Onrec Explores / AI & Automation

AI agents are entering recruitment. Where should human judgement begin?

The promise is less administration and more time with people. The real test is whether automated hiring workflows improve decisions—not simply accelerate them.

Onrec Explores · 17 September 2026 · Analysis · 5 minute read

A recruiter opens a vacancy, gives an AI tool a brief and receives a suggested shortlist. The same system may also prepare outreach or help organise the next step. The attraction is obvious: fewer hours moving information between screens, more time speaking to people.

But each step raises a different question. Drafting a message is one thing. Deciding who gets to receive it is another. As recruitment tools take on more of a workflow, employers need to decide which actions can happen automatically and which require a person to stop, check and take responsibility.

Why the conversation has moved beyond writing job adverts

LinkedIn’s January 2026 research reported that 93% of recruiters surveyed planned to increase their use of AI during the year, while 66% said finding qualified talent had become harder. The underlying research included 6,554 HR professionals across multiple markets, including the UK; these are international findings, not UK-only figures. Read LinkedIn’s research and methodology.

The same announcement describes AI agents for finding, shortlisting and contacting candidates. That is a useful indication of the direction of recruitment products, although a supplier’s product claims should not be treated as independent evidence that every employer will achieve the same results.

For this discussion, an AI agent means a system that can carry out a sequence of tasks towards a goal, within the access and permissions it has been given. The practical distinction is between asking a tool to suggest something and allowing it to act.

Speed is easier to measure than a missed opportunity

Time spent screening is visible. A qualified applicant who never reaches the shortlist may be much harder to spot. That is why a faster process, by itself, is an incomplete measure of success.

In a May 2026 analysis, ILO senior economist Janine Berg argues that recruitment systems need clear objectives, suitable data and meaningful oversight. Her discussion includes a system that attempted to assess a growth mindset through candidates’ use of particular words—an example of how an apparently sophisticated assessment can measure a poor substitute for the quality an employer actually wants. Read the ILO analysis.

Our view is that employers should start by defining what a good decision looks like for the role. A confident explanation from a tool is not a replacement for evidence that its recommendations are useful.

Give the tool a task—and a boundary

A sensible pilot might let a tool draft interview invitations for a recruiter to approve. A more consequential pilot might generate a shortlist, but require reviewers to compare it with a sample of applications the system did not select.

Those are different levels of responsibility. Treating them as separate decisions makes it easier to choose a useful starting point and to understand what additional checks would be needed before expanding the tool’s role.

Before a pilot, agree who can change the criteria, who reviews errors and what would trigger a pause. Give the reviewer enough time and information to challenge the output. A person clicking “approve” on an unexplained recommendation is a weak checkpoint.

Five questions to ask before expanding a pilot

Be careful with the bigger jobs narrative

Automation within a hiring team should not be confused with proof that AI is driving the wider recruitment slowdown. In June 2026, LinkedIn said its EU and UK analysis had not found evidence of a broad AI-driven fall in hiring at that point, pointing instead to economic conditions and business uncertainty. That is a finding from its data, not a guarantee about future employment. Read the EU and UK analysis.

For an individual recruitment team, the more useful question is specific: what happens to the time you recover? If it goes into better briefings, more thoughtful candidate conversations and clearer feedback, the benefit is easier to explain than a claim about processing more applications.

The next step should be evidence, not a bigger promise

AI agents deserve a serious trial, but a trial needs a test that the system could fail. Start with a defined task, record the result and make room for recruiters and candidates to explain what the numbers miss.

The strongest case for automation will come from teams that can show both the time saved and the quality protected. That is the conversation Onrec wants to bring together: practical accounts of what worked, what needed changing and what should still be done by a person.

This feature combines linked research with Onrec analysis and suggested questions for teams evaluating recruitment technology. No sponsor funded this article. Industry contributions are invited below.

Please include your name, role, company and any relevant commercial connections. We’ll contact you before publishing your contribution.