The headline is easy. The integration is hard.
Every enterprise has announced an AI initiative. Fewer have one running in production every day, validating field reports against site equipment, digitising avometer readings, and flagging cleanliness deviations against a uniform standard. The gap is not the model — pre-trained vision models for these tasks are commodity. The gap is integrating those models into the workflows where the work actually happens: on-device, in poor connectivity, with imperfect captures, on phones held by tired technicians at the end of a shift.
Lesson 1: design for the capture, not the inference
Most AI projects start by improving the model. The bigger leverage is improving the input. On-device capture guidance — framing overlays, blur detection, angle validation — improved downstream inference accuracy more than any model tuning did. Build for the moment of capture, not for the moment of inference.
Lesson 2: confidence routing is the product
Pure automation rates are misleading. What matters is which decisions are routed to humans, and how cleanly. We design pipelines where exceptions surface to the right reviewer with the right context — making the human-in-the-loop step take seconds, not minutes. The system is measured by how quickly it stops needing the human, not by raw automation percentage.
Lesson 3: ship the simplest model that closes the loop
The temptation in any AI engagement is to optimise the model. Resist. Ship the simplest version that delivers measurable operational change, then iterate against real production signal. The closed loop — capture → infer → action → feedback — is worth more than any single model improvement.
Closing
AI doesn't transform organisations. People using AI well do. Our role is to build the connective tissue that makes AI accessible to the people closest to the problem.