AI Is Now in Your Interview Process. Here Is Where.

Three places it has already arrived

Automated assessment in hiring is no longer speculative. Three forms are common enough that most people running a serious job search will meet at least one.

One-way recorded interviews. You read a question on screen and record an answer with nobody present. Usually timed, often with limited retakes.

Transcription and scoring. Live interviews recorded and transcribed, with a model summarising or scoring against a rubric before a human reviews.

AI note-taking. The interviewer runs a tool that transcribes and summarises. Increasingly ordinary, and it changes what happens to your words afterwards.

None of these replaces the human decision at the end. All of them change what performs well earlier in the process.

The one-way interview, which is harder than it looks

Candidates dismiss these as a formality and then perform noticeably worse than their live interviews suggest they should. The reasons are mechanical.

There is no interviewer reacting, so you lose every cue that normally tells you the answer landed or that you have talked long enough. There is often a countdown. Sometimes there is one take.

What to do differently:

• Use the practice question properly. Most platforms offer one and most candidates skip it. It is the only chance to hear what you actually sound like on the recording.
• Put notes on paper beside the camera, not on screen. Reading from a screen is visible and reads as evasive.
• Answer to the lens. Watching your own image on screen means never appearing to make eye contact.
• Keep answers shorter than you would live, around ninety seconds. Without an interviewer to interrupt, digressions run long.
• Start with your answer, not with a preamble. If the recording cuts off, you want the substance already said.

Test everything the day before. Camera, microphone, lighting, browser. These platforms are often the only time you will use that software, and a technical failure mid-answer is unrecoverable on a single-take format.

How automated scoring actually works

It is worth being accurate here, because the marketing on both sides is unreliable.

Most systems in common use work from the transcript. They assess what you said against a rubric derived from the role: whether you addressed the question, whether you gave a specific example, whether you covered the competency being probed.

Claims about assessing personality from facial expression or vocal tone have attracted regulatory attention in several jurisdictions and some vendors have withdrawn them. Treat any suggestion that a system reads your character from your face with scepticism, and know that it is restricted or prohibited in a growing number of places.

The practical implication of transcript-based scoring is simple. Clear audio matters more than a good background. Complete sentences matter more than charm. Naming the thing directly matters more than implying it.

One clear takeaway: if it is scoring a transcript, say the words rather than conveying the impression.

Answering in a way that survives both readers

The useful news is that you do not need two versions of yourself. What works with automated assessment is a slightly more disciplined version of what works with a person.

Answer the question that was asked, explicitly, in the first sentence. A human forgives a slow build; a rubric checks whether the point was made.

Use the specific noun rather than the pronoun. "When I rebuilt the reporting process" is more legible in a transcript than "when I did that".

Give one concrete example per answer with real detail: what the situation was, what you did, what happened. Both readers reward specificity, and it is the thing generic answers lack.

Name the competency if you know it. Where a question is clearly about conflict, saying "this was a disagreement about priorities" gives both readers the label they are looking for.

Avoid filler. "Kind of", "sort of", "you know" survive into transcripts and read badly there in a way they do not in speech.

What you can ask, and reasonably expect

Candidates assume these processes are take-it-or-leave-it. Often they are not.

You can ask whether AI is used in the assessment and how. Several jurisdictions now require disclosure, and companies that are on top of their obligations will answer plainly. A company that will not is telling you something.

You can ask for an alternative format. This matters most for accessibility, and where a disability makes a timed one-way recording unreasonable, an adjustment is a legal obligation in many places rather than a favour.

You can ask whether a human reviews the output, and at what stage. The answer changes how much weight to give the exercise.

Asking is not held against candidates in my experience of how these processes are run. It reads as someone paying attention.

When the interviewer is running a note-taker

Increasingly, a live interview is transcribed by a tool that produces a summary afterwards.

Two consequences. Your answers are read later, in text, by people who were not in the room, so the same legibility points apply. And what you said is on the record in a way it previously was not, which is worth remembering when discussing a former employer.

You are generally entitled to be told. Where a recording notice appears and you would rather not be recorded, saying so is reasonable, though be aware that declining sometimes ends the process at companies that have standardised on it.

The assessment before the interview

Increasingly the automated stage arrives even earlier, as a test sent with the application. Situational judgment questions, a short cognitive test, a work sample, or a personality questionnaire.

These are used because they are cheap to run at volume, and they are genuinely predictive for some roles. They are also where a large share of candidates are eliminated without ever speaking to anyone.

A few practical points. Situational judgment tests reward the answer that fits the employer's stated values rather than the answer you would give a friend, and the values are usually published on their careers site. Timed cognitive tests reward practice, because the format is unfamiliar and speed matters more than difficulty. Personality questionnaires reward consistency, since most contain repeated questions phrased differently to detect people answering strategically.

Do them when you are alert rather than at eleven at night, and do them in one sitting. Fatigue shows up in the results more than most candidates expect.

Where these processes are unfair, and what to do

It is worth being straight that automated stages disadvantage some people for reasons unrelated to capability.

Non-native speakers are penalised by transcript-based scoring in ways a human conversation would forgive. People with poor connectivity are penalised by video platforms. Anyone with a speech difference, a stammer, or a condition affecting timing is disadvantaged by a countdown.

Two things help. Ask for adjustments early rather than after a poor result, since most employers can extend time limits or offer a live alternative and many will not think to offer unprompted. And where a process feels arbitrary, weigh what it tells you about the employer. A company that will not discuss its own hiring process is unlikely to be more transparent once you work there.

Using AI to prepare, without arriving over-rehearsed

The obvious move is to have a model generate likely questions and draft answers. The first half is useful, the second half is a trap.

Generated questions are genuinely close to real ones, particularly when you supply the job description. Rehearsing against them is time well spent.

Generated answers are the problem. They arrive fluent, generic and unconnected to your actual history, and they collapse on the first follow-up because you have not lived them. The pattern is recognisable to interviewers now, in the same way AI-written cover letters are.

Use it for the questions and for interrogation. Paste your own answer and ask what a sceptical interviewer would probe. That produces the follow-up you are unprepared for, which is the point of practising.

Our interview prep tool generates questions against your actual role rather than generic lists.

What has not changed

Worth saying, because the topic invites overreaction.

The hiring decision is still made by people. Automated stages filter and summarise; they do not choose. Every process still ends in a conversation with someone who will work with you.

Referrals still bypass most of this. A candidate introduced by someone inside frequently skips the automated stages entirely, which remains the highest-return action in any job search.

And the underlying preparation is unchanged. Six to eight real examples, known in detail, told clearly. That works in a recorded interview, in a transcript, and across a table. Our career guides cover what each role's process typically involves.

Questions people ask

Do companies really use AI to score interviews?

Some do, mostly at larger employers and in high-volume hiring, usually working from the transcript. Smaller employers largely do not, and a human makes the final decision either way.

Can I ask whether AI is being used?

Yes, and in several jurisdictions companies must disclose it. Asking is reasonable and reads as attentiveness rather than suspicion.

How long should answers be in a one-way interview?

About ninety seconds, shorter than you would speak live, with your actual answer in the first sentence rather than after a preamble.

Does my background or appearance affect an automated score?

Most systems in common use score the transcript. Claims about reading personality from face or voice have drawn regulatory action and are restricted in a growing number of places.

Can I request a different interview format?

Yes, particularly where accessibility is involved, in which case an adjustment is usually a legal obligation rather than a favour.

Related reading

• What a Recruiter Reads in the First Ten Seconds
• What Interviewers Are Actually Deciding About You