The human in the loop: An employer’s guide to leadership in the age of autonomy

For as long as we’ve built machines that make decisions, we’ve reassured ourselves with the fact that there is a human in the loop.

What began as autopilot in airplanes, or automated control systems in factories – the uniting philosophy is that a person remained in the chain. They were able to see what the machine was doing and, if necessary, to override it. However capable the technology became, a human would always be there as backup.

That promise is now being tested in a new way and in almost every workplace. For most of the last three years, the conversation about AI was a conversation about tools – systems that draft, summarise, and suggest, but wait to be asked. That’s changing.

At our Future Workspaces Conference, we tackled this topic with one of our speakers, Paul Bentham, who said that agentic AI “has agency, it makes decisions.” Ask it to “book my travel for next week” and it doesn’t return a suggestion; it acts. Around four in five organisations now report adopting AI agents in some form, and Gartner projects that by 2028, roughly 15% of day-to-day work decisions will be made autonomously.

When the machine stops advising and starts doing, the loop changes shape.

And that raises a question: who, exactly, is the human we’ve left in there?

And this is the challenge we’re tackling at our next Future Workspaces event, held during Cambridge Tech Week in September: can technology ever deliver its full value without leaders who bring context, judgement, and empathy?

What follows here introduces the potential answers to those questions – and what employers need to consider, practically and legally.

The machine that acts on its own.

Agentic tools – that book travel, reconcile the invoice, or control the inbox – have already acted by the time anyone looks.

Jenna Goldstein of The Berkeley Partnership asked the simple, important governance problems of this setup:

“How do we manage bots? How do we manage the people who are managing bots?”

For most organisations, the honest answer is: they’re not. And Gartner expects that more than 40% of agentic AI projects will be cancelled by the end of 2027, with inadequate risk controls among the leading causes. This is the control gap – wherein recent survey data suggests that almost two-thirds of organisations cannot enforce purpose limitations on their AI agents, and 60% cannot reliably shut down one that misbehaves. An agent you cannot stop is not a tool you are using; it is a decision you’ve delegated outside of the human sphere.

None of this is an argument against adoption. Speed, as Paul Bentham argued, matters – “work fast with empowered tools.” But the faster an organisation moves from AI-that-suggests to AI-that-acts, the more the burden shifts onto whoever remains in the loop, and the more it matters that the loop has been designed rather than assumed.

Someone has to know what good looks like

Putting a person in the loop protects you only if that person can tell when the machine is wrong. Rafah Knight of Secure AI Consultancy put it plainly at our conference – when a model hallucinates, “how do you know, if you don’t have the experience to know what good looks like?”

If you have done your human job long enough, you can tell whether the output is good; if you are junior, you cannot.

Automation bias is the well-documented tendency of people, including experts, to defer to an automated recommendation and reduce their own scrutiny. The counter-intuitive part is that better systems make it worse: the more useful an AI system is perceived to be, the stronger the pull to accept even its flawed recommendations. Researchers have a blunt term for what results – “rubber-stamp oversight”: fast, efficient, but fragile. The law has noticed too. The EU AI Act’s human-oversight provisions go so far as to require that those assigned to oversee high-risk systems are kept aware of their own tendency to over-rely on the machine’s output.

A signature on an approval is not really oversight.

Oversight is the capacity to look at what the system produced and say, with authority, no.

Context, judgement, and the things you can’t automate

If the human in the loop has to be someone with the expertise and standing to overrule the machine, then oversight is not an administrative task to be pushed downwards. It is fundamentally a leadership function. This was another thread running through our conference. David Bellamy of Harkn reduced it to five words – “the human is the most important asset” – and located responsibility firmly at the top. Leadership, he argued, owns the conditions in which people work. KPMG’s Steve Nathan made the same point, placing Leadership and Ways of Working alongside Technology as equal elements of workplace design, not subordinate to it.

What leaders supply is precisely what the machine cannot:

• Context – knowing which cases are the exceptions
• Judgement – knowing when a confident answer is a wrong answer
• Empathy – knowing that a decision about a person is never only a data problem.

David Terrar of the Tech Industry Forum warns that models themselves carry the biases of the data they learned from, which means an unsupervised output can encode outdated assumptions into the way a system thinks from that point on. The scarce skill is no longer producing the work – it is judging it.

Yet somehow, accountability for AI has been slow to climb to where it belongs. In a McKinsey survey, only around 28% of organisations said their CEO takes direct responsibility for AI governance, and just 17% said their board does. Governance is currently being treated as something that happens somewhere in the organisation rather than something leaders own. But the loop needs a leader precisely because a checkbox will not do.

The vanishing bottom rung

There is a sting in the tail of all this. If the human who oversees the machine has to be experienced enough to know when it is wrong, where will the next generation of such people come from? Adam Clements of Freshminds put the question directly: junior and low-level tasks – transcription, summaries, first-pass analysis – are exactly the work AI now absorbs.

“We need fewer entry-level hires. Great for productivity, not great for the world. Where are future managers and future leaders going to come from?”

The early data suggests this is not hypothetical. Analysts have begun describing a ‘seniorisation’ of work: entry-level roles do not so much vanish as get quietly promoted up the skills ladder, out of reach of the people trying to start . A Harvard paper analysing 62 million workers found that junior hiring fell by nearly 8% within six quarters at firms that adopted AI – not through layoffs, but through a freeze on new positions.

In the United States, recent-graduate unemployment reached 5.7% late in 2025, with underemployment above 40%, and the World Economic Forum reported that 40% of employers expect to reduce headcount where AI can automate the tasks involved.

Some economists point out that the junior hiring slowdown began before most firms had meaningfully adopted generative AI – making AI at least partly a convenient explanation for decisions driven by the wider economy. But the structural risk stands regardless of the cause: entry-level work has always been where judgement is grown, and if that ground is automated away, organisations may find they have optimised the present at the expense of the people they will need to run the loop in a decade.

You cannot, in the end, build senior judgement in a workforce that was never allowed to be junior.

Where to get started

1. Design the loop; don’t assume it. Before deploying an AI system that acts – ask where the human sits, who that human is, and what power they have. Start with the unglamorous basics: which tools are approved, a mandatory human review for consequential decisions, logging capability, and an override function.
2. Put your most capable people, not your most available, in the loop. Oversight only protects you if the overseer can recognise a wrong answer. That means matching the seniority of the reviewer to the stakes of the decision, and resisting the temptation to treat AI review as low-value administrative work to be handed to whoever has capacity. The reviewer needs the expertise to spot the error and the authority to act on it.
3. Guard against your own success. The better your AI performs, the more your people will trust it – and the more likely they are to wave through the occasional bad output. Build in deliberate friction where it matters: require justification for acceptance on high-stakes decisions, and reward the person who challenges the machine rather than the one who clears the queue fastest.
4. Protect the pipeline. If AI is absorbing the entry-level work where judgement has always been learned, you need to be deliberate about how junior people now develop it. Redesign early-career roles rather than removing them – you’ll be investing in the future supply of people capable of running the loop at all.
5. Get the legal foundations in place. Accountability for what an AI system does in your name sits with you. Understand your obligations under UK data protection law, the duties that flow from the EU AI Act if you operate there, and the employment and equality implications of restructuring around automation. Seek advice specific to your size, sector and risk appetite – and put the register, the reviews, and the audit trail in place before, not after, something goes wrong.

Continue the conversation in Cambridge

This article previews the themes of our next Future Workspaces event, a Cambridge Tech Week fringe session at Emmanuel College on Wednesday 16 September, where business leaders and experts will explore how organisations lock in the leadership skills and human perspective that allow technology to deliver real value.

Places are limited, register here.