non-negotiable rule: it is sold to the candidate, and only to the candidate.

Concretely:

Your practice sessions are never shared with an employer, a recruiter or a hiring platform. Not as scores, not as summaries, not in aggregate. Nothing you do in practice can reach anyone making a decision about you.

The audio of your voice is never stored. It’s analysed as you speak and then it’s gone. What remains is your transcript and the measurements of how you sounded, visible to exactly one person.

And we tell you precisely what is analysed, in plain English, on screen, before and during every session. Since 2 August, EU law (Article 50 of the AI Act) has required systems like this to disclose themselves to the people they are used on. I would like to say we were ready on the day. We were not: the disclosure went live the following week, written by one person who read the regulation and then went and built it. What I will say is that nothing about it was reluctant. The product was designed on the assumption that you have a right to know what is being measured, and the law moved the deadline rather than the intention.

One thing I won’t claim, because overclaiming is half the disease in this category: PrepCall measures how your voice sounds, not what you feel. It doesn’t read minds and it won’t pretend to. Your voice can go flat on a day you feel perfectly fine. But the person across the table only ever gets the sound, and the sound is trainable. That’s the entire proposition.

Why this matters beyond one product

Roughly six in ten of the job seekers Greenhouse surveyed said disclosure of AI assessment should be a legal requirement. In Europe, it now is. Over the next year I think the tools in this space will sort into two camps: those that treat the new transparency rules as friction to be minimised, and those that treat them as the product.

We’re in the second camp, on the candidate’s side of the table, permanently. If the machines are going to listen either way, you deserve one that reports only to you.