The Recruiter Is a Bot: How to Pass AI Interviews (Zara, Mercor & Co.)

2026-09-16 · Expert Match AI team

If you have applied to an AI training platform in the last year, there is a good chance the first "person" you spoke to was software. micro1 routes nearly every applicant through Zara, its recorded AI interviewer. Mercor's screening interview is AI-conducted and then scored before any human reads it. Several smaller platforms have bought the same tooling. On a board where micro1 alone accounts for 145 of our 222 priced roles, the AI interview is now the single most common gate between you and paid work.

That is less alarming than it sounds, because the bot is doing something narrow and predictable. Our resume post covered writing for the parser; this is the spoken version of the same idea.

What the interview is actually doing

An AI interviewer is not forming an impression of you. It is transcribing what you say, then matching the transcript against the role's requirements - the same requirements the resume screen used - and against a rubric for how clearly you explain things. Three things drive the score:

Coverage. Did the required nouns show up? A listing for a "Corporate Attorney (BigLaw Firms)" wants to hear the firm, the practice area, the deal types and the years. A listing for a "Senior Full-Stack Engineer" wants the stack named and the scale of what you built. If you talk for four minutes and never say the words on the listing, the transcript does not match, however impressive the story.

Specificity. Rubrics reward concrete detail - numbers, names of tools, dates, outcomes - over adjectives. "I led a team" scores below "I ran a team of six through a nine-month migration off a legacy billing system." The model cannot verify the claim, but it can tell whether the answer contains verifiable-shaped content.

Communication. Most rubrics include a plain-language dimension: was the answer structured, did it answer the question asked, was it the right length. This is where expert-tier candidates most often lose points, by giving the answer they would give a peer instead of the answer they would give an intelligent outsider. Remember what the paid work is: explaining your judgment clearly enough that a model can learn from it. The interview is a sample of that.

What it is not scoring, in our experience and from worker reports: your face, your accent, your lighting, or whether you smiled. Those are the things people spend their prep time on.

Preparing in an hour

You do not need a coach. You need the listing and a notepad.

1. Pull the exact listing text. Every role on our jobs board shows the description the platform published. Highlight every requirement and every skill noun. That is your keyword list.
2. Write three stories, each with a number in it. One that proves the core credential, one that proves the core skill, one that shows judgment under ambiguity. Each should take 60 to 90 seconds to tell and should contain at least one figure - a headcount, a budget, a duration, a result.
3. Say the credential string out loud, the way the resume says it. "Board-certified in internal medicine, licensed in Ontario since 2014." "CFA charterholder, eight years in fixed income." The transcript needs the string, not a paraphrase.
4. Answer the question, then stop. AI interviews are usually fixed-length per question. A crisp 75-second answer scores better than a rambling three-minute one that gets cut off mid-point.
5. Test your audio, not your background. Transcription quality is the one technical thing that matters. A bad microphone loses words, and lost words are lost keywords.

Platform by platform

micro1 (Zara). Recorded, typically 15 to 25 minutes, a mix of background questions and role-specific technical ones, sometimes with a live coding or written component for engineering roles. Zara will follow up on vague answers; treat the follow-up as a second chance to add the specific you left out. micro1 uses the interview to route you across a very wide board - $6 to $400/hr on today's listings - so the same interview can land you in a much better queue if the transcript matches a better listing. See our micro1 analysis for how those tiers are structured.

Mercor. Resume first, then an AI screening interview, then scoring into talent networks that labs draw on. Mercor's 17 priced roles today average $120/hr, and the interview decides which of its networks you sit in. Mercor questions lean toward "walk me through how you would approach X" - it is grading the reasoning, so narrate it in steps.

Surge, Braintrust, Turing, DataAnnotation. Mostly no AI interview: Surge and Turing screen on resume and written assessment, Braintrust on profile and client interview, DataAnnotation on a written starter assessment. The assessments guide covers those gates.

The line

Preparing stories, rehearsing your credential string and reading the listing carefully is what every good candidate has always done. Reading scripted answers off a second screen, using a live AI to generate answers during the interview, or having someone else take it for you is fraud against the platform and, on most of them, a permanent ban when detected - and detection is getting easier, because the same models that conduct the interview are good at spotting the cadence of read text. It is also self-defeating: the interview is a sample of the paid work, and if you cannot do the sample unaided, the queue will find that out in the first week.

Before you book it

The cheapest dry run is our own matcher. Upload a resume on the homepage and it ranks every live role against your background and tells you which requirement strings you match and which you do not - the same matching an AI interviewer does, run before the interview instead of during it. Then check the salary report to see whether the queue you are interviewing for is worth the prep.

Published by the Expert Match AI team. Descriptions of platform interview formats are drawn from platform documentation and public worker reports and may change without notice; check the platform's own instructions before your interview. Listing counts and rates are from our tracker as of September 15, 2026 and are listed rates, not accepted offers. Some outbound application links on this site carry disclosed referral codes; rankings and recommendations are never influenced by referral payout.