Calculating the ROI of an AI agent: a simple model

Most AI ROI presentations fail for a boring reason: nobody measured anything before the project. So the « gain » is compared to a guess, and the CFO is right not to believe it. The fix is a model simple enough to fit on one line, applied with a little discipline.

The formula

Monthly ROI = (hours saved × loaded hourly cost) + revenue recovered − subscription. Hours saved: the tasks the agent completes, valued at the time they took a human before. Revenue recovered: what only exists because the agent acted — the follow-up that got an invoice paid, the lead answered within the hour instead of within the week. Then subtract the full monthly cost of the agent. No projected productivity, no « employee experience » line — things a finance director can check.

Measure the baseline before, not after

The whole model rests on one prerequisite: measure before deploying. One week of honest measurement is enough — how long the tasks take, how many there are, what falls through the cracks. Skip this and every later number becomes an estimate against an estimate. It is the single most common mistake we see, and it is irreversible: you cannot measure the « before » afterwards.

Once several agents run, the same logic scales up: sum the saved hours across the fleet and express them in full-time equivalents, the unit executives actually read. We published our exact methodology for that fleet-level measure separately — the principle is identical, only the scale changes.

Actionable version: open a spreadsheet with four columns — task, occurrences per month, minutes per occurrence, hourly cost. Fill it this week for the process you are considering. Multiply, compare to the subscription you were quoted, and you have your answer before talking to any vendor.