How would you know if a candidate can use AI?

I've been on both sides of hiring this year. For most of my career I've been the one reading CVs and deciding who gets to progress through a recruitment process, and since March I've been the one sending them. From the candidate side, it's hard not to notice AI creeping into job adverts. Some of the roles that mention it most are at tech start-ups wanting a particular kind of AI recruiting background I don't have, so I've ruled myself out of those. But it's showing up in the talent acquisition roles I do go for too, from a ratings and research firm asking for experience of implementing AI tools in recruitment to a law firm role where AI came up in my conversations about it. I haven't yet seen it mentioned in a single People and Culture role. That's got me thinking about it as a recruiter rather than a candidate. When someone says they can use AI, how would you know?

Most of what people know about AI right now, they've taught themselves. There are plenty of courses and certificates, and I've done a few myself, but there's no recognised standard yet that tells an employer what someone can do in practice, and no job title that reliably does either, so we're back to the problem recruiters have always had with self-taught skills. I've worked with recruiters who taught themselves Boolean search. When a CV said "Boolean", I'd give the candidate a real brief and ask how they'd go about finding that person. You learn far more from how someone talks you through a search than from the word on the page, and I think the same test works for AI.

Paper can undersell people as easily as it oversells them, which is why I'd always rather test the skill. One of the best hires I was involved in at my last firm was a lawyer who didn't fit the profile the firm usually hired. He came from a smaller, less well-known firm and his CV undersold him. I screened him, thought he was right for the team and put him in front of the partners, who tested his drafting. He did really well and he's still there. If we'd gone on the CV alone, he wouldn't have got through the door.

So what does my own evidence look like? My starting point was Copilot at my last employer. It was rolled out gradually, with some people getting it first and everyone else applying for a licence, and I was one of the first in my team to go to a training session, which gave me a head start. I worked out what it could do and then helped colleagues informally, talking through what they wanted out of it and what might work better. Most of the team could see it helping with an email or tidying an inbox, but asking it to help design a gender equity programme for EMEA felt like a step too far. A lot of our work depended on human conversations, like sitting down with someone to talk about racism in an office, and AI can't have those for you. That's what held the team back from using it more. What it could do was help with the thinking around that work. It helped me plan and structure our EMEA affinity groups for gender and LGBTQ+, for example, and by the time I left I was trying to get the team looking at how we might use agents. That was a harder conversation, because it was a long way from how most people were using it day to day.

Since March I've carried on experimenting, trying out different tools to see where each one fits and whether it earns a place in my day-to-day work. Along the way I've paid for Claude, ChatGPT, Perplexity, HeyGen, Apollo, Lovable, n8n and Apify, and I've dropped some of them. HeyGen went because I wasn't making avatars, and Perplexity went because its deep research kept burning through credits and the other tools had caught up with what I was using it for. Claude has stayed because it's part of my working day, from tailoring CVs to building the job feed below, and ChatGPT has stayed too, as a support tool alongside Claude and for the things it does better, like creating images. Knowing what to stop paying for has turned out to be as useful as knowing what to keep.

The thing I'm proudest of building is my job search feed. Looking for a job in 2026 means vacancies coming at you from LinkedIn, job boards and recruitment agencies across more than one country, and a lot of what LinkedIn was sending me was already out of date.

I'd been using n8n since an Ivee course on vibe coding in the spring, for things like form completion and routing emails, so last month I started building a feed there with ChatGPT. The build got slow and long-winded, so I moved it to Claude Code. It now:

• pulls roles from the job boards and recruitment agencies I care about, from Bristol, Cardiff and London to Dublin, Zurich and Barcelona
• scores each one against criteria I defined around location and what makes a role suitable
• sends the best one or two to my phone each day

Because I'm flexible on location and the type of role, it covers far more ground than I could search properly by hand. Not everything it finds is suitable, but it's turned up roles I hadn't seen anywhere else, and every application now sits in one tracker.

None of that involved me writing code. I'm not a technologist and I don't pretend to be, but it's a long way past "help me write an email", and that gap is what I'd want a recruiter to probe.

So if I were hiring now, I'd want to see some evidence first, whether that's the Featured section on LinkedIn, a personal website or a GitHub page. Then I'd ask three things.

1. For recruiter roles, how would you run this search? I'd give them a hard-to-fill profile and ask them to talk me through it with whatever AI tool they use: how they'd break the brief down, where they'd look, what search strings they'd write and what they'd still check themselves. I'd be judging the thinking, not whether they'd found anyone in twenty minutes, because without an ATS or a proper sourcing tool nobody could.
2. What have you stopped using, and why? You usually only drop a tool once you've used it enough to know it isn't worth it.
3. What would you never hand to AI? My own answer is candidates' CVs.

That last one is where the rules come in. Putting CVs into a tool the organisation hasn't approved, with no idea where that information ends up, is a GDPR problem before you've even started. Using AI to screen, rank or decide on candidates is a separate question, and under the EU AI Act that kind of use is classed as high-risk, with the obligations that come with it now due from December 2027. Anyone using AI in hiring should know which of the two they're dealing with. I'm cautious about screening tools anyway, because if you feed CVs in and can't see what the scoring is based on, you need a person checking what comes out.

There's another side to this that I'm acutely aware of, perhaps partly because of my background in inclusion and culture. The basic paid tools are relatively cheap, and open-source models will bring the cost down further, but access is about more than price. A lot of people never get past using AI for novelty, like making pictures of themselves in the '80s, often because nobody has shown them what else it can do. Inside a firm, there's training, there are colleagues to swap notes with, and the firm pays for the licence. The moment you leave, you lose all of that. Serious use costs more too, and Claude Max alone is $100 a month for me. I'm happy to pay it because I think the frontier models are worth it, but not everyone can, and free tools often mean trading your data for the access. If recruiters start screening for AI skills, they risk screening for who could afford to keep experimenting, and that doesn't seem equitable to me.

If you're hiring, ask candidates to show you what they've built and what they've dropped. And before AI skills become another line on the person spec, think about who's had the chance to build them.