Most AI projects I get called into are demos wearing production clothes. Impressive on stage, useless at 2am when a real customer hits the edge case. The companies actually winning with AI are doing something much less glamorous: automating the boring middle of their operations, one workflow at a time, until the machine runs most of the day.
The best case I have seen up close got roughly 80% of operations automated. Not 80% of the interesting work — 80% of all of it. Here is what that path actually looks like.
Start with the work nobody wants
Forget the moonshot. List the workflows your team does every single day: support triage, order reconciliation, data entry from documents, status chasing, weekly reporting. Rank them by hours spent times how much judgment they need. The winners are high-hours, low-judgment. That is your automation queue, and it is never the sexy stuff.
Production AI is a systems problem
The model is maybe 20% of the work. The rest is what wraps it: structured inputs and outputs so the rest of your stack can trust the result, evals that run on every change so quality regressions get caught before customers do, confidence thresholds with a human fallback, logging that lets you replay any decision, and a kill switch. A demo needs none of that. A business needs all of it — and that gap is exactly why most AI initiatives stall after the demo.
The 80% is a ratchet, not a leap
Nobody automates 80% in a quarter. It ratchets: automate intake, watch it for a month, move the human to exceptions only. Then the next workflow. Each win frees hours that fund the next one. Eighteen months of that compounds into a business where the team touches only the exceptions — and the exceptions are where the humans were always most valuable anyway.
What the team does instead
The point was never headcount. The point is that the people who used to copy data between systems now work on the next big thing — the product bet, the new market, the partnership that was always postponed because operations ate the week. AI does not replace the team. It hands the team back their calendar.
If you are picking a first project this month: choose the workflow with the most hours, wrap the model in evals and a fallback, run it shadow-mode for two weeks, then flip it live for the easy 60% of cases. That first ratchet click teaches you more than any strategy deck.