AI can make recruitment dramatically more efficient.
I work with a small manufacturing business that has been growing quickly. Like many SMEs, recruitment used to become a bottleneck whenever several vacancies opened at the same time. One person could realistically manage around three concurrent vacancies. After introducing AI into the recruitment process, that increased to around ten.
That is a significant productivity gain for a small company. But it also raises a more interesting question: just because AI can make more recruitment decisions, should we allow it to?
I don’t think we should.
What we learned from automating recruitment
With our current system, we upload a job description, and the software extracts the skills it believes are relevant. A human reviews them, decides which are mandatory and assigns different weights. When somebody applies, their CV is assessed against those criteria and receives a percentage score.
We spent a lot of time testing the system before deciding how to use that score. Below 65%, a candidate is rejected, although they receive a personalised explanation and are invited to reapply if they believe the assessment is wrong. Above 75%, they are shortlisted. Between the two, a human reviews the application and decides.
Once shortlisted, candidates receive a WhatsApp message from a bot. It may ask them to send something that was missing from the application or clarify a point about their experience. Nothing more than that. The decision on whether to take the candidate further remains with us.
It works, and the productivity improvement has been significant. But using the system has also made us think more deeply about what we want AI to do in recruitment, and what we don’t. If I were designing the next iteration today, I would take a different approach.
A percentage looks more objective than it is
Imagine two candidates receive scores of 76% and 64%. It is tempting to conclude that the first is simply better. But where did those numbers come from?
In the end, the score is only as good as the choices behind it. We decide which skills matter and how much they matter. Then the system has to recognise those skills in a CV, which is not always straightforward.
Two people can have very similar experience and describe it completely differently. And the person with the better CV isn’t necessarily the better candidate. A consistent process helps, but consistency and fairness aren’t the same. This is why I would design the system differently now.
Break the process into smaller jobs
Rather than asking one AI system to assess candidates from beginning to end, I would use several agents, each with a narrow responsibility.
The first would do something unrelated to recruitment: security. CVs are external documents entering an AI-enabled process, so an agent should first check for malicious content such as prompt injection.
A second agent would anonymise the CV, removing personal information and replacing it with a unique candidate reference. The objective is simple: assess what somebody can do before knowing who they are.
Deciding what we need from an employee should remain a human responsibility. The hiring manager should define the skills, decide which are essential and determine their weighting. AI can help us process information, but it shouldn’t quietly redefine the job we are recruiting for.
Only then would another agent assess the anonymised CV against those human-defined requirements. Then I would stop the AI.
Don’t just give me a score
I would also change the output.
A percentage is useful, but on its own it tells the recruiter remarkably little. I would want the assessment agent to provide the score, three positives, three negatives and, importantly, explain why the candidate didn’t score 100%.
That turns the score from a decision into information that can be challenged. Finally, a quality assurance agent would check that every previous agent has performed its task correctly before anything reaches the recruiter. The human would then decide what happens next.
Keep the human in the lead
There is a lot of discussion about keeping a “human in the loop” when deploying AI. In recruitment, I would go further. The human should be in the lead.
AI is exceptionally useful for repetitive recruitment work: processing applications, anonymising information, checking evidence, and applying the same criteria repeatedly. That can give a small business the recruitment capacity of a much larger organisation.
But efficiency isn’t the only objective.
We also need to know why somebody received a particular assessment, minimise the information that could introduce bias, protect candidates’ personal data and make sure somebody remains accountable for the final decision. The best recruitment automation knows exactly where to stop.
