What Counts as Meaningful Human Involvement in UK Recruitment?

Summary: Under the post-DUAA framework, whether a hiring process counts as solely automated turns on whether human involvement is meaningful. The ICO's March 2026 recruitment report found that human review in AI-assisted screening was often perfunctory, with reviewers rubber-stamping outputs without access to the underlying reasoning. That finding, not the statutory text, is the practical standard employers now have to meet.

Why the question decides everything else

The Article 22A to 22D framework applies to decisions based solely on automated processing. If human involvement is meaningful, the decision is not solely automated, and the mandatory safeguards apply differently.

This makes the definition load-bearing. Almost every employer using AI screening asserts that a human is involved. The ICO's position is that most of those assertions do not survive examination.

The DUAA itself does not define the term, and the Explanatory Notes indicate it may be clarified later by secondary legislation. Until then, the ICO's draft guidance published on 31 March 2026 is the operative steer, and it is unusually concrete.

The three-part test

The ICO's standard is review by a competent person who has the authority, the information, and the time to overturn the tool's output, and who actually considers the individual case.

Authority. The reviewer must be able to change the outcome. A coordinator who can flag a case for someone else's attention has escalation rights, not authority. If overturning the model requires approval from a person who is not doing the review, the reviewer does not hold the authority the standard requires.

Information. The reviewer must be able to see why the tool produced its output. This is where most processes fail. A recruiter looking at a ranked list with scores but no drivers cannot form an independent view of the individual case. They can only accept or reject a number they do not understand, which is not consideration.

Time. A reviewer processing several hundred candidates in an afternoon is not considering individual cases regardless of their authority or information. Volume and review quality are in direct tension, and the ICO has not accepted volume as an excuse.

Plus the behavioural test: they must actually consider the case. The ICO is explicit that a consultant who only ever confirms what the machine suggested is not providing meaningful human involvement, so the process is still automated decision-making.

Two things that expressly do not count

Designing or building the system. Human judgment exercised when the model was configured is not human involvement in any individual decision the model later makes. This argument is made surprisingly often and does not work.

Rubber-stamping. Confirming the output is not reviewing it. The ICO's recruitment report found this pattern directly, and the letters sent to 16 organisations in March 2026 suggest the regulator is prepared to act on it.

The measurable signal most employers ignore

There is a straightforward diagnostic: what proportion of the tool's recommendations does your human review actually overturn?

If the answer is near zero across a meaningful volume, one of two things is true. Either the tool is performing so well that no correction is ever warranted, which is an extraordinary claim, or the review is confirmatory rather than substantive.

An override rate of zero is not evidence of a good model. It is evidence of a review process that cannot change anything, which is exactly what the ICO says does not qualify. Employers who track this figure have both a management signal and, if it is healthy, a piece of evidence.

Redesigning review so it qualifies

Two viable routes. The ICO frames them as a choice, and both are legitimate.

Route one: accept that the process is solely automated. Acknowledge it, and implement the Article 22C safeguards properly: advance transparency, a functioning right to human intervention, a right to make representations and contest, and meaningful information about the decision. This is often the more honest position for genuinely high-volume screening, and it is now permitted in ways it was not before February 2026.

Route two: build review that is genuinely meaningful. This means giving reviewers the drivers behind each output, the authority to overturn without escalation, and enough time per case to consider it. It usually means reviewing a subset properly rather than all cases nominally.

The failure mode is choosing neither: asserting meaningful review, not resourcing it, and therefore having neither a defensible human process nor the safeguards that would apply if you admitted the process is automated.

Practical steps

1. Measure your override rate by tool and by reviewer, over a defined period.
2. Audit what reviewers can actually see. Sit with one and watch. The gap between the documented process and the screen is usually large.
3. Time the review. Cases per hour tells you whether individual consideration is physically possible.
4. Check who can overturn without escalation, and whether that matches your process documentation.
5. Pick a route per tool and document the reasoning. Different tools in the same funnel can legitimately sit on different sides.

Frequently asked questions

Is there a required override rate? No. There is no threshold in law or guidance. The figure is a diagnostic for your own assessment, not a compliance test, but a rate at or near zero invites the question.

Can one reviewer provide meaningful involvement across thousands of applications? Unlikely at scale. The time element of the ICO's test is not waived by volume. High-volume employers are generally better served by acknowledging automation and implementing safeguards.

Does a human interview later in the process cure automated screening earlier? No. The rejected candidate never reaches the interview. Each decision point is assessed on its own terms.

Does this apply to agencies? Yes. A recruitment agency screening candidates with automated tools is subject to the same analysis, and the ICO's recruitment report drew on evidence from over 30 employers across the sector.

What if the tool vendor says a human is in the loop by design? Vendor design claims describe the product, not your implementation. The test applies to what your reviewers can actually do.