You join a video call and there is no one on the other side. A synthetic voice greets you, confirms your name, and tells you the first problem is on the shared editor. It waits. When you finish, it asks why you chose a hash map over sorting, whether your solution holds if the input is empty, and what the complexity is. It never interrupts, and it never gives you the small “yeah, keep going” cues a human interviewer leaks without meaning to. This is the AI-conducted screen, and by 2026 it has become a common gate before you ever talk to an engineer.
The format borrows its bones from the structured, recorded, rubric-scored phone screen Karat popularized: a fixed question bank and consistent scoring across candidates. What changed is that the interviewer is now software, and it comes in two flavors. Live conversational agents, like Eightfold’s AI Interviewer and the screens companies such as Mercor run at scale, listen and then branch their next question off what you just said. Asynchronous one-way assessments, the classic HireVue video format, have you record answers to fixed prompts with no live back-and-forth. This piece is about the live conversational screen, where the follow-up loop lives, though the narration habits below help in both. Either way, you are not judged by the machine alone: it collects evidence and a person reads it later.
What the hour actually looks like
Most AI screens run 45 to 60 minutes and open with an identity check. Some platforms fold in fraud detection: ID verification, a webcam check, sometimes a second-monitor scan. On what you are allowed to use, confirm with the recruiter, but commonly you get a real editor with syntax highlighting, sometimes autocomplete, a scratchpad, and a run button that compiles and executes against sample cases. Some environments lock down copy-paste and external tabs. Ask before the timer starts.
There is a short warm-up, one or two questions about a resume project, meant to get you talking and calibrate a baseline. Then the first coding problem lands in the editor, typically easy-to-medium difficulty, the kind of thing you would see early in a set organized by problems by difficulty. You write the solution while narrating. When you stop, the follow-up loop starts: the agent probes edge cases, asks for the complexity, and sometimes whether another approach would be faster or lighter on memory. Then problem two, same rhythm, often a notch harder.
What is missing tells you as much as what is present. There is rarely a genuine “do you have questions for me” segment, because the bot has no team or salary band to describe. A few platforms tack on a canned slot, but it counts for nothing. Budget your energy for the code and the reasoning, not for charming an interviewer who isn’t there.
Where it diverges from a human phone screen
The instinct to treat this like a normal screen is half right: the problems and clock are similar, but the feedback channel is gone and the scoring is more literal. A human interviewer infers that you understood the edge case from your tone; the agent needs you to actually say it. The table below is the mental model that matches how these systems work.
| Dimension | Live AI-conducted screen | Traditional human phone screen |
|---|---|---|
| Who runs it | A live AI agent (Eightfold, Mercor-style); a recruiter reviews the recording and transcript after | A live engineer or recruiter who decides in real time |
| Format | ~45 to 60 min: identity check, resume warm-up, two coding problems in a shared editor | ~45 min: brief resume chat, one or two coding problems |
| Follow-ups | Automated follow-ups triggered by your answer: edge cases, complexity, alternative approach, restated if your reply is incomplete | Improvised by the interviewer, who reads your body language and adjusts |
| Scoring | Answers matched against a fixed rubric; content over delivery; transcript and summary sent to a human | Holistic human judgment, written up in notes after the call |
| What to do differently | State every assumption and complexity out loud, close each clarification explicitly, keep talking through silence | Read the interviewer’s cues, ask them questions, let rapport carry some of the signal |
How the bot scores you
Vendors differ and none publish the exact rubric, so treat specifics as directional. The common pattern: the agent transcribes what you say and write, matches it against predefined, job-related criteria, and produces a structured summary for a recruiter, not a single pass/fail verdict. Eightfold’s own materials say its interviewer focuses on the content of your responses, your skills and reasoning, and does not grade accent, tone, pauses, or facial expressions. The scoring language carries across both formats: HireVue, describing its recorded assessments, says much the same, matching structured responses against predefined job-related criteria with a human reviewing after.
One thing vendors rarely spell out: whether your code is executed against test cases or scored only from the transcript of what you typed and said. Assume both. Write code correct enough to pass hidden tests, and narrate the reasoning behind it, so you score well whether a test harness or a transcript reader is grading.
Unspoken reasoning tends to earn nothing. Add a guard clause for an empty array without comment and a human might catch the intent from your keystrokes; a transcript-based scorer sees only a code change with no explanation. Say it out loud: “I’m adding a check for the empty input because otherwise this indexes out of bounds.” On a transcript-scored platform, that sentence is the artifact that earns credit, and the guard clause alone may not. The vendor may not grade your accent, but the speech-to-text still has to hear you: because scoring runs off that transcript, speak deliberately, especially names and numbers. If your accent trips up dictation, lean on the editor: typed code and comments are usually captured verbatim, so a short comment backs up a spoken point that might get garbled.
The same logic applies to complexity. When the agent asks for the runtime, give the bound and the reason, not the bare “O(n).” Something like “O(n) time because I touch each element once, O(n) space for the hash map” reads as complete and closes the branch. Recite the letter alone and you invite a follow-up asking you to justify it. A mental map of the common bounds, the kind on a Big-O cheat sheet, keeps these answers fast and correct.
Talking out loud when nothing nods back
The strangest part of the AI screen is the silence. In a human interview, small acknowledgments tell you your approach is landing. The bot gives none, and candidates fill the vacuum two bad ways: they go quiet and code in their head, or they over-explain and burn the clock.
The fix is to narrate the way you would in a design review, in short declarative beats. State the approach before you write it: “I’ll use a sliding window here since we want the longest substring without repeats.” Write. State the invariant you are maintaining, the condition your loop keeps true on every pass. Write. When you hit a decision, say which way you are going and why. That narration is the raw material the scorer grades, and it happens to be the same habit that makes you legible to a human panel too. If narrating under pressure feels unnatural, it is a rehearsable skill: a structured study plan with mock runs out loud fixes it faster than grinding problems silently.
Will the bot prompt you if you go quiet? Sometimes. A few nudge after a silence, some just wait, some flash a timer before a segment ends. Do not count on any of it: assume no nudge and no warning, keep a light commentary going, and watch your own clock.
Closing the clarification loop
AI interviewers re-ask. If your answer is partial, the agent circles back with a narrower version of the same question, because its branch logic, the script that picks the next question, is waiting for a specific piece of information before it moves on. Candidates read the repeat as “I got it wrong” and second-guess a correct solution. Usually it just means you left the target unsaid.
It cuts the other way too. When you have a real clarifying question, “can I assume the array is sorted?”, a live agent may not answer, and many will not engage the way a human would. Do not wait on a reply that may never come: state your assumption out loud, “I’ll assume the input is unsorted and handle that,” and proceed. If the agent had a constraint in mind, it surfaces in a later follow-up.
When it does ask, answer the literal question and mark it done. If it asks about edge cases, enumerate them as a set: empty input, single element, all duplicates, integer overflow if the values are large. Naming the category and giving concrete instances signals coverage and stops the loop. Vague answers (“I’d handle edge cases”) reliably trigger another pass, because the agent cannot confirm you named anything real. Think of each follow-up as a checkbox with an explicit condition, and say the thing that ticks it.
Pacing the follow-up gauntlet
The follow-ups are where the hour is won or lost, and they commonly come in a familiar order: edge cases first, then complexity, then “is there a better approach.” That last one is a trap if you treat it as an accusation. Often the agent is checking whether you know the tradeoff, not demanding a rewrite. A clean answer names the alternative and its cost: “I could sort first and use two pointers, which drops the space to O(1) but makes it O(n log n) time; I chose the hash map because the problem favors speed.” That shows range without rebuilding a working solution. Recognizing which coding pattern a problem fits, and which the alternative uses, lets you answer in one sentence instead of freezing.
Watch the clock yourself, because the bot will not rescue you. Most platforms hold strictly to timing and move on when the segment ends, finished or not. Get a working solution down first, even brute-force, then optimize if time allows. A correct O(n²) answer you explained well beats an elegant O(n) you never finished and never narrated.
None of this replaces knowing your data structures; the AI screen has not changed what makes an engineer legible in an interview, only the medium. It is also a different question from whether you are allowed to use AI tools while you code, which is a separate call each company makes; here the AI is the interviewer, not your assistant. For the wider shift, the AI-era interview guide covers how these rounds are changing, and if you want the counter-trend, plenty of companies are pulling final rounds back onsite because remote screening got noisy. One thing worth remembering while you talk to the empty screen: the transcript the agent produces is the first document a human reads about you before the next round, so the narration you did for the scorer is also your opening argument to the person who decides.
Practice the behavioral round:
