A familiar steering committee scene: a pilot six months in, its budget spent, and no result anyone can measure. The model is rarely the reason. The pilot involved too many stakeholders, its success criteria were written as a business case instead of a dated hypothesis, and so much had been invested by month three that nobody wanted to be the person who said it was going wrong. MIT NANDA's 2025 study of more than three hundred enterprise deployments found ninety-five percent of generative AI pilots produced no measurable profit-and-loss impact, and Gartner forecast that at least thirty percent of generative AI projects would be abandoned after proof of concept by the end of 2025.
Lean thinking was built for exactly this situation. Toyota developed its production system from the 1950s to shrink batch sizes, keep workers close to outcomes, and let anyone on the floor stop the line by pulling the andon cord. Eric Ries later turned the same instincts into build, measure, learn for software startups. AI has all the conditions Lean was designed for: behaviour nobody can fully predict in advance, deployments that are expensive to unwind at scale, and frontline staff who know where the work actually breaks but are rarely asked.
Five practices translate directly. Go and see the real workflow before automating it, the practice Toyota calls genchi genbutsu. Define value from the user's problem rather than from what the model can do. Start with one real workflow and real users instead of four departments at once. Set stop-or-go metrics and a review date before the pilot begins. And give one named person an andon cord: the authority to pause the deployment without first writing a business case to justify it.
Small batches reduce risk before they increase speed. A small pilot limits exposure at the point where the team knows least. A small failure is survivable, so people report it honestly, while a large one pushes everyone involved to quietly redefine success so the project can be called a win. In practice the first few weeks look like this: pick one workflow with a named owner and recurring friction, watch the work being done, write a one-page plan with a success metric, a stop threshold, a review date and a named andon cord holder, then build the smallest version and test it on real transactions. If the evidence says stop, the pilot has done its job.
No new governance infrastructure is needed to start. Take the widest-scoped AI initiative on the current roadmap and ask what its smallest honest version would look like.
Reference
This piece is adapted for Praxora Lab from the original. Originally published at terencekok.com ›
