The six layers of first-round screening

The ATS knockout and resume screen

What happens

Software reads the application and rejects most of them before any person is involved.

Sold on

Throughput. Fewer resumes reaching a human.

Who buys it

Whoever already owns the ATS. It is rarely a separate purchase.

This is the real first round for most companies and almost nobody sells against it, because it does not feel like a product decision. It arrives as a feature of the applicant tracking system you already bought. A keyword filter, a knockout question, a match score, and the majority of a pile is gone before a recruiter opens the list. In a ResumeBuilder.com survey of 948 United States business leaders fielded in October 2024, 82% said their company already uses AI to review resumes. Treat that as one commercial survey rather than a universal rate, but the direction is not really in dispute, and as of 2026 all five of the major systems named below ship AI screening of their own.

The reason it deserves the top of this map is volume. Every other layer on this page operates on the survivors of this one. If the knockout is tuned badly, no downstream tool can recover the people it removed, because those people are never seen again. A strong candidate with an unusual title, a career break, a resume written in the wrong vocabulary, or no resume-writing skill at all is filtered by a mechanism that never explains itself to anyone.

It is also the layer with the worst feedback loop in hiring. You can measure who you hired. You cannot measure who you rejected in error, so the filter never learns it was wrong.

Who plays here: Workday, Greenhouse, SmartRecruiters, Bullhorn, JobAdder.

The recruiter phone screen

What happens

A recruiter calls the candidate for ten to fifteen minutes and works through a checklist.

Sold on

Nothing. There is no software here. This is a person and a calendar.

Who buys it

Nobody buys it. It is the default.

This is still the most common first round on the planet, and it is the thing worth understanding properly, because it is what an AI interviewer is actually competing with. Not another vendor. This.

It is good in ways that get overlooked. It is a conversation, so it adapts. A vague answer gets a follow-up. A candidate who is nervous gets a second run at the question. Someone whose CV looks wrong for the role can explain in thirty seconds why it is not. No test, no game, and no recorded video prompt does any of that.

Its problem is not quality, it is arithmetic. Ten minutes each, and a role with four hundred applicants is sixty-six hours of calls that no coordinator has. So the phone screen does not scale down gracefully, it gets rationed. In practice, most applicants are cut on a resume skim precisely so that the few who survive can get the call, which means the best round in the process is applied to the smallest number of people, chosen by the worst one.

Psychometric and gamified assessment

What happens

The candidate spends roughly twenty minutes on tests or games measuring traits and cognitive ability.

Sold on

Prediction. A validated score that forecasts performance.

Who buys it

Enterprise graduate programs and large volume employers, usually with an occupational psychologist involved.

This is the oldest and most scientifically serious layer here. Its moat is validation: published studies, adverse-impact analysis, and defensibility if a rejected candidate challenges the decision. That is a real moat, and it is the reason this layer survives inside big graduate programs.

It buys prediction rather than proof. The output is a probability about how someone will perform, inferred from traits, not a record of them doing anything related to the job. For a graduate intake with thousands of applicants and no relevant work history to assess, that trade is often correct, and dismissing it is a mistake.

Where it strains is everywhere else. It asks a candidate for twenty unpaid minutes before anyone has spoken to them, it produces a score a hiring manager frequently cannot interpret, and it says nothing about whether this specific person can do this specific job next week. For a frontline role with a shift to fill, that is a long way round.

Who plays here: SHL, Aon Assessment (cut-e), Arctic Shores, Harver (pymetrics), Criteria (Revelian), Talent Games, Pulsifi.

Skills tests

What happens

The candidate completes a test or work sample that demonstrates a specific skill.

Sold on

Proof, not prediction. Evidence the skill exists.

Who buys it

Anyone hiring for a role where the skill is discrete and testable, heavily weighted to engineering.

The strongest layer here when the skill is genuinely testable. If a job needs someone to write working code, a coding test settles it better than any interview will, and it settles it the same way for every candidate. That is a real advantage and it is worth conceding plainly.

The layer is also the most honestly priced part of the market, with published tiers and self-serve free plans, which is why it dominates the small-company end.

Its limit is the shape of the job. Most of the world's hiring is not for discrete testable skills. A test cannot tell you whether an aged care worker will be patient with a distressed resident at 3am, whether a support worker has genuinely done the shifts they listed, or whether a candidate can explain a decision they made under pressure. That is a conversation, and outside desk and engineering roles it is most of what a first round has to establish.

Who plays here: TestGorilla, Vervoe, Xobin, iMocha, Mercer Mettl, Alooba, HackerRank, Codility, CodeSignal.

One-way async video

What happens

The candidate records answers to fixed prompts, alone, with no one on the other end. A recruiter watches later.

Sold on

Recruiter time. Watch on your schedule, at double speed.

Who buys it

High-volume employers, and the layer with the deepest enterprise install base.

This was the previous decade's answer to the scaling problem, and structurally it is a recording, not an interview. The prompts are fixed, so nothing adapts. A candidate who misunderstands a question answers the wrong one perfectly. A vague answer stays vague, because there is nobody there to ask the obvious follow-up. A strong candidate who needed one clarifying sentence to shine never gets it.

It does solve the recruiter's calendar, which is why it scaled, and it is the honest reason it still has the install base it does.

It is now the incumbent that most of the market is attacking, and the attack is coming from the candidate side rather than the buyer side. Talking to a camera in an empty room is a fundamentally strange thing to ask someone to do, and it asks it at the exact moment a candidate is deciding whether your company is worth the effort.

Who plays here: HireVue, Spark Hire, VidCruiter, Willo, Jobma, Hireflix, myInterview.

Conversational AI screening (chat)

What happens

An AI screens the candidate by text chat, often over SMS or a web widget.

Sold on

Speed, reach and completion. Meet the candidate where they already are.

Who buys it

High-volume hourly and frontline employers: retail, hospitality, logistics, banking, care.

The most commercially successful application of AI in the first round so far, and the reason is unglamorous: text has the lowest friction of any format on this page. No camera, no scheduling, no download, answerable on a bus. Completion rates reflect that, and this layer is genuinely excellent at scheduling, screening questions and knockout logic at a scale nothing else here matches.

The consolidation is worth noticing on its own. When a major ATS buys the leading chat screener, it signals the layer is becoming table stakes inside the system of record rather than a standalone purchase.

The ceiling is what text can carry. You can establish facts by chat: availability, right to work, certifications, years of experience. What you cannot establish is how someone speaks about their work: the hesitation on a claim they cannot back up, the detail they add unprompted about a resident they cared for, the difference between a rehearsed answer and a lived one. For roles where communication is the job, that gap is the whole assessment.

Who plays here: Paradox (Olivia), Sapia.ai, Humanly, XOR, Sense.

AI live interviewers

What happens

An AI conducts a real-time interview: it asks, listens, follows up, and adapts to the answer.

Sold on

Give every applicant a real interview, not a recording and not a test.

Who buys it

Still being decided. This is the newest layer and the least settled.

Ployo's own layer, and the most crowded one. It exists because the previous three all made the same trade in different directions: the test dropped the conversation, the one-way video dropped the listener, and the chat screen dropped the voice. A live AI interview is the attempt to keep all three at a cost per candidate that lets you run it on everyone.

The interesting differences inside this layer are not about who has an AI interviewer, because within a year everyone will. They are about what comes out the other end. A ranked list with a confidence score is a very different product from a per-requirement verdict with the candidate's own words attached as evidence, even though both are sold with the same sentence. The first asks a hiring manager to trust it. The second lets them check it in ten seconds.

The second real difference is what happens when the AI is wrong, because sometimes it will be. A system that hands back the recording, the transcript and the reasoning is auditable by the person accountable for the decision. A system that hands back a number is not.

Who plays here: Ployo, Apriora, Micro1 (Zara), HeyMilo, Ribbon, Classet, Wayfaster, Maki, Alpharun, Braintrust AIR, Mercor, Interviewer.AI.