FIND OUT WHO’S ACTUALLY GOOD WITH AI

Every résumé perfectly matches the role. The interviews pass with flying colors. The task of choosing seems impossible. A month later, the hire grades as a bust.

Upon closer inspection, the cover letter, the take-home, and the portfolio also read like AI. Every artifact a candidate hands over may tell you nothing useful.

What’s missing in today’s evaluation process? Assume everyone uses AI, and the vetting changes from skill to judgment. What they question, check, and take on faith.

At Olive, we evaluate candidates on how they apply judgment to guide AI.

ZipRecruiter’s economists survey more than 1,000 verified US hiring managers and talent-acquisition professionals. 74% call AI skills a strong advantage or a flat requirement.

The demand is not new. In a Microsoft and LinkedIn survey from February and March 2024, 66% of leaders already say they would not hire someone without AI skills, and 71% said they take a less experienced candidate with AI over a more experienced one without. Two years ago before recent findings emerge (on the next screen).

The problem: employers fail to check for the skills they want. The disconnect results in bad hires, failed AI projects, and wasted investment. The gap is observable.

ZipRecruiter’s own report says employers are changing what they screen for and automating the postings. The results still say nothing about how to verify the skill ask before an offer goes out.

Olive measures every active posting in its own corpus — 2,151,213 of them. 34,743 carry an AI-usage requirement specific enough for the skill extractor to name: only 1.62% show enough detail to define a skill.

Indeed’s economists run a version of the same test on the wider market. When they put a full year of AI-mentioning postings through a topic model, one in four gives little indication of how AI may surface in the role. Specific language about the work remains rare.

Uniquely, Olive counts the same corpus two ways — one reading retrieves what the extractor found, the other reading directly pulls the raw text. A mistake in one cannot hide inside the other. See below.

Two confident facts point in the same direction. Postings in human resources that mention AI double across 2025, climbing from 4.4% to 8.8%. Nationally, postings asking for AI skills grew 144% in a year while postings overall grew just 7%. Olive finds that the listings predict the drift before the process confirms.

Candidates remain ahead by doing the same thing employers ask about. In Gartner’s survey of 3,290 applicants, from late 2024, 39% say they use AI somewhere in the application process. Capterra’s survey of 2,997 candidates across twelve countries, run separately, puts AI use somewhere in the search at 58%. And 59% of job seekers, by Resume Genius in March 2026, use AI to write their résumé. Today, AI saturation seems obvious and inevitable.

Hiring managers describe what they see when the same kind of résumé lands on the desk again and again. 69% say résumés read perfectly but more generic than they did five years ago. 77% say many now look AI-generated, and 39% name perfect grammar with no variation as the tell. 86% say AI makes it too easy to embellish, 42% call the trend a serious hiring risk, and 34% report résumés often or always fail to match the skills behind them. Résumés carry less and less value for information.

The obvious fix — grade what a candidate hands in, and let the artifact speak for itself. The next record chronicles when a vendor tries exactly such a plan, on its own AI-run interviews, at scale.

The statistical version finds cheating useless. Among 5,179 customer-support agents, AI raises output 34% for the least experienced and barely moved the most experienced — the gap between them narrows. Among 453 professionals in a study published in Science, AI raises output quality 18% and narrows the spread between workers. A test that grades just the artifact loses its power to tell people apart.

On a task squarely inside what a model handles well, the gain can be enormous: one Copilot trial finishes a task 55.8% faster. The honest word here becomes jagged, not better or worse.

The fundamental question becomes not whether employees will use AI at work.

Can the hiring process find candidates best capable of leveraging AI for productivity?

What separates the helped from the harmed was not prompting. The advantage surfaces when the person interrogates what came back — inquires of the LLM on a fine point, checks a figure, or refuses an answer that merely sounded right. The secret identifies a behavior, not a skill with the tool. And Olive excels at recording behavior.

What Olive reads is the judgment a person applies while working with AI, not how well they prompt. The assessment rests on the candidate’s own reasoning, spoken (typed as an accommodation), as it happens; a task specific to one of twelve occupations; dimensions aim at how a claim gets checked rather than how an answer gets phrased; and evidence excerpts in place of a single number.

Others try pieces of our techniques. Canva’s engineers write that candidates unable to tell when a model was wrong struggle in their interviews. Zapier judges how a candidate pushes back on an output. CodeSignal names a skill it calls GenAI Limitation Awareness. Bryq names one such dimension in its assessment. Each names a different piece of the same judgment.

Eleven of twelve roles sit in the Bureau of Labor Statistics’ top exposure tier — 206 of 831 occupations. Software engineering, financial analysis, management consulting, data & analytics, product management, marketing, journalism, healthcare revenue cycle, legal operations, accounting & audit, underwriting & claims, supply chain.

These places, within an hour with an AI colleague, receive inquiries during the work rather than placed afterwards. The next section shows what that hour looks like, from the opening document to the written debrief.

The session begins in a browser, one real task from the role’s case set, a packet of documents, and a provided (or bring your own) AI assistant in the same workspace. The candidate thinks aloud throughout. The clock runs 55 minutes for a loan case, 60 for an engineering ticket, with a hard cap of 75 either way. Olive, an AI colleague, joins partway through with scripted questions about approach and workflow, spoken and shown in text at once. A written debrief follows submission.

Olive and the voice exist to get the thinking out while it happens, not after the fact.

Six dimensions:

• Problem framing — Did the first move go after understanding, or straight for output?
• Evidence sourcing — Did they demand evidence for the claim that mattered?
• Delegation boundary — What did they keep, and what did they hand over?
• Working structure — Did anything exist between the brief and the answer?
• Output rejection — Was anything the assistant produced refused, and on what grounds?
• Verification — Was anything tested against the world, and did the result change something?

The assessment records the commentary, against the case’s own answer key.

Separately and represented for context: prose quality, prompt syntax, tool trivia, completion speed, tone of voice, accent, pace or hesitation, and AI use. Think-aloud audio becomes available for a reviewer. Never captured: a camera, the screen outside the task, keystroke rhythm, a face, a room, an identity document.

You are a credit analyst at a private credit fund. It has been asked to hold $110.0M of a senior term loan, with a second $20.0M line — the revolver — unused at close. 55 minutes, 75 at the hard cap, for a 600–900-word pre-read memo to the committee.

Seven documents sit in the packet: the tasking email, the audited accounts, two extracts from the seller’s pitch document — the CIM — the term sheet, the rate sheet, and a sector file. Six were built to carry one checkable wrong fact; the audited accounts plant nothing, since the other six get checked against them.

You serve as a software engineer at a parcel carrier. A tool prints each depot’s end-of-day delivery sheet. The ticket adds failed deliveries, the re-attempt fee — the fee for trying a delivery again — and refunds. The engineer who owns this left recently. 60 minutes, 75 at the hard cap: working code, plus a written note to the operations lead.

Six items sit in the packet: the ticket, a departed engineer’s handover notes, two emails, a pay schedule, and the code itself. Every one of them carries a planted claim. Nothing here checks against a clean copy — in code, any file can be wrong.

Six dimensions, each with what a reviewer finds and the excerpts from the reading.

Each finding below comes from the record. Each excerpt quotes the moment and the source. The candidate reads the same page an employer does, at the same time.

The single comparison this page allows is a property of the procedure. What an employer reads before anything else is a finding — the shape a reviewer wrote about a session. Reading a section and stopping tells you something about that session.