In 2013 the American surgeon and writer Atul Gawande wrote a New Yorker piece in which he looked at the reasons why some innovations spread almost instantly whilst others can take decades to be adopted. He used a fascinating medical example, telling the story of how the use of anaesthesia (introduced in 1846) had spread globally within months. In contrast, antiseptics, in the form of Joseph Lister’s sterilisation methods (published in 1867) took a generation to be adopted.
Gawande noted that both innovations involved some degree of technical complexity, but whilst antiseptic practices remained stubbornly slow to take hold, anaesthesia birthed a whole new medical specialty – anaesthesiology. The key reason for the different rates of adoption, he says, was that one combatted a visible and immediate problem (pain) whilst the other combatted an invisible problem (germs) where the benefits wouldn’t be seen until well after the operation. Antisepsis also carried a cost for the surgeon since it required them to significantly change their routines and involved the use of carbolic acid which would sometimes burn the physician’s hands.
Gawande’s story is a great example of what Ronald Heifetz describes as the difference between ‘technical’ and ‘adaptive’ challenges. Technical problems can be fixed by experts applying known solutions, and change happens visibly and quickly. With adaptive challenges the problem is messy, progress requires learning, and solutions take time. Adaptive challenges can only be solved by the people that have the problem, since it requires them to change habits, behaviours and priorities.
Most AI transformation programmes are being designed as a response to a technical problem. Licences, Copilot rollouts, training on features, adoption metrics. These things matter of course, but the real challenge is an adaptive one – behaviour change. This technical framing is failing in some predictable ways including expected productivity gains not materialising, but being accompanied by replacement planning, to the disquiet of the workforce. Writer and Workplace Intelligence’s 2026 Enterprise Survey found that 29% of employees admitted to sabotaging their company’s AI strategy (!), with two of the main reasons given being a fear that AI would take their job and a poorly executed strategy. Then there’s over-estimating what the models can do end-to-end, which is leading to reversals in headcount reduction. Forrester’s Predictions 2026 report said that 55% of employers who reduced headcount citing AI now regret those decisions. And of course there is the all too common mistake of forgetting that constraints shift from volume to judgement. Bottlenecks in processes don’t just disappear when you introduce AI, they simply move to another part of the process, often the critical parts that involve humans. There was a lucid example of all this recently when Ford rehired over 300 veteran engineers after discovering that its AI-powered quality control systems and 900 plant cameras missed subtle defects that experienced human experts had been catching.
So what to do? BCG have a helpful way of framing technical vs adaptive change. They talk about taking a 70,20,10 approach to scaling AI, based on what separates the companies getting real value out of AI from those that aren’t. Around 10% of the effort should go into the models, 20% into the technology and data needed to make them work, and 70% into people and processes – redesigning workflows, rethinking incentives, building skills and changing how decisions get made. In other words about a 30/70 split between technical and adaptive work. Right now it’s probably the other way round – functional training focused on how to use the tools, adoption metrics, and a high expectation that somehow this will lead to significant behaviour change and productivity benefits.
If this ratio indicates how we need to roughly apportion effort, then Ethan Mollick’s ‘Leadership, lab, and crowd’ concept is a useful way of thinking about organising it. AI transformation, he argues, starts as a leadership problem. Leaders need to paint a clear and specific picture of what an AI-enabled organisation looks like, what it means for the people in it, and to realign incentives so that experimentation is rewarded rather than punished. The business theorist and psychologist Edgar Schein had a useful way of explaining why this framing matters. People only learn new ways of working, he said, when survival anxiety (the fear of what happens if they don’t change) outweighs learning anxiety (the fear of looking incompetent or losing status whilst they learn). The mistake that many leaders are making is to ramp up the former rather than reduce the latter. One study from late last year reported that 59% of executives said that they would replace staff who resisted adopting AI tools. Threatening to replace people is the bluntest form of ramping up survival anxiety.
The ‘crowd’ in Mollick’s model is the majority of the company population and it’s where the real discovery happens because the people doing the work are the best placed to figure out how AI can help them do it better. But the problem is often that most of this stays hidden or siloed. The ‘lab’ is a concentrated team of AI champions whose job it is to take what’s being discovered, to build and test new workflows, and spread what works. This references Heifetz’s idea of giving the work back to the people that have the problem, but it also helps avoid the ‘top-down fantasy, bottom-up trap’ of transformation, and to show change rather than just raising awareness. The ‘lab’ helps to turn individual discoveries into organisational capability.
The final piece of this puzzle is how behaviour change can spread effectively, and the work of the social psychologist Damon Centola is particularly useful here. He distinguishes between simple contagions like the spread of information (or a virus) which move from one person to the next through single contact, and complex contagions (like behaviour change) that are more costly or require people to give something up. Simple contagions travel best through broad networks with lots of reach but complex contagions require many examples and reinforcement from multiple people we trust before we will adopt them. In a 2010 experiment Centola found that a new health-related behaviour spread further and faster through tightly clustered networks where people heard about it from many overlapping contacts, rather than through random networks that had wider reach. In other words, for complex change depth beats breadth.
Anaesthesia was a simple contagion. Antisepsis was a complex one. Genuine AI transformation means fundamentally changing how you work which is about as complex a contagion as you can get because it means changing your routines, exposing what you don’t know, and perhaps even accepting that your role is going to look very different in the future. Yet the standard approach is a breadth strategy – raising awareness, Copilot for everyone, all-staff functional training, adoption metrics that tell you very little about how people are actually using it. A much more effective approach is to go deeper with AI champions, a sharper focus on a smaller range of key teams until new ways of working become normal there, and then to build what Centola calls ‘wide bridges’ (multiple connections rather than a single link) between those teams and the next. That’s how change actually spreads.
Towards the end of ‘Slow Ideas’ Gawande describes a programme in India trying to get birth attendants to adopt a handful of basic safety practices like washing hands and keeping newborns warm. Guidelines, posters and classroom training had all been tried and none of them had really stuck. What made the difference was a coach visiting in person again and again, building a relationship with them and working through the practices alongside them. His conclusion was that, even in an age of instant communication, new standards still mostly spread through people talking to people. AI technology will keep on improving, in most cases faster than organisations can absorb it, but the real change will happen when we stop treating this like a technical challenge and start dealing with it as an adaptive problem.
A version of this post appeared on my weekly Substack of AI and digital trends, and transformation insights. To join our community of over thirteen thousand subscribers you can sign up to that here.
To get posts like this delivered straight to your inbox, drop your email into the box below.
Leave a Reply