AI recruiting tools are good at the part of hiring nobody enjoyed: sorting high volumes, matching CVs against keyword lists, drafting adverts and scheduling. They are not good at the part that actually decides a hire: spotting the person who can diagnose a problem that is not in the ticket, who will own a decision when ownership is unclear, and who will still be useful in twelve months when the tool stack has changed. Rely on them for the first part and you will miss the second.
Who this is for
- Recruiters and hiring managers in UK tech who are being sold an AI screening product.
- Candidates trying to understand why a process that claims to be smarter still feels like a keyword lottery.
- Leaders who have just inherited a hiring funnel and need to decide where the automation helps and where it hides risk.
What the tools actually do
Most AI recruiting products are a speed layer over the old keyword filter. They parse CVs, score them against a job description, generate outreach messages, and sometimes produce interview questions. A few go further: voice screening, coding tests, sentiment analysis. The common thread is that they handle volume and pattern matching well.
That is genuinely useful when you have 300 CVs for a generic role and most of them are clearly off-target. It is less useful when you have forty CVs for a specialised role and the best candidate is the one whose previous title does not match your JD.
Where the noise comes from
The first source is the CV-to-JD match. The tool rewards the candidate who used the exact words in your advert. The candidate who rebuilt a small firm's identity system but called themselves an "IT support engineer" scores lower than the candidate who copied your tech stack into their personal summary. The more generic the role, the worse this gets.
The second source is flattened context. A candidate who spent two years untangling a legacy integration looks less relevant than a candidate who listed three tools your team uses, even if the first role required far more judgement. The tool cannot weigh what the person actually did, only what they wrote down.
The third source is false coverage. Having 50 AI-screened candidates does not mean you have considered the market. It means you have considered the market that fits your filter. The strongest people are often outside that shape.
What to do instead
Write the brief around outcomes, not tool brands. "Reduce repeat tickets by diagnosing root cause" is a better filter than "three years of ServiceNow". Tool experience can be trained; diagnostic judgement is harder to teach.
Use AI for the parts that do not decide the hire. Let it draft the advert, summarise CVs and send rejection emails. Do not let it rank the shortlist. A human who has done the job should hold a ten-minute screen before any technical assessment.
Source deliberately from adjacent roles. The best cloud support hire may currently sit in internal IT. The best detection engineer may come from a sysadmin background. The best identity engineer may have a title that does not contain the word identity. If your only source is inbound CVs scored by an algorithm, you are fishing in a small pond.
Finally, audit the output. Run a monthly check: of the people you hired in the last quarter, how many would have made it through the AI filter if you had applied it before the human screen? If the answer is "not many", your filter is costing you talent.
When automation is the right call
If you are hiring at genuine volume for a genuinely standard role, automation helps. The danger is using the same tool for roles where judgement is the actual product. Most technical hires fall into the second category.
Where this connects on POST
The Wrong Pool essay covers what happens when a JD defines a candidate who does not exist. The rest of the Hiring Perspectives series looks at the seven failure modes that keep technical roles open despite strong candidate supply. For candidates, the expectation compression piece explains why the job market can feel crowded and empty at the same time.