Section 1
Write the equation before you buy
Return on a workflow automation is the annual cost of the current process, minus the annual cost of the new one, minus the transition cost, divided by what you invested. Every term needs measuring rather than estimating. The current cost is the one people skip, because it requires watching the work for a week. Count runs per month, minutes per run, and how often the output has to be redone. That is your baseline, and it is also your evidence later when someone claims the automation changed nothing. Without a baseline you are not measuring return. You are comparing a memory with a dashboard. Definitions and mechanics are in [What Is AI Automation? A Plain-English Guide for Founders](/blog/what-is-ai-automation-a-plain-english-guide-for-founders).
Section 2
The cost lines founders forget
Five costs sit outside the subscription price and routinely double the real figure. Integration: connecting the tool to your systems, usually the largest single item. Review time: someone checks the output, and that time is a permanent operating cost, not a temporary one. Usage variance: per-run pricing scales with volume and with input size, so long documents cost more than the pilot suggested. Maintenance: templates change, formats drift, the model provider updates something. Budget a few hours a month per live workflow. Change cost: the fortnight where the team is slower because they are learning. It is real and it is always underestimated.
Section 3
Saved time is not saved money by default
This is the part that separates a real return from a slide. Four hours a week freed across five people is twenty hours, and it is worth exactly nothing until one of three things happens. You avoid a hire you would otherwise have made. That is cash, and it is countable. You redirect the hours into revenue-generating work and the revenue moves. Countable, with a lag. You reduce paid capacity, through contractors or overtime. Countable and immediate. If none of the three happens, the hours are absorbed and your cost base is unchanged. There may still be a good reason to do it, better quality, faster response, less tedium, but do not call it return. Related mechanics for on-device work are in [Edge AI Automation: What Founders Need to Know](/blog/edge-ai-automation-what-founders-need-to-know).
Section 4
Payback, and an honest range
State payback as a range, not a point. Best case assumes the automation handles the share of cases you saw in testing. Worst case assumes it handles half of them and someone still reviews everything. If the worst case still pays back inside a year, the decision is easy. If only the best case works, you are betting on the pilot generalising, and pilots systematically flatter, because the hard cases arrive later. Also set a stop rule in advance: if the correction rate has not fallen below a stated level by a stated date, the project ends. Decided beforehand, that is discipline. Decided afterwards, it is an argument.
Section 5
When return is the wrong frame
Some automations exist to reduce variance rather than cost, and forcing them into a return calculation understates them. Risk work in the NIST sense spans trustworthiness, design, evaluation and use. A workflow built to catch errors before they reach a customer, or to keep a record of who approved what, is buying reduced downside. Its value is the incident that does not happen, which will never appear in a spreadsheet. The honest way to present those is by exposure: what one bad outcome would cost, and how often it currently nearly happens. Do not disguise them as efficiency projects. They compete for budget on different grounds. [Storytelling in the Age of AI and Automation](/blog/storytelling-in-the-age-of-ai-and-automation) is useful for making that case internally.
Section 6
The review that keeps the number honest
Measure once at ninety days and again at a year. The ninety-day number tells you whether it works. The one-year number tells you whether it stayed working, which is a different question and the one that decides your next investment. Four figures per workflow: hours consumed now versus baseline, correction rate, total cost including usage, and what happened to the freed capacity in specific terms. Publish them internally even when they are disappointing. A company that can say this one returned nothing and here is why is a company that will get the next three right. One that only reports wins is accumulating unexamined systems.