Two years ago the standard AI strategy question was where do we start? Today it is quieter and more uncomfortable: we started — why didn’t it stick?
The pattern is remarkably consistent across companies we talk to. A promising pilot, real enthusiasm, a demo that impressed leadership — and then six months later the team is back to the old spreadsheet, and the pilot lives on only as a line item nobody wants to cancel. The problem is rarely the model. It is almost always one of four things.
1. The pilot automated a demo, not a workflow
Pilots get scoped around what shows well: a chatbot over documents, a summarizer, a content generator. Production value lives somewhere less photogenic — inside a specific, repeated, measurable business process with an owner, an input queue, and a definition of done. If you cannot name the person whose Tuesday changes when the system ships, you scoped a demo.
The fix: pick one core workflow — singular — and follow it end to end. One input, one transformation, one output someone consumes. Everything we build, from client automation layers to our own internal tooling, starts with this constraint, and it is the same discipline behind scoping an MVP to one core workflow.
2. Nobody owned the last mile
The gap between “works in the notebook” and “works every morning without a babysitter” is where pilots die: error handling for malformed inputs, fallbacks when the model is uncertain, logging someone actually reads, and a human escalation path. None of this is research. All of it is engineering, and it is usually unassigned.
The fix: staff the last mile explicitly. Production-izing a workflow is an integration project — APIs, data plumbing, monitoring — that happens to have a model in the middle.
3. The humans were an afterthought
A workflow is a social system. If the people running the old process are handed a tool that silently judges their work, they will find its failures fast and its uses never. Adoption is designed, not announced.
The fix: put the current process owners in the loop from week one — reviewing outputs, correcting them, and visibly shaping the system. The correction loop is not just change management; it is your evaluation dataset.
4. Success was never defined in operational terms
“Explore AI opportunities” cannot fail, which means it cannot succeed either. Production systems need operational definitions: hours redirected, turnaround time, error rates against the human baseline — measured on the workflow you actually shipped, not the technology category.
The fix: write the success metric before the first line of code, and make it a number the workflow’s owner already tracks.
The one-workflow rule
If there is a single transferable principle, it is this: breadth is the enemy of production. A company with one AI-run workflow in daily use learns more — about its data, its edge cases, its people — than a company with five pilots in perpetual evaluation. The first workflow is the hard one; it forces every unglamorous question. The second is dramatically easier, because the questions have answers.
Start smaller than feels ambitious. Ship all the way. Then widen.