Reskilling programmes fail predictably when they run in parallel with unchanged work. People attend, they enjoy the course, they return to the same queue with the same targets, and within a month the new skill has no place to land.
The sequence that works is the reverse of the intuitive one. Redesign the process with AI in it, define the new roles and the decisions those roles own, and only then build the learning path — targeted at the delta between what people do today and what the redesigned role requires.
“If the work does not change shape, the training does not change behaviour.”
Three design choices matter more than curriculum quality. First, teach in the workflow: short interventions attached to real tasks outperform day-long workshops. Second, make the new behaviour observable — if managers cannot see the new behaviour, they will manage the old one. Third, protect the transition: people moving into AI-augmented roles need a period where their old throughput targets are formally relaxed, or they will quietly revert.
There is also a harder truth about composition. Most organisations will need fewer people in roles defined by routine processing and more in roles defined by judgement, verification, and exception handling. Being honest about that in workforce planning is uncomfortable and necessary; discovering it through attrition is worse.
The organisations that do this well treat reskilling as an operating change with a training component, not a training programme with an operating excuse.
