When organisations talk about adopting AI, the conversation usually starts with capability: what the tools can do, which tasks they can take on, how much time they might save. But the harder and more consequential story is rarely technical. Introducing AI changes how people work day to day, it redraws who is responsible for what, and it stirs up how people feel about their jobs and their worth. For professional service firms and public institutions, where the work is the expertise, and where trust and accountability are critical, the human dimensions are the main event. The encouraging news is that there is now a solid and growing body of evidence, from academic field experiments and from real deployments, about what leading that change well actually looks like.
How people work is changing
The most rigorous study of AI in knowledge work to date makes the point vividly. In a pre-registered field experiment with Boston Consulting Group, Dell’Acqua and colleagues (2023) gave 758 consultants realistic tasks to complete with and without GPT-4. On tasks inside what they called the “jagged technological frontier” – the uneven set of things AI does well – consultants using AI were faster and produced markedly higher-quality work. But on tasks that fell just outside that frontier, AI could make performance worse, because people trusted polished output that was subtly wrong. The researchers observed two adaptive styles: “centaurs”, who divide labour cleanly between human and machine, and “cyborgs”, who interleave the two at every step. The lesson is not that AI makes work easier. It is that AI changes the texture of work, and the new core skill is judgement about where the frontier lies, and when to interrogate the machine rather than defer to it.
From doing to judging
As AI absorbs routine drafting, search, and first-pass analysis, the centre of gravity of professional work shifts towards judgement, review, and the handling of exceptions. That is a change in roles, and often in identity. When the law firm Allen & Overy (now A&O Shearman) rolled out the legal AI tool Harvey to around 3,500 lawyers from 2023, what stood out was not the technology but the choreography. The firm trialled it from late 2022 through a dedicated innovation group led by senior champions, began with a narrow set of high-value, lower-risk tasks (contract drafting, due diligence, regulatory work), and kept lawyer review firmly in the loop, positioning the tool as an assistant rather than an autonomous decision-maker. Rollouts of this kind typically run over six to nine months and have to win over partners, associates, and knowledge-management lawyers separately, because each group has different reasons to adopt or resist. The reframe matters: the AI is not coming for their expertise, it is taking the routine work that was never the point of their expertise in the first place. Whether the loss of that routine work is a problem for junior staff – who have traditionally learned the craft by doing it – remains a huge open question.
How people feel changes
Feelings about AI are consequential. A long line of behavioural research documents “algorithm aversion”: Dietvorst, Simmons and Massey (2015) found that people abandon an algorithm m far faster than they abandon a human after seeing it make a mistake, even when the algorithm is more accurate overall – and the effect tends to be stronger among domain experts. Yet the picture is not uniformly negative. Logg, Minson and Moore (2019) documented the opposite tendency, “algorithm appreciation”, in which people often prefer algorithmic advice for analytical judgements. But these reactions are fragile and highly context-dependent. A 2025 study of patient preferences is a striking case in point: patients preferred clinical replies drafted by AI over those written by a human – yet once that AI authorship was disclosed, their satisfaction fell (Cavalier et al., 2025). How people feel about AI, in other words, tracks framing and transparency as much as the quality of the output itself. Which reaction dominates depends on exactly these things: how openly the AI’s role is disclosed, how much control people keep, and whether they have seen the system get things wrong.
The same ambivalence surfaces when these tools reach real workplaces. In the UK government’s cross-departmental Microsoft 365 Copilot trial – 20,000 civil servants across twelve organisations over three months – users saved an average of 26 minutes a day, and 82% said they would not want to give the tools up. But the gains were uneven: 17% reported no time savings at all, and users repeatedly flagged that the technology struggled with the nuanced, complex or sensitive work that calls for human judgement. Adoption is uneven, and trust has to be earned task by task.
Leading the change well
The cautionary tale of Klarna. The fintech froze hiring, leaned heavily on an AI chatbot it said did the work of 700 agents, and framed the move publicly as cost-cutting. By 2025 its chief executive acknowledged that the company had pushed automation too far, that quality had dropped and trust had eroded, and it began rehiring human agents, concluding that customers must always be able to reach a person. The pattern across the sector is consistent: AI handles routine, high-volume work well, but humans remain essential for escalation, complexity, and anything emotionally or reputationally sensitive. Framing AI as pure replacement tends to backfire; hybrid models tend to win.
So, what does leading the change well involve? The classic change-management literature still applies. Take Lewin’s (1947) unfreezing, changing and refreezing. Unfreezing is making the case for AI and surfacing the anxieties it provokes – the very aversion and identity worries described above; changing is piloting on a few well-chosen, lower-risk tasks with support, so people can find where the technology is reliable; refreezing is embedding those new ways of working, and the judgement they demand, into everyday routines and training. Kotter’s (1996) emphasis on a guiding coalition, a clear vision, and early visible wins is exactly what the Allen & Overy approach delivered and what Klarna’s top-down approach lacked. And the technology-acceptance literature (Davis, 1989) is a reminder that people adopt tools they find genuinely useful and genuinely easy to use – so part of the job is to make the tools that good, and part is to show it.
Drawing the threads together, four things separate good transitions from bad ones. Be honest that the goal is augmentation, not quiet replacement; people sense the difference and respond accordingly. Start narrow, on high-value and lower-risk work, and let early wins build confidence. Lead through credible champions rather than by mandate from the top. And take the human response seriously: create the psychological safety for people to experiment and, crucially, to flag when the machine is wrong; invest in the judgement and skills the new work demands; and keep people in the loop where the stakes, the nuance, or the accountability are high – which, in professional and public-sector work, is most of the time.
The technology, increasingly, is the easy part. The organisational and human transition is the hard part – and the evidence is clear that it is something to be carefully led, not merely deployed.