Pipeline
Years of AI Experience: What to Require Instead
Require zero years of AI experience. A year count is a proxy for accumulated exposure, and it only reports something once a tool has been stable long enough for exposure to accumulate, which tools a few years old are not. Ask instead for recency plus evidence: the candidate has used an assistant on the role's kind of work in the last six months, and can walk through a specific piece of it. If the requisition demands a number, attach it to the occupation rather than to the tool.
The takeAsking how many years of AI experience you can require draws ridicule, and the ridicule is fair and useless. Somebody still has to type a number into the field, and telling them the field is absurd leaves them typing three. The real objection is not that the tools are young. It is that a year count measures exposure, and exposure to something that changes every few months buys habits about as often as it buys skill. Recency answers the same question the field was asking and ages correctly.
Where Olive fits
Open a role and see what the work shows
A recency claim on a resume is still a claim. Olive turns it into an observation: one 50-to-70-minute session on the candidate's own clock, written up by a human reviewer as six evidence-anchored findings, with the candidate granted the identical report.
Rank your shortlistWhy a year count stops working on a young tool
Because a year count is not a measure of skill. It is a proxy for exposure, and it works when a tool has been stable long enough that more months with it reliably mean more competence at it. That condition holds for a general ledger system and does not hold here. Someone who has used an assistant heavily since 2023 has three years of practice at interfaces and constraints that have since moved.
The vocabulary itself is young enough to date. Indeed Hiring Lab counted distinct US job titles carrying AI or a related term in the employer's own wording: 264 in the first quarter of 2022, rising to 822 by the first quarter of 2026, with 63% of them now outside tech occupations 1. That counts titles rather than postings or hires, so it measures how far the words have spread and not how much demand sits behind them. Still, the words the requirement gets written in are only a few years old at this scale, which is thin ground for a five-year bar.
The measured pattern also runs against the assumption underneath a year count. In a staggered rollout of a generative AI assistant to 5,179 customer support agents at one software firm, issues resolved per hour rose 14% on average, with a 34% improvement among novice and low-skilled agents and close to no effect among experienced, highly skilled ones 2. That is one firm and one occupation, so read it narrowly. What travels is the direction: the people who gained most were the ones with least, which is the opposite of what a seniority proxy assumes.
What to ask for instead of years
Two things, both of which a candidate can answer in a sentence. First, recency: has used an assistant on work of this kind within the last six months. Second, specificity: can name one piece of that work and say what the assistant did, what came back wrong, and what they changed. Recency dates correctly on a tool that keeps moving, and specificity is the part that cannot be borrowed from a template.
Written into a posting, that reads as a duty rather than a filter. For an analyst role: drafts the monthly variance commentary with an assistant, reconciles every figure against the source system, and can say which parts of the draft were discarded. For a paralegal role: uses an assistant for first-pass document review and can explain how the cited authority gets checked.
Notice what those lines do not ask for. No tool name, no year count, no certificate. They describe an artifact and the person's relationship to it, which is a thing a hiring conversation can actually get at. If you want the underlying vocabulary sorted out first, what counts as an AI skill is the argument beneath this one.
How to fill the field when the system demands a number
Attach the number to the occupation and leave the tool out of it. Most requisition templates want a years-of-experience value per requirement, and the honest answer is that the occupation has a defensible number while the tool does not. Four years in claims adjudication is a real bar with real reasoning behind it. Four years of AI experience is a bar you cannot explain and did not derive from anything.
So the field gets the occupation, and the AI moves into the description of the work attached to it. The requirement reads: four years adjudicating auto claims, including recent work drafting and checking adjudication summaries with an assistant. One number, one recency clause, and nothing anyone has to invent a history for.
Watch what the system does with that field downstream. Plenty of requisition tools turn a years value into a knockout question on the application form, so a number typed to satisfy a template quietly becomes an automatic rejection rule nobody chose. A rule that cuts a group protected by Title VII at a higher rate has to be defended as job related for the position in question and consistent with business necessity 5, and a number nobody derived from the work will not carry that. If the field cannot be left blank and cannot be un-wired from the knockout, put the lowest defensible occupational number in it and carry the real bar in the duties.
That also survives the follow-up question better. A rejected applicant asking where the AI bar came from gets a real answer about the tasks, rather than an answer about a template. The same decision in the other column, required or preferred, turns on exactly this: whether you can say where the number came from.
Test the recency claim once, or nothing has changed
A recency line on a resume is still a line on a resume. If the process never looks at it, you have swapped one unverified claim for a fresher-sounding unverified claim, and the requirement has cost you applicants for nothing. Give one stage of the loop the job of looking: ten minutes on a phone screen, or a short piece of the work with the assistant open, whichever the role can carry.
Self-report is the weakest place to spend that stage. In a study of 288 teachers who took both a self-report and a knowledge-based test of AI literacy built on the same framework, correlations between the objective and self-reported factors ran from r = 0.07 to r = 0.24, with profiles for people who overrated themselves and people who underrated themselves 3. Those are teachers in Taiwan rather than job candidates, so it is not a hiring finding. It does say the two instruments measure different things.
The same gap shows up in practitioners with years of the work behind them. Sixteen experienced open-source developers working on repositories they had known for about five years forecast that AI tooling would cut their completion time by 24%, and still believed it had helped by 20% afterwards, while the measured result was 19% slower 4. Small sample, one setting, and it does not transfer to juniors or unfamiliar code. Keep the gap rather than the magnitude: people can be wrong about their own AI productivity in the direction, not only in the size.
Ten minutes is usually enough to see whether a recency claim is real. Verifying an AI claim on a resume in ten minutes is the version of that stage for teams with no assessment budget, and checking AI-proficient claims more generally covers the pile-level version.
Common questions
Is asking for three years of AI experience illegal?
No US statute names it as an off-limits qualification as of August 2026. The exposure is indirect: a year count on a tool of this age can filter by when someone entered a particular kind of work. A qualification that disproportionately screens out a group protected by Title VII has to be shown job related for the position in question and consistent with business necessity 5, and age claims run under the ADEA, which sets its own standard. The practical risk is that almost nobody derived the number from the work, so there is nothing to point at when asked. Have counsel read the final wording.
What if candidates have genuinely been using these tools since 2021?
Some have, in research groups, machine learning teams and early product work. That history is real and worth hearing about in an interview. It is still a poor screening field, because the population who can honestly claim it is small, skews to a few employers, and is not the population most roles are hiring from. Ask for it in conversation where it can be explored, rather than in a requisition field where it becomes a gate nobody meant to build.
Does six months of recency work for every role?
It is a starting default, not a rule. Pick the window from how fast the work itself changes. A role that touches an assistant weekly justifies a short window, because the habits that matter turn over quickly. A role where AI touches one quarterly deliverable justifies a longer one, or none at all. State the window in the posting so a candidate who paused for parental leave or a contract gap knows where they stand.
How do I compare two candidates when neither has a year count?
Compare the work, not the tenure. Give both the same task and read what came back, or ask both the same question about a specific piece of their own recent work and listen for what they threw away. Comparison needs a common measure taken the same way, which a year count never was: two people writing three years meant entirely different things by it, and nothing in the process ever asked.
Should the posting say the tools are new and no experience is needed?
Say what the work involves and let candidates draw their own conclusion. A line announcing that no experience is needed reads as reassurance and tends to invite applications from people who read nothing else. A duties line saying the person drafts with an assistant and is accountable for what ships tells a beginner they can apply and tells an expert what the standard is, which is the same information without the coaxing.
References
- 1. AI Is No Longer Just a Tech Occupation Story: It's Spreading Across Job Titles in the US and Europe hiringlab.indeed.com Supports the count of distinct US job titles carrying AI terms and the share of them sitting outside tech occupations.
- 2. Generative AI at Work (NBER Working Paper 31161) nber.org Supports the claim that measured gains concentrated in novice workers and were close to nil for experienced ones, against the assumption behind a seniority proxy.
- 3. How to Assess AI Literacy: Misalignment Between Self-Reported and Objective-Based Measures arxiv.org Supports the claim that a self-rating of AI skill and a demonstrated measure of it do not track each other.
- 4. Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity (arXiv:2507.09089) arxiv.org Supports the gap between what practitioners believed about their own AI speedup and what was measured.
- 5. 42 U.S.C. 2000e-2(k) - Burden of proof in disparate impact cases uscode.house.gov Supports the job-related and consistent-with-business-necessity standard, which is the Title VII test and covers race, color, religion, sex and national origin only.
5 sources, numbered by first appearance. How Olive sources claims
General guidance for hiring teams. What works at one company and one volume may not transfer to yours.
Olive assesses how a person works with AI. It does not detect AI-written documents, and it never produces a score, a ranking, or a match percentage for a person. Candidates read the same report the employer reads.