Measuring AI ROI: A Practical Framework

Adoption is a property of the tool. Return is a property of the workflow. Enterprises measure the first because it is easy and hope it implies the second. It does not.

Most large enterprises can now produce a detailed account of AI adoption: seats provisioned, weekly active users, prompts per employee, pilots in flight. Far fewer can produce a defensible account of AI return. The gap is not a reporting problem. It is a measurement design problem, and it has two specific causes: almost nobody captures a baseline of workflow economics before rollout, and almost nobody commits to an attribution method that separates AI-driven change from everything else that changed at the same time.

This paper presents the framework we use to close that gap. It has four parts. First, baseline the workflow before automation, using six measures that take days, not months, to capture. Second, choose an attribution method before rollout: a holdout comparison where the work is divisible, a before-and-after design with an explicit confound ledger where it is not, and a modeled estimate only where effects are too diffuse for either. Third, account for total cost of ownership, including the lines most models omit: inference spend, permanent human review time, rework on AI errors, and ongoing maintenance. Fourth, report the result on one page, in a format finance can audit.

A worked example runs through the paper: an accounts payable automation whose vendor pilot implied roughly $590,000 a year in savings, and whose measured, attributed, fully costed return is a net $332,000 a year. The second number is smaller. It is also real, and it survives scrutiny. That is the trade this framework asks you to make.