What actually happens in an AI technical screen

This round barely existed a few years ago and is now the first filter at a lot of companies, including ones hiring senior engineers. Candidates hate it, partly because it is unfamiliar and partly because it gives you nothing back. No nodding, no follow-up smile, no sense of whether that answer landed.

It is also, once you understand what it is measuring, one of the more predictable rounds to prepare for. Here is what is going on.

Why companies started doing this

A single open backend role can attract several hundred applications. Traditionally a company would resolve that with a recruiter reading CVs, which is fast and enormously noisy, and then a thirty minute engineer phone screen for maybe fifteen people, which is accurate and expensive.

An automated technical screen sits in between. It gives every applicant the same length of technical conversation without burning a senior engineer's afternoon, and it produces something a CV cannot: a record of how you talk about your own work. Vendors in this space, micro1, Karat, HireVue and others, differ in the details, but the shape is broadly the same.

The uncomfortable implication for candidates is that the round is designed to compare you against a lot of other people on the same axis. Sounding fine is not enough, because everyone sounds fine.

What actually happens in the thirty minutes

1. You get a link, usually with a window of a few days to complete it. Camera and microphone permissions up front.
2. A short warm-up, often about your background or a project on your CV.
3. A sequence of technical questions, typically anchored on the stack in the job description.
4. Follow-up questions generated from what you just said, not from a fixed list.
5. Sometimes a small coding or debugging exercise.
6. It ends. Your answers are transcribed and scored, and a human at the company reads the summary later.

That fourth step is the one that catches people out. Most candidates prepare a set of answers. In this format, the answer you prepared is only the doorway, and the thing that gets scored is where you go when it asks you why.

Five ways it differs from a human phone screen

1. There is no charity

A human interviewer who likes you will fill in your gaps. They will hear a half-formed answer, recognise the shape of the right idea, and move on. Automated scoring does not do you that favour. If you did not say the thing, you did not say it.

This is the single biggest adjustment. Say the mechanism out loud even when it feels obvious. "The event loop is single threaded, so a synchronous call there blocks every other request on that process" is worth stating explicitly, even though with a human you might have shrugged and said "yeah, blocking".

2. It is scoring your transcript, not your vibe

Warmth, rapport and the fact that you both used to work with the same person do not exist in this round. What exists is the text of what you said. That is bad news if charm was carrying part of your interview performance, and good news if it never was.

3. The follow-ups are relentless and unembarrassed

A human interviewer will often stop drilling once they are satisfied, or once they sense you are uncomfortable. Automated follow-ups have no social brakes. If you use a term loosely, expect to be asked about that exact term, and then about the answer you gave to that.

4. Silence is expensive

Thinking time reads as nothing at all. You do not have to fill every second, but narrate the shape of your thinking rather than going quiet for twenty seconds: "Two ways to approach this, I'll take the second one because...".

5. You cannot ask what they meant

Some systems handle clarification well, many do not. Assume you get one interpretation of the question. State the interpretation you are answering under, in one clause, then answer it. "Assuming you mean at the API layer rather than the database, then...". Nobody ever lost points for that sentence.

Why strong engineers fail this round

The failure modes are consistent, and mostly they are not knowledge failures.

• Buzzword without mechanism. Saying "we used a queue for backpressure" and having nothing underneath when asked what happens when consumers fall behind.
• Rambling. Long, unstructured answers that never state a conclusion. Every extra sentence dilutes the ones that mattered.
• Answering the CV, not the question. Describing a project when you were asked how something works.
• Deferring to the team. "Our platform team handled that" is true and reads as a gap. Say what you know about it anyway.
• Treating it as casual. It is a recorded technical interview that decides whether a human ever reads your CV. It is not a chat.

How to prepare for this specific round

Prepare depth, not answers

Take the five or six technologies in the job description that are also on your CV. For each, be ready to go three layers down: what it is, how it works underneath, and what happens when it fails. The screen will find whichever of those three you skipped. If you want to see the difference that makes concretely, these eight side-by-side answers are the same questions answered with and without that second layer.

Rehearse out loud, into a microphone

Speaking a technical answer is a different skill from knowing it, and the first time you discover the difference should not be during the real screen. Record yourself answering three questions and play it back. Almost everyone finds the same two things: they took forty seconds to reach the point, and they used a term they cannot define.

Fix the room before you fix the answers

• Wired headset or a decent external mic. Laptop microphones in a room with hard walls transcribe badly, and a transcription error costs you an answer.
• Quiet room, door shut, notifications off, phone in another room.
• Stable connection. Restarting halfway through is not something you want to test.
• Do it when you are alert. This round has no small talk to warm up on, so you are cold from the first question.

Have your own projects loaded and ready

Pick three things you have built. For each, know the numbers, the hard part, the thing that broke, and the decision you would make differently now. Most technical questions in this format can be answered through a real example, and a real example with a number in it beats a textbook definition every time.

Is it fair?

Partly. It is more consistent than a human phone screen, which is genuinely worth something, because human screens vary enormously with who you get and what kind of morning they had. Every candidate gets the same length of conversation and the same attention, which is more than most CV piles get.

It also flattens things a human would have caught. Accents and non-native English can be penalised by transcription rather than by substance. Nerves cost more than they should. And a candidate who is brilliant on a whiteboard and awkward on a recording will read worse than they are.

Fair or not, it is now the door. The pragmatic move is to treat speaking about your own work under follow-up pressure as a skill worth practising deliberately, because it is one, and almost nobody does.