Human in the Loop Is the Wrong Way to Think About AI
By Professor Ashley Braganza
Human in the loop and AI in the loop are the phrases leaders reach for to sound careful about AI. Both smuggle in an assumption worth challenging: that the human and the machine sit in separate places, and the only question is where to put the checkpoint between them. Kerri O'Neill puts it more sharply than most. The more we imagine humans over here and AI over there, she argues, the more risk we take on. On this account the loop is not a safeguard. It is a category error.
My guest on The AI Adoption Podcast is Kerri O'Neill, Chief People Officer at Ipsos UK and Ireland and its Global AI Workforce Transformation Lead, a role that spans 20,000 people across 93 countries. She is the author, with Alex Bailey, of Rapid Reculturing, published by Wiley. Her argument is that most of what goes wrong with AI is not technical at all. It surfaces in the cultural layer, where the technologists rarely look.
Adoption fails in the cultural layer, not the technology
Kerri is direct about where AI implementations come apart. There is nothing wrong with the technology, she says, and nothing wrong with the technologists who bring it in. The culture around it fails instead, the shared understanding that decides whether people trust a new tool or quietly route around it. This matches what I put to boards and senior managers I work with: the model is rarely the binding constraint. The organisation is.
Her prescription starts with purpose. Kerri makes the case that purpose is not a backward-looking statement of who a company has been, but a forward description of the future it is trying to create, and that agents need this direction of travel as much as people do. An agent given a task without a purpose optimises the task. An agent given the purpose can weigh the task against it. Whether a system that predicts language can participate in a culture, rather than supply more of the words a culture is built from, is an open question of capability. She extends the same logic to mindsets, which she treats as trainable in a way that deeply held values are not. A team can adopt a pace mindset for a quarter without rewiring its values, and agents can be framed the same way, designed on how they should work and not only on what they should do. The verdict is uncomfortable for anyone selling adoption as a software rollout: the hard part was always the people, and it now includes the agents.
Agents need the certainty that people learn to live without
Here is the tension at the centre of the conversation. People are not comfortable with uncertainty, but they are practised at it. We meet ambiguity, feel the anxiety of it, and work through it, often by talking to someone else. Agents are built the other way. They are given a context and a frame, and they operate inside guardrails that narrow ambiguity by design. Place the two side by side and the friction is plain. The way people usually close that gap is by talking to one another, which is why dialogue keeps returning as the thread running through this conversation.
Kerri complicates the easy reading, which is where the conversation earns its keep. The instinct is to give agents as much certainty as possible. She warns against it. Give an agent too much certainty, she argues, and you risk accidentally narrowing the organisation's ambition, because the unexpected is part of any change worth making, and a system tuned to remove surprises will remove the productive ones too. Her more interesting move is to reframe resistance itself. Rather than treating pushback as friction to be overcome, she reads it as a signal about what people are trying to protect. That reframing sits awkwardly beside a technology built to resolve ambiguity quickly. An organisation learning to sit with uncertainty, and one deploying agents to eliminate it, are pulling in opposite directions, and leaders will have to hold that tension rather than resolve it early.
Empathy is the capability leaders keep assuming they can automate
For leaders, the practical question is not where to place a human in a loop, but what becomes of human connection when most of the working day is spent with agents. Kerri does not soften this. She points out that loneliness, in medical terms, is as much of a threat to health as obesity, and that heavy reliance on agents can feel like connection in the moment while thinning it over two or three years. She describes the human brain as the supercomputer behind your eyes, and the serendipity of people thinking together as something AI struggles to replicate.
This is where the loop framing does real damage. When a person hits an ambiguous moment, they turn to another person, and an agent can answer in fluent, sympathetic language. The empathy a colleague brings to that moment is not something a rules-based system supplies. Kerri and I discussed a live version of this: people already lean on chatbots as advisers and, in some cases, as a form of therapy in their private lives. Organisations will face the same question at work. Whether a company should provide such agents to its employees, and what that does to the culture, is a leadership decision, not a procurement one. The organisations that treat connection as something to design, rather than assume, will be the ones whose cultures survive the transition.
Trust cannot be bought off the shelf, and neither can adoption
Kerri's account of adoption at Ipsos gives the argument a concrete shape. The company has set out a vision it calls the augmented Ipsos, led by its CEO, Kelly Beaver, built on a foundation of trust, truth and transparency, with four operating principles adapted for the AI age: security, simplicity, substance and speed. Kerri reports higher than average levels of adoption inside the organisation, and attributes it to something other than the tools. People trust the leaders on where they are taking the work.
That link between trust and adoption is easy to assert and hard to engineer. Trust, like culture, is intangible. You cannot take it off a shelf and put it on a conveyor belt, as Kerri puts it, and a company that treats AI as one more productivity tool will not build it. The organisations reporting real adoption tend to be the ones that have described where they are going and why, and have earned the belief that the destination is worth the disruption. The lesson for leaders is not that trust is nice to have. It is that trust is the means by which adoption happens at all, and no amount of capability compensates for its absence.
Culture is carried in the millions of ordinary conversations that agents are now poised to reshape. Leaders can let that happen by default, or they can decide, deliberately, which conversations must stay human. The augmented organisation is an appealing destination, but the route there runs through a question most adoption plans avoid. On this evidence, the organisations that keep asking the purpose of their culture, and who it serves, will adapt faster than those still debating where to place the human in the loop.
Listen to the full conversation
YouTube: https://youtu.be/2QESjAQRWdw
Comments