The most common reason an ROI claim collapses is that the "before" figure was estimated after the fact, by the people who wanted the project. Nobody is lying; memory is simply not a measurement instrument, and estimates made in hindsight bend toward the outcome.
Measuring a baseline is usually two weeks of light instrumentation: count instances, time a sample properly, record the error and rework rate. If that is impossible to arrange, it is worth asking how the saving will be demonstrated afterwards, because the same obstacle will still be there.
Twenty minutes saved across forty people is not fourteen hours of capacity; it is forty people with a slightly easier day. That is a real benefit and it is not a financial one, and calling it one is what makes finance distrust the whole model.
Hours become money in three ways: a role that is not backfilled, a hire that is not made as volume grows, or capacity redirected to work that produces revenue. If a saving does not fall into one of those, report it honestly as a quality or capacity improvement and let it stand on its own.
Headcount avoided. The cleanest case: volume grew and the team did not. Defensible because the counterfactual is documented.
Cycle time. Days-to-cash, time-to-quote, time-to-first-response. Often worth more than the labour saving and easier to measure.
Error and rework cost. Duplicate payments, mis-shipments, penalties. Usually undercounted because nobody aggregates them today.
Capacity redirected. Legitimate, but only if you can say what the freed time is now doing.
A gross saving is not a return. Model and infrastructure usage, licences the flow needs, the exception desk that still handles the 20% that do not go straight through, and the maintenance the system needs when a supplier changes a format all belong in the model.
Maintenance is the line most often left out and the one that most often makes a marginal project negative. A reasonable planning assumption is a modest ongoing percentage of the build cost per year for a stable flow, more where the surrounding systems change often.
If including honest run costs turns the case negative, that is the model working. Far better to find it in a spreadsheet than in year two.
Every ROI model rests on a handful of assumptions: the straight-through rate, the volume forecast, the loaded hourly cost, the time an exception takes. State each one, with its source, in the model itself.
This does two things. It lets a sceptical reader argue with an assumption instead of dismissing the whole model, which is the difference between a conversation and a rejection. And it makes the number reconstructable later, when someone asks why the projection said €35,000 and the outcome was €28,000 — and the answer turns out to be that straight-through came in at 62% rather than 75%, which is useful information rather than an embarrassment.
The metric, its baseline, the measurement window and who reports it should all be fixed before the pilot runs. Choosing after the fact means the pilot cannot fail, which also means it cannot prove anything.
One number, agreed by both sides, measured the same way before and after. Everything else is supporting detail.