Wunmi Alimi
Managing Director · 7 min read
A familiar pattern: a technical pilot succeeds under controlled conditions, gets declared a win in a steering committee deck, and then very little changes in day-to-day operations six months later. The model worked. The transformation didn't. These are not the same achievement, and treating them as equivalent is the single most common failure mode we see.
Model accuracy is close to irrelevant if the people meant to use the system don't trust its output, if their actual workflow was never redesigned to incorporate it, or if their incentives still reward the old way of working. A technically excellent recommendation engine that nobody's workflow has room for is, in business terms, indistinguishable from a broken one.
The fix is to run change management in parallel with technical delivery from day one rather than treating it as a training module bolted on after launch. Stakeholder mapping, incentive redesign, and workflow redesign need to happen concurrently with model development, not sequentially after it, when the system is already built around assumptions nobody validated with the people who'll actually use it.
There are reliable early warning signs that a transformation is failing even while the technical metrics look fine: usage that declines after the initial novelty wears off, frontline staff quietly building workarounds that route around the new tool, and no movement in the underlying business metric the AI was meant to move in the first place.
The practical recommendation is unglamorous: track adoption metrics with the same rigor and in the same steering report as model metrics. If precision and recall get a slide, so should weekly active usage by the intended frontline population, and so should the business metric the whole initiative was justified on.
Technology proves that the AI works under the conditions you tested. Change management proves the organization will actually use it under the conditions it really operates in. Most transformations fund the first generously and the second as an afterthought, and then are surprised when only one of them shows up in the results.
“A model that works and a workforce that won't use it produce the exact same business result: zero.”