An honest guide to AI workflows that pay for themselves
Not the demos. The three places AI has measurably paid for itself in delivery work, and the places it quietly cost more than it saved.
article hero image
Where it pays
Reading unfamiliar code, drafting the first version of a document nobody wants to start, and turning a messy transcript into structured notes. All three share a shape: high-volume, low-stakes, verified by a human in seconds.
Where it costs
Anywhere the verification is harder than the work. Generated tests nobody reads, architecture opinions with no accountability, and code in the parts of a system where being subtly wrong is expensive.
Make the team keep it
The difference between adoption and abandonment is whether the workflow survives the week I leave. That means it lives in their tooling, their prompts, their repo — not in my head.
If you cannot say what it replaced, it did not save you anything.