Section 1
Why most ROI stories collapse under questioning
Three questions do the damage. Compared to what? Over what period? Net of which costs? Compared to what is the counterfactual. If the sales team also hired two people and changed the pricing page that quarter, the automation cannot claim the whole delta. Over what period exposes seasonality: a support automation launched in a quiet month will look extraordinary until the busy month arrives. Net of which costs is where most published numbers quietly fail, because the build hours, the review hours, and the ongoing usage bill are treated as free. None of this means the return is not real. It means an unqualified number invites the reader to discount all of it, which is a poor trade for a founder who did the work.
Section 2
Capture the baseline before you build, because you cannot go back
The baseline is perishable. Once the workflow changes, the old timings are gone and you are reconstructing them from memory, which flatters the result and everyone knows it. Spend a week measuring before anyone builds. For the target workflow, record volume per week, elapsed time from trigger to completion, touch time versus waiting time, rework rate, and the error rate that reached a customer. Pull what you can from timestamps in the CRM, helpdesk, or accounting system rather than asking people to estimate, because self-reported time is consistently wrong in the direction of the story people want to tell. Store that snapshot somewhere unedited. It becomes the before column, and later it becomes the reason a sceptical CFO believes the after column.
Section 3
The four-line comparison
Any honest before-and-after fits on four lines: what the work cost before, what the automation cost to build and to run, what the work costs now, and what else changed in the same window. The model below lays that out with the inputs for each line. For an example of the same discipline applied to a whole company, read [Startup Success: How AI Automation Transformed Our Business](/blog/startup-success-how-ai-automation-transformed-our-business).
Section 4
Running the comparison so the result means something
Run the old and new paths side by side for a fortnight if you can. Parallel running is unglamorous, costs a little, and removes almost every argument about whether the improvement is real. Pick a measurement window long enough to include a bad week. Report elapsed time as a median and a worst decile, because averages hide the tail where customers actually get annoyed. Where a metric moved for several reasons, say so and split the credit roughly rather than claiming the whole thing. Then convert to money only where money genuinely changed hands or changed timing. Faster invoicing that shortens days sales outstanding is cash. Faster quoting that raises win rate is revenue. A reclaimed hour that got absorbed into the day is a capacity gain, and should be labelled as one.
Section 5
The costs that never make it onto the slide
Build time is the obvious one, usually undercounted because it is founder time and founders do not invoice themselves. Then the recurring items: usage bills that scale with volume, the seat that gets added every time someone new joins, and the annual price rise you agreed to without reading. The subtler costs are operational. Review time on every output, which is real work. Exception handling, which lands on whoever is most conscientious. Maintenance when an input format changes upstream. And the cost of being wrong, which is small in an internal report and large in a customer invoice. Name the switching cost too. A workflow rebuilt inside a vendor's product is cheap to run and expensive to move. That is not a reason to avoid it. It is a reason to know the number before the renewal conversation, and it is why [The Rise of Autonomous Startups: Fully Automated Companies](/blog/the-rise-of-autonomous-startups-fully-automated-companies) is worth reading before you go deeper.
Section 6
The numbers worth publishing
Four hold up in front of an outsider: elapsed cycle time before and after, cost per transaction fully loaded, rework or error rate, and volume handled per week without adding people. Two supporting numbers make them credible: the measurement window, and what else changed during it. Publish the ones that went sideways as well. A before-and-after with one honest disappointment in it is read as a report. One with only wins is read as marketing, and gets discounted accordingly.