Pipeline
An AI Skill Is a Decision, Not a Tool You Have Open
An AI skill is a repeatable decision, not a tool you have open. Four of them carry almost all of it: what work you hand over, how much context you give it, what you check before the output leaves your desk, and what you refuse to delegate at all. Naming the tools separates nobody, because the words are free to add and nothing in an application can contradict them. Write the requirement as decisions and the tool names become optional detail.
The takeThe annual AI skills list is the wrong artifact, and its publishers half know it. Tools turn over every quarter, so the list has to be rewritten to stay current, and the rewrite is the product. What the work asks of a person has barely moved in three years: frame the problem, give the assistant enough to go on, check the part that would embarrass you, keep the call that is yours. Write that down once and it survives the next model release.
Where Olive fits
Open a role and see what the work shows
Olive assesses those decisions directly: a 40 to 60 minute occupational assignment done with an AI assistant that will take the whole thing over if nobody stops it, written up as six findings with the timestamped moment behind each one. The candidate is granted the same report, free, on every tier.
Rank your shortlistWhat separates a skill from a tool you have open?
A skill has a failing version you could watch. Opening an assistant does not: everyone can do it, nobody can do it badly, and it produces no evidence about the person. The decisions around it do fail visibly. Handing over work that needed a specialist fails. Sending an output nobody checked fails. Those failures are what a hiring process can actually see, so those are the skills.
Run the test on a real posting and most of the AI line disappears. "Experience with ChatGPT, Claude and Copilot" survives contact with any candidate who has spent an afternoon in a free tier. "Prompt engineering skills" is closer to real but still names a technique rather than a call. "Knows which parts of a draft still need a source" is the first phrase in that list that a competent person could fail in front of you.
Tools and decisions age at different rates, which is what makes the distinction practical. A tool list written eighteen months ago names products that have been renamed, merged or replaced, and a requirement anchored to it now excludes people who are current. The decisions have not changed at all. Nothing about deciding what to verify was different two model releases ago, and nothing about it will be different two releases from now.
Put that distinction against a resume and the resume loses. A tool name is a claim the document itself cannot test, so an application tells you which products a candidate has heard of and stops there.
Why the tool list stopped filtering anyone
Because AI vocabulary went mainstream faster than AI practice did, and the vocabulary is what a list captures. Employers price the language into what they advertise: postings mentioning AI skills carry salaries about 28% higher than postings that do not, roughly $18,000 more a year, and in 2024 over half of the postings requesting AI skills sat outside IT and computer science 1. That is a posting-level figure about what gets advertised, not about what anyone was paid.
The same spread shows up in role names. Indeed counted distinct US job titles carrying AI language and found them rising from 264 in the first quarter of 2022 to 822 in the first quarter of 2026, with 63% of them now outside tech occupations 2. A title counts once whether five postings carry it or fifty thousand, so the number tracks how far the word has travelled across role names and says nothing about volume of demand. The travel is the point: the language arrived everywhere before the practice did, so naming the language filters on who reads industry press.
What has actually changed inside the work is more specific and more useful. Indeed's index rated almost 2,900 skills found in US postings and put 40% in a hybrid category, meaning the work changes shape but human oversight stays necessary, against 1% rated as fully transformed 3. Large language models produced those ratings by scoring skill descriptions, so the number measures judged capability and reads as an upper bound. Even as an upper bound, the shape is clear. Oversight is the part that survives, and oversight is a decision.
A requirement built on tool names is therefore filtering on the one attribute that spread fastest and means least. Naming behaviors instead of tools is the same argument at the level of individual phrasing.
Name the four decisions the work asks for
Four questions cover nearly all of it, and each one has a wrong answer you can point at. What goes to the model and what does not. How much context the model gets before it starts. What gets checked, against what, before the work carries someone's name. And which calls stay with a person no matter how good the draft looks.
The fourth is the one most lists omit and most managers care about most. It is the refusal: the moment someone declines to hand over a judgment they own, or throws out a plausible output because it was answering a question nobody asked. In a randomized trial with 640 Kenyan business owners given an AI assistant, there was no detectable average effect on business performance, but high performers gained just over 15% while low performers did about 8% worse, and the researchers attribute the split to how owners selected from and implemented the advice rather than to differences in the advice itself 4. Small businesses over chat are not office knowledge work and the average effect could not be distinguished from zero, so leave the percentages with the study. What survives is the demonstration: identical assistance produced opposite results depending on the judgment applied to it.
Written out for a specific role, the four questions stop being abstract. For an analyst: which parts of the model get built by hand, what the assistant is told about the business, which figures get traced to source, and which recommendation is never delegated. For a recruiter: which outreach gets drafted, what the assistant knows about the role, which claims about a candidate get verified, and which rejection is written by a person.
Rewrite one requirement this week
Take the AI line out of one live posting and rewrite it as a decision with a failing version. It takes about twenty minutes. The test to apply on the way out is simple: could a hiring manager describe someone doing this badly, and would two managers recognise the same failure. If not, the phrase is decoration and it belongs in the onboarding plan instead.
A worked pass on a single line:
- Before: "Comfortable using AI tools to improve productivity."
- After: "Uses an assistant for first drafts and can say which parts of a draft still need a source before it goes to a client."
- Before: "Prompt engineering experience preferred."
- After: "Gives an assistant enough context about the account to produce something usable, and edits rather than sends."
The rewrite is longer, which is the usual objection, and it is the wrong objection. A posting has room for one long sentence that does work and no room at all for four short ones that do none. The phrasing also has to survive a legal read, and how to write an AI-skills requirement that is not legally vague covers what a posting can commit to safely.
One thing to settle before the rewrite: what the assistant is actually doing in this role today. A requirement can easily describe AI work that nobody in the team performs, and the fastest correction is to ask the two people doing the job. How to find out what AI actually does in the role is that conversation. The three terms people reach for while writing these lines pull apart in literacy, fluency and prompt engineering.
Common questions
Is prompt engineering an AI skill or not?
It is a technique inside one of the four decisions, specifically how much context the model gets before it starts. That makes it real and worth teaching, and a weak thing to select on, because it is learned in a fortnight by anyone who already understands the work. Treat it the way you treat spreadsheet formulas: expected, trainable, and not the reason you hired someone.
Do AI skills mean something different in a non-technical role?
The decisions are the same; the content changes. A lawyer decides which research gets traced to a primary source, a marketer decides which claims about a product can be drafted and which need sign-off, a recruiter decides which candidate claims get verified. The tool question would give you four different answers and none of them useful. The decision question gives you one frame that transfers, which is why it works for roles where nobody has written an AI competency yet.
How do you check for this before an interview?
Not from the resume, and not from a tool list on the application. What works is a short piece of work with something wrong in it: a task where an assistant will produce a plausible draft containing an error the candidate should catch. The signal is whether the error survived, and that is visible in the output without asking anyone to describe their process.
Should the requirement say required or preferred?
Depends on whether the role can be performed without it today. A behavior nobody in the team currently performs cannot be required, because you have no way to evaluate it and no defensible reason to exclude anyone for lacking it. If the work genuinely depends on the decision, make it required and assess it. If it is aspiration, mark it preferred and mean it, because required reads as a condition of eligibility.
Does a certificate count as evidence of an AI skill?
It is evidence about vocabulary and one vendor's product line, which is worth something at onboarding and little at selection. An exam records which answers a person could recognise under time. What a candidate refused to accept from an assistant never appears in it. Read a certificate as a signal that onboarding will be shorter rather than as a signal that the judgment is present.
What if candidates just describe the decisions back to you?
They will, and well-prepared candidates describe them beautifully. Describing a verification habit and having one are different, and the gap is only visible in work rather than in conversation. Any question that can be answered from a blog post about AI collaboration will be, so put the signal somewhere a description cannot reach: an actual task, with an actual wrong answer in it.
References
- 1. Beyond the Buzz: Developing the AI Skills Employers Actually Need lightcast.io Supports the advertised salary gap on AI-mentioning postings and the claim that most postings asking for AI skills sit outside IT.
- 2. 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 AI-touched job titles and the share of them sitting outside tech occupations.
- 3. AI at Work Report 2025: How GenAI is Rewiring the DNA of Jobs hiringlab.indeed.com Supports the split of rated skills across transformation categories and the point that hybrid means human oversight stays necessary.
- 4. The Uneven Impact of Generative AI on Entrepreneurial Performance escholarship.org Supports the claim that identical AI assistance produced opposite results depending on which advice the user chose to act on.
4 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.