The pattern
Most AI pilots follow the same arc: an impressive demo, a wave of enthusiasm, a quiet death. Six months later the company has a slide that says "we explored AI" and an operation that works exactly like it did before.
Why it happens
Pilots stall because they're designed as experiments, not as systems. Nobody owns them after the demo. They run on sample data instead of the real, messy feeds. They have no error handling, no escalation path, no answer to "who fixes this at 2am". The model was never the hard part — the operations around it were, and the pilot skipped them.
A pilot that can't survive a holiday weekend was never going to survive production.
What works instead
Start from one real workflow with a named owner and a measurable cost. Build the boring parts first: data access, logging, the human handover. Ship to production on week one at tiny scale, then widen — instead of polishing a sandbox for a quarter. A small system that runs beats a large demo that doesn't.