Turn a Repeated Workflow Into an AI-Assisted Process
A two-week field method for converting one recurring workflow into an AI-assisted process without betting anything you cannot afford to lose.
The gap between an AI demo and an AI process is the gap between “it can” and “it does, reliably, on Tuesday afternoons, when the person running it is busy.” This guide crosses that gap deliberately slowly: one workflow, two weeks, and a decision at the end that you can defend with your own evidence.
Setup (90 minutes)
Pick the workflow like an investor, not a fan. The right first candidate repeats often enough to generate evidence quickly, hurts a little (annoying, time-consuming), and fails cheaply. Do not start with the workflow you most wish were automated — start with the one that can teach you fastest.
Write the current process as it actually happens. Ten to fifteen steps, in the words of the person who does it. The step they describe with a sigh is usually where the AI assistance belongs. The steps involving judgment calls about people are usually where it doesn’t.
Design the checkpoint before the automation. Decide, in writing: which step the AI performs, what the human reviews before anything leaves the team, and what “obviously wrong” looks like so review stays fast.
The two weeks
Run every instance of the workflow through the new process. Keep a plain log: date, what the AI produced, what the human changed, minutes spent. No dashboard, no tooling project — a note file is enough. The log is the deliverable; resist the urge to optimize mid-experiment.
The decision
At the end of two weeks you have something rare: local evidence. Three honest outcomes are possible — adopt (corrections are light and getting lighter), adapt (the assistance belongs at a different step than you guessed), or abandon (review costs more than the original work, which is a successful experiment that cost you two weeks instead of a year of quiet drag).
Whichever way it goes, you have also built the more valuable asset: a team that has watched AI succeed and fail at close range on its own work, and has calibrated accordingly.