The HR calls that always need a human in the room

Why these four, specifically

Plenty of HR tasks are perfectly suited to automation or even AI assistance. These four are different. Each one can end someone's job, or come close to it, and each one is the kind of decision a tribunal or a regulator will look at closely if it is challenged. That combination, real consequence and real scrutiny, is exactly where a tool that produces a plausible answer rather than a correct, explainable one becomes a genuine liability.

It is also worth being plain about the exposure involved. Discrimination claims, unlike some other tribunal awards, carry no statutory cap on compensation. A scoring system that quietly reproduces a pattern of bias is not a small process error if it ends up in front of a tribunal, it is the kind of claim that has no ceiling on what it can cost.

Recruitment scoring

There is a real difference between checking whether a CV meets a fixed criterion, a qualification, a number of years' experience, a professional certification, and asking a tool to judge a candidate's overall fit. The first is a fact with a correct answer. The second is a judgement, and letting a model make it quietly turns an opinion into something that looks like a measurement.

The practical fix is not to avoid software here, it is to keep the software doing the first kind of task and keep a person doing the second. A tool that flags "does not meet the stated criteria" for a human to check is doing something useful. A tool that outputs "72% match" for a role with no defined criteria behind that number is doing something that looks precise and is not.

Performance ratings

A rating built purely from monitoring data, output logged, hours tracked, response times measured, misses everything that gives a rating context: illness, a difficult quarter, a team going through change. A human manager weighs that context, imperfectly but genuinely. A model trained only on the numbers cannot, because the context was never in the data it saw.

This is where the workplace monitoring consultation covered earlier in this series and this piece meet directly. A monitoring tool that only logs activity is one thing. The moment its output feeds into a performance rating with real consequences, promotion, a warning, a rating tied to pay, it has crossed from measurement into decision-making, and the same human-judgement standard applies.

Disciplinary outcomes

Natural justice in a disciplinary process depends on someone actually hearing an explanation and being able to change their mind because of it. If an automated tool recommends an outcome and the human sign-off becomes routine, a rubber stamp rather than a real decision, the human is not really deciding anything. The ICO's own recruitment guidance names this problem directly: human involvement has to be genuine, not a formality dressed up as oversight.

The same logic applies just as squarely outside recruitment. A system that flags an attendance pattern as grounds for a disciplinary hearing is doing useful, rule-based work. A system whose flag becomes the outcome, with the hearing itself reduced to confirming what the system already decided, has quietly removed the one thing a disciplinary process is legally supposed to guarantee: a genuine hearing.

A human who cannot change the outcome is not really the decision-maker.

Redundancy selection

Scoring matrices for redundancy selection are ordinary and lawful when a person builds and applies them with clear, defensible criteria. Automating the scoring itself raises the same questions as automated recruitment scoring, and the same risk: criteria that quietly correlate with age, disability, or another protected characteristic, scored by a system nobody in the room can fully explain.

It is worth naming the specific trap here: a criterion like "flexibility" or "adaptability to change", innocuous on its face, can correlate with age or caring responsibilities once it is scored at scale by a system optimising for a pattern in past data rather than for fairness in this specific selection round.

What a human in the room actually requires

It is not enough that a person's name is on the decision. The person needs real authority and enough information to genuinely change the outcome, and enough time to actually use it, not five seconds to approve a screen before the next one loads. That is the practical test worth applying to any of these four decisions in your own business: could the human who signed off actually have said no.

Where the record does the proving

Jamie HR keeps a clear record of who made a decision, on what basis, and when, which matters here for a specific reason: if one of these four decisions is ever challenged, you need to be able to show a human genuinely decided it, not just that a human was technically present.