June 21, 2026 · ← Notes
When I map a client's AI stack, I use two kinds of arrows. Solid lines for flows that run automatically. Dashed lines for flows that require a human to move something from one place to another.
Most stacks I look at have more dashed lines than solid ones. And that gap — between what's automated and what isn't — is almost always where the unrealized value lives.
A 10-minute task done twice a week is 17 hours a year. That's before you account for the context-switching cost, the delay it introduces into a downstream process, and the fact that the person doing it is usually someone whose time is expensive. The math is brutal once you write it down.
So why do these handoffs persist? Three reasons, in my experience.
Most operators don't have a documented map of their own workflows. The handoffs are invisible not because they're hidden, but because nobody's ever drawn the diagram. Day-to-day, you're building, selling, and supporting customers — not auditing the connective tissue between systems. The manual steps just become part of how things work, and they stop being visible as a cost.
Individual tasks feel too small to justify automation. Ten minutes here, five minutes there. Each one, considered alone, seems like a rounding error. But these tasks don't exist in isolation. They repeat. They compound. Across a team, across a year, what looks like a minor friction point is often a significant drag on capacity — one that never shows up as a line item because nobody added it up.
The honest reason many handoffs stay manual is that operators formed their automation assumptions a few years ago and haven't revisited them. The calculus has shifted. AI can now handle judgment-based tasks that previously required a human — classifying edge cases, drafting responses to ambiguous inputs, making routing decisions based on unstructured text. The tools caught up. The mental models didn't.
Write down your five most frequent workflows. For each step, mark it as automated or manual. Then count the dashed lines and estimate the time per week each one consumes.
That's it. No consultant required. The exercise itself usually surfaces two or three handoffs that have an obvious fix — and the annual time value of eliminating them tends to be larger than anyone expected before they ran the numbers.
The automation opportunity in most AI stacks isn't in the AI layer. It's in the manual steps surrounding it.