The oldest paper on my doctoral reading list is about coal mining. It was published in 1951, and it explains the newest problem I see in AI programmes better than anything written this year.

The problem goes like this: an organisation introduces a better technology, carefully and competently, and the results get worse. Or they stall, which in a board report looks almost the same. The technology works. Something else broke.

For the rest of us: what “sociotechnical” means

Every system of work has two halves. The technical half is the tools: machines, software, now AI. The social half is everything else: who works with whom, who depends on whom, who decides, what gets rewarded, what people feel safe admitting.

Most technology programmes change the technical half and treat the social half as fixed. “Sociotechnical” is the academic word for a simple observation: the two halves are one system. Change one without redesigning the other and you can make the whole thing worse, even when the new tool is better.

What happened at the coal face

Eric Trist, a social scientist at the Tavistock Institute, and Ken Bamforth, a former miner, studied British collieries that had moved to mechanised “longwall” mining.

In the older method, miners worked in small, self-organising groups, typically pairs who shared the whole job of getting coal from a short stretch of face. Each depended on the other, and each could see the whole task.

The longwall method was technically superior. An electric coal cutter undercut the coal and a conveyor carried it away. But it reorganised the work around the machines. The job was split into a three-shift cycle: one shift cut, one shift loaded, one shift moved the equipment forward. On the loading shift, some twenty men each worked their own identical 8–10-yard section of a face 180–200 yards long. The filler who had once had a partner now depended on an “absent group” of workers from earlier shifts whom he rarely saw, and who had no stake in whether his stretch of coal was workable when he arrived.

The social consequences were severe, and Trist and Bamforth documented them in detail. Men developed “reactive individualism,” competing for better places instead of helping each other. Groups fell into mutual scapegoating across shifts. And absenteeism became a coping technique. When conditions got bad, absences piled up until the remaining men had to stay down extra hours to clear the face, and sometimes, if too few arrived, everyone went home. The authors described a “degree of social illness costly alike to productivity and to personal well-being.”

The machines had not failed. The system around them had. The Tavistock researchers’ later work was titled simply Organizational Choice. The same technology could be run with a very different social organisation, and the choice mattered.

Why this keeps happening

You might expect that seventy years of this lesson would have sunk in. Enid Mumford, one of the people who spent a career trying to make it sink in, concluded in 2006 that sociotechnical design had been widely praised and rarely practised. In each new technology cycle, it lost out to rollouts optimised for efficiency first.

AI looks like the next cycle. RAND interviewed 65 experienced data scientists and engineers about why AI projects fail. The first root cause they list is a leadership failure: stakeholders misunderstanding or miscommunicating what problem the AI is supposed to solve. That isn’t a technical failure. It’s a failure in the social half of the system.

McKinsey’s 2026 State of AI survey puts it in consulting language, but the point is the same. After a long list of practices that distinguish the small group of companies getting real returns from AI, one of the authors concludes: “The limiting factor is increasingly the organization’s ability to absorb change.”

What the social half looks like in an AI programme

The McKinsey data is useful here, because the practices that separate its high performers are mostly social, not technical:

  • Leaders visibly own it. High performers are twice as likely to say senior leaders demonstrate real ownership of AI initiatives: championing them, role-modelling, keeping them funded.
  • The incentives change. Among the practices McKinsey finds high performers following more often: making employees’ use of AI a factor in performance management, and completing strategic workforce planning on how roles and skills shift.
  • Someone has authority to remove bottlenecks. Another practice on the list is a transformation office, or an equivalent, empowered to clear the organisational obstacles that stop pilots from scaling.
  • The work is redesigned, not just equipped. Nearly three-quarters of high performers have fundamentally redesigned workflows. Among everyone else, it’s about a quarter.

Read against the coal mine, those aren’t “soft” extras. They’re the redesign of the social half.

The culture shifts that follow

From my own work, three shifts matter most.

From “use the tool” to “redesign the work.” A rollout that measures licence activation is optimising the technical half. The question that matters is what the team’s work looks like once AI is in it, and whether the team had a say.

From hiding to disclosing. When people use AI quietly because they’re unsure whether it’s allowed or respected, you get the collieries’ “absent group” problem in a new form: work depends on things nobody can see. I’ve written before that shadow AI is a verdict on your culture, not your security. The fix is making disclosure safe, not just banning tools.

From deciding for people to designing with them. Sociotechnical design had a principle called participation: the people affected help design the system. With AI, that group now includes people whose tasks an agent takes over, people who use its outputs, and people who have to check its work. They know where the process will break, and they’re usually the last to be asked.

What this means

The mines didn’t need worse machines. They needed the work organised around the machines in a way people could sustain. AI programmes are no different. The tools will keep getting better on their own. The social half of the system won’t, and in most organisations that’s now the half limiting the return.

If your AI programme has a technology roadmap but no plan for how roles, incentives and team structures change, you’ve designed half a system. Seventy-five years of research says the other half doesn’t just go along with it.

References