The number most agent ROI calculators still hide

Nearly every AI automation proposal runs the same calculation: hours saved, multiplied by a loaded hourly rate, minus what the agent costs to operate. It is simple to verify, simple to present, and wrong in a specific, predictable direction.

What happens to the tasks that do not clear

No agent clears everything, and that is not an engineering failure, it is the nature of the work itself. The real question is what happens to the remainder and what that remainder costs. A failed task costs more than an untouched one, because a person now has to read the agent’s attempt, judge it wrong, and complete the job regardless.

At five hundred tasks a month and ninety-two percent accuracy, that works out to forty escalations. If each one takes twelve minutes rather than the original six, exception handling alone consumes eight hours a month, taking a headline first-year figure of fourteen thousand down to roughly eleven.

Why volunteering the number wins the work

Finance has typically seen the naive number before, sometimes more than once. Volunteering the exception line yourself does two things: it makes the rest of the estimate credible, and it survives the meeting where finance tries to break it. A smaller, defensible number beats a larger one that falls apart.