Across most of the business, the question about AI is how much of the work you can redesign. In the people function, the more important question is the opposite one and getting it wrong is now a legal risk as well as a human one.

Ask most people where AI is headed in HR and you will hear the same answer: it will screen the CVs and rank the candidates. That is the loudest story in the market. It is also the one I would be most careful about.

Careful about two reasons. The first is the argument this series has made throughout: pointing AI at an existing task, rather than redesigning the work, is the low-value move. Automated CV screening mostly speeds up a process that was already flawed - it inherits whatever bias and blunt filtering the old process had and runs it faster.

The second reason is newer. From 2 August 2026, the EU AI Act treats AI used for recruitment, selection and worker management as high-risk by default. Systems that screen or rank candidates, or profile employees, carry hard obligations: human oversight, explainability, candidate notification, documentation, and the ability for a person to override the machine. In other words, the regulators have reached a conclusion the industry is still arguing about - that some decisions about people should not be handed to a machine unsupervised.

So, the interesting question in HR is not “how much can we automate?” It is “what must we refuse to hand over?” That inversion, I would argue, is the whole strategy.

In every other function the question is how much you can redesign. In HR, the more important question is what you refuse to hand over.

The real prize is quieter than the headline

The valuable transformation in HR is not glamorous, and it has nothing to do with hiring decisions. It is about the administrative connective tissue that has quietly eaten the function for years.

Consider onboarding. A single new joiner touches HR, IT, payroll, facilities and a hiring manager. Today someone is the human glue between those systems - chasing provisioning, uploading forms, tracking completion, sending reminders. It is slow, error-prone, and a poor first impression.

Redesigned, an onboarding agent recognises a signed offer, provisions the accounts, sets up the payroll record, schedules the induction and briefs the manager - coordinating across every system and escalating only what stalls. It is the same sequential-to-parallel shift we described earlier in this series, and of all the HR use cases it consistently produces the highest satisfaction.

The same holds for policy questions, leave requests, benefits changes and first-line case management - high-volume, low-judgement work that keeps HR from the work only HR can do. Take that load away and you do not shrink the function. You give it back the hours it has been asking for.

Draw the line, and make it explicit

The discipline, then, is a clear line between what an agent may do and what it may not. The teams doing this well tend to define three tiers, and every HR function should write down its own version.

• Auto-execute: low-risk, reversible actions - answering a policy question, booking leave within entitlement, sending a reminder.

• Escalate: anything needing human judgement - flagging a possible flight risk, proposing a shortlist, suggesting a development move - where the agent assembles the context and a person decides.

• Block: the decisions that must stay human, full stop - who is hired, promoted, rated or exited, and anything touching a grievance, a disciplinary matter or someone’s wellbeing.

The principle underneath is simple: the agent handles the logistics; the human handles the relationship. An agent can compile everything a manager needs for a difficult conversation. It should never be the one to have it.

Restraint is not caution. It is the strategy

It is tempting to read all this as a brake - the risky, regulated function that has to hold AI at arm’s length. I would argue the opposite.

The scarce asset in every people process is trust. Employees have to trust that a decision about their career was made by a person who was accountable for it. Candidates have to trust that they were treated fairly. Regulators now require you to demonstrate it.

An HR function that automates its admin ruthlessly and its judgement not at all is the one that earns that trust - and, not by coincidence, the one that captures the most durable value. The organisation racing to automate screening is buying a short-term efficiency and a long-term liability. Restraint here is not the absence of ambition. It is what ambition looks like when the subject is people.

HR’s real job in the agentic era

There is a larger point, and it is the one I would leave a people leader with.

As agents spread across operations, finance and commercial, someone has to own the human side of that shift - how roles change, how work is redesigned without breaking the people doing it, how a workforce of humans and agents is actually managed, how people build the skills the new work needs. That is not a technology problem. It is an HR one.

So, the most important thing AI does in HR may not be anything it does inside HR at all. It is that it hands the people function the clearest mandate it has had in years: to steward how the whole organisation learns to work alongside machines - humanely, accountably, and with the judgement calls kept firmly in human hands.

That starts with deciding, deliberately, what not to automate. If that is a line your team is trying to draw right now, it is a conversation we would be glad to have.

‍Contact Us.