The Challenge
The client makes steel. Blast furnaces, electric arc furnaces, casters, hot strip mills, mines, ports, and rail, across sixty countries, in a business where a single percentage point of yield is worth more than most software companies earn in a year.
For six years it had bought its artificial intelligence the way it bought most transformation, from two tier-one strategy consultancies and their embedded AI units. Roughly $40 million a year. More than sixty consultants on site at peak. The output was real, roughly thirty models built, some of them genuinely good. Four were running in production. The other twenty-six had been handed over as notebooks, decks, and a slide that said "operationalize." Nobody internally could maintain them. Models drifted. When a model broke, the consultancy was re-engaged to fix the model it had built. When an engagement ended, the people who understood the work got on a plane, and the understanding went with them.
The problem was never the consultants' competence. It was that the client had outsourced a capability rather than buying a project, and had no mechanism to ever take it back. Every year the dependency deepened, the internal skill base thinned, and the cost of insourcing rose. The board had begun asking why a company that builds its own blast furnaces could not build its own software.