AI automation ROI calculator, with every assumption shown.

Any ROI calculator is an argument dressed as arithmetic, so here is the argument in full. Six assumptions sit behind the four numbers above. Three of them flatter the result, one deliberately understates it, one is a definition you supply, and the last is a promise about your data rather than a calculation.

The one worth arguing about is the 80%. Vendors in this market routinely model 95% or 100% automation, which is how a business case survives the pitch and dies in month three. In every workflow we have shipped, a stubborn tail of exceptions stayed with a person, and that tail is a feature: it is where the judgement lives. Model it, price it, and the rest of the number holds up.

Coastline Logistics ran order operations by hand: fourteen steps across three teams, two days end to end, and 698 hours a month measured from their own timesheets. After a six-week build it became a single automated run of three minutes and fifty-six seconds, with only genuine exceptions reaching a person.

That is the number to put through the calculator, and it comes with a catch worth understanding, because it is the same catch in your own estimate.

The fourteen steps take ten hours and fifty-five minutes of elapsed time per batch, and four of those hours are spent waiting for supplier confirmations while nobody works. Strip the waiting and 63% of the elapsed time is active labour: 698 hours a month becomes 442. Waiting hours are real and they are worth removing, but they do not cost salary the way active hours do, so a cash business case built on the full 698 is overstated.

Both columns below are honest. The left one is what the published figure gives you; the right one is what a finance director will accept. Coastline did not publish its loaded hourly cost, so $40 is used as an illustration in both columns rather than as their number.

Running costs are the first omission. An automation of this size costs something to keep alive every month: model API calls, platform seats, monitoring, and occasional engineering when a vendor changes an interface. Between $150 and $600 a month covers most of what we ship. Subtract it before you show the payback figure to anyone who controls a budget.

The second is your own time during the build. Someone on your side answers questions, tests the run and signs off the exceptions. Two to four hours a week for the length of the engagement is normal, and pretending otherwise is how automation projects acquire a reputation for being disruptive.

The third is the upside, which the calculator also ignores. Coastline stopped being two days late on dispatch, and their weekly report went from a day out of date to current on every run. Neither shows up in an hours-saved figure, and for some workflows the cycle time is worth more than the labour. The calculator is deliberately the pessimistic half of the case.

Start by writing the workflow down step by step, with the person and the duration on each line. Nobody guesses this correctly from memory, and the list is always longer than the person doing the work expects. That list is also the artefact an agency needs to quote accurately, so the hour it takes pays for itself twice.

Then split the hours into active and waiting, put the active figure through the calculator, and subtract a year of running costs from the saving. What you are left with is a defensible annual number and a payback period in months. If payback lands inside a year, the project is usually worth doing. If it lands past two, either the workflow is too small to automate or the wrong part of it was measured.

Bring both to the first call. An agency that argues your number down rather than up is telling you something useful about how it prices. We will send back the version of the arithmetic we believe, including the cases where our answer is that the workflow is not worth automating yet.