From stalled pilots to a working AI programme at a manufacturing group

A diversified manufacturing group had run three AI pilots in eighteen months — demand forecasting, predictive maintenance, and an internal chatbot — and none had reached production. Each pilot had worked in demonstration and died in handover. The board was losing confidence, and the executive team could not articulate why technically successful pilots kept failing commercially.

The diagnostic found the pattern we see most often in this band of readiness: pilots were scoped around what vendors could demonstrate rather than around the group's data reality. Production sensor data lived in incompatible historian systems across plants; the forecasting pilot had been trained on a hand-assembled extract that no one could reproduce monthly; and no one owned the decision of what "production" meant — there was no route from a working model to a maintained system.

Rather than starting a fourth pilot, the engagement restructured the programme. We ran a readiness assessment to make the gaps measurable and visible to the board, then sequenced the work: a data-foundation phase consolidating the two highest-value data flows into reproducible pipelines; a single production target — the use case with the strongest data foundation — carried through deployment with defined ownership, monitoring, and a maintenance budget; and an executive and engineering training track so the group could evaluate vendor proposals against its own data reality.

The advisory retainer gave the executive team a standing, vendor-neutral counterparty for the decisions between board meetings — vendor evaluation, scope control, and honest kill/continue calls on the remaining pilot backlog.

The group moved from three stalled pilots to one production system with an owner, a budget, and monitoring — and a repeatable route for the next use case. The board-level readiness picture converted AI from a source of scepticism into a governed programme with measurable milestones, and internal teams now run vendor evaluations against their own reproducible data rather than vendor-supplied samples.