Your job ad says "AI skills." So what?

AI skills demand has doubled in a year, but most job ads still don't know what they're asking for.

New Zealand employers are twice as likely to mention AI in a job ad as they were twelve months ago, but almost none of them say what they mean by it. That is not a criticism, it is a timing problem: the technology has moved faster than the language we use to hire for it, and most of us are writing role descriptions in the gap.

SEEK's June Employment Report, released on 24 July, is quietly one of the more useful documents to land on a hiring manager's desk this year. The headline is the recovery: job ad volumes rose 10.7% year on year, now sitting at their highest point in more than two years, with growth in every month since December 2024. Applications per ad have eased from their mid-2025 peak, which means candidate competition is loosening slightly, though the market is still tough if you are the one applying.

But the number worth sitting with is further down. Job ads referencing AI skills rose 3.9% month on month and are up 107.3% year on year.

That's a big statistic, but it's important to pay attention to scale: scale: only 3.5% of all job ads mention AI at all, and those are concentrated in ICT, marketing and communications, and consulting and strategy. This is a doubling off a small base, not a market-wide shift, and anyone telling you otherwise is not looking at the whole picture.

Also, if you break the growth down by term , references to agentic AI are up 134.4% year on year and generative AI is up 100.7%, but mentions of AI ethics and governance are up 330.1%.

Read that again, because it reframes the whole question. The fastest-growing AI skill in New Zealand job ads is judgement.

Our read is that these mentions are keyword counts, not a skills requirement. They tell us a small group of employers is out in front on this and starting to specify what they want. From the briefs crossing our desks, though, the term "AI skills" is still doing an enormous amount of heavy lifting for a set of very different things, and several of those things are not skills at all.

So it is worth asking the question properly. If you are writing a role description this quarter, what should "AI skills" actually mean for this role?

AI is a toolbox, not a tool

On a recent episode of our Find Your People podcast, Dr Michelle Dickinson made a point that reframes the whole problem. Most people saying "AI" mean generative AI, meaning ChatGPT or Claude or Gemini. But AI is a category, not a product. It includes machine learning, reinforcement learning (the thing shaping your Netflix homepage), optical recognition, and what she works in day to day: physical AI, teaching robots that the world is three-dimensional and that a fish is heavy, wet and inclined to wiggle.

Her framing was to treat AI as a toolbox rather than a tool. "There's a spanner and a hammer, and you find the right thing."

Which means if your job ad says "experience with AI tools", you have said close to nothing. You have told candidates you are aware AI exists. Every applicant who has opened ChatGPT once will now tick that box, and your screening has gained no information at all.

The fix is unglamorous. Name the task, not the technology. "Can use an LLM to draft and iterate first-pass client documentation" is a requirement. "Has evaluated vendor AI tools and can explain the tradeoffs" is a requirement.

The actual skill is problem framing

The second Find Your People conversation worth reading alongside the SEEK data is with Joe Davis, who works with leadership teams on what AI means in practice.

His point about capability is the one we would put in front of every hiring manager. You do not need to be a programmer to build a tool that does your work for you. What you need is the ability to understand the shape of the problem and to ask good questions about it. The technology will handle the code.

Which has an uncomfortable consequence for how we have historically assessed people:

"Being a spreadsheet wizard, like man, I'm good at Excel, I'm the guy who writes all the formulas and makes these beautiful spreadsheets, is no longer going to differentiate you."

For interviews, that means the useful question is not "which AI tools have you used". It is: tell me about a problem you solved recently. How did you work out what the actual problem was? What did you try that did not work?

That last part matters more than it looks. Davis is direct that judgement is mostly built through failure, and you just hope the lessons were cheap ones. A candidate who can walk you through a decision that went wrong and what they extracted from it is showing you something a certification cannot.

The skill is knowing when not to use AI

If problem framing is the first capability we should be looking for, judgement is the second, and it is the one most organisations are not screening for at all.

Davis borrows an analogy from Cory Doctorow. The good version is the centaur: human head and torso on the body of something much faster and stronger, with the technology providing the horsepower and the person providing the intellect. The bad version is the reverse centaur, where automation drives every decision and the human becomes the soft squishy bits doing the last-mile tasks the machine cannot manage yet.

His warning is that nobody is going to make that choice for us. It is not the natural outcome of the technology getting smarter. It is a leadership decision, made repeatedly, in small moments.

Practically, that means the person you want on your team is the one who can say: we could automate this, and we should not, and here is why. Because just because you can build something in ten minutes does not mean building it is consistent with how you treat your people, handle your data, or want to be experienced by your customers.

That is a values question dressed as a technical one, and it does not scale unless the whole team can make the call, not just the executive. Which is a hiring problem before it is a governance problem.

Have you actually used the thing?

Both guests landed on the same word, independently: play.

Davis is blunt that if you have not played with these tools, you have very little of value to contribute to a strategy conversation about them. You will offer "efficiency" and "cost saving" and the whiteboard session will end. As he put it: "We don't get strong by looking at weights, we get strong by lifting them."

Dickinson's version is more specific and, we think, better advice. Do not build your first thing for work, because the pressure to justify it will kill the learning. Build something for home. Her own example is an AI tool that handles her family's weekly meal planning and grocery list around two toddlers, a travelling job and a husband who will not eat beetroot.

She also offers the single most useful prompting tip we have heard: ask the tool how to ask it. Most people use ChatGPT like Google, which is close to the worst way to use it. Ask it a question, then tell it that was probably not the right way to ask, and get it to explain how you should have framed it.

For hiring, this gives you a clean signal - ask a candidate what they have built or automated for themselves, outside work, for no particular reason. The people who have done this have real capability. The people who have completed a course have a completion certificate. These are not the same thing, and the difference will show up in the first month.

The awkward part: screening is probably the weak link

Here is where Dickinson's assessment gets genuinely uncomfortable for our industry, and we would rather print it than dodge it.

Most CVs are now written with AI. Most CV screening software is built on AI. Which means, as she put it, we have "bots talking to bots in what essentially is a people wanting people decision."

The consequence is that we are almost certainly all filtering out strong candidates because the bot on their side and the bot on your side did not align on keywords.

If AI capability is genuinely a hiring priority for you this year, the most valuable work is probably not in your job ad at all. It is in reducing how much your process depends on keyword matching against a document that a language model wrote. Work samples. Scenario questions. Structured conversations about tasks rather than titles. Dickinson's reframe is worth stealing wholesale: instead of asking someone what they do, ask what tasks make up their work, and which of those are most powerful for them as a person.

And the risk we should all be watching

One number sits awkwardly next to the recovery. Even as job ads rise, employment among New Zealanders under 30 has fallen 1.4% over the past year.

That is worth taking seriously, because the work being automated first is disproportionately what Davis calls intellectual manual labour - opening emails, hunting through systems for the information needed to answer them, moving data from one spreadsheet into another. Which is also, historically, entry-level work.

The uncomfortable question for anyone doing workforce planning: if judgement is built through experience and cheap failures, and we automate away the roles where people used to accumulate both, where does the next generation of senior capability come from? Dickinson's talent pipeline warning and the digital divide she describes, where learning AI increasingly requires having the technology at home, both point the same direction.

We do not have a tidy answer. We would rather flag it as an open risk than pretend otherwise.

What we would do in the next quarter

Five things, all small enough to actually happen:

1. Rewrite one job ad. Delete the phrase "AI skills" and replace it with the specific task you want done.

2. Add one scenario question to your interview process. Not "have you used AI" but "walk me through a problem you framed badly at first, and how you worked out you had it wrong."

3. Give your team explicit play time in AI tools. Small, unmeasured, no ROI question attached. You cannot lead a conversation about tools nobody in the room has touched.

4. Audit your screening for keyword dependence. Pull twenty rejected applications from your last hire and read them properly. See what you missed.

5. Protect one entry-level role you were considering automating, and be deliberate about what judgement that person is meant to build.

One last thing

Davis offered a line we would hang above every AI steering committee in the country: never believe anyone who claims to be an expert in AI. You cannot be an expert in something changing this fast. We are all learning.

Which is, in the end, the answer to what "AI skills" means. It is not knowledge of a tool, because the tool will be different next quarter. It is the ability to keep learning when there is no end state to arrive at, and the judgement to know which of the things you could now build are things you should.

That has always been the hard part of hiring.
It just used to be easier to hide behind the software.

If this piece raised any questions for you, reach out to Tribe Group today. With 11 specialist teams, offices in Auckland and Wellington and recruiting nationwide, we can help you find your people.

Find Your People is Tribe Group's podcast on the future of work and where the human fits in all of it, hosted by Bruce Pilbrow. The conversations with Joe Davis and Dr Michelle Dickinson referenced here are available now.

Data source: SEEK New Zealand Employment Report, June 2026, released 24 July 2026. Under-30 employment figure via Stats NZ filled jobs data.