AI Agent ROI Worksheet
A practical AI agent ROI worksheet for measuring cost per accepted result, human hours released, rework, model spend and quality retention.
ROI worksheet
Use this worksheet before and after a pilot. It focuses on accepted business outcomes rather than raw agent activity.
| Metric | Before pilot | During pilot | After rollout |
|---|---|---|---|
| Tasks/month | |||
| Human minutes/task | |||
| Accepted outputs % | |||
| Rework minutes/task | |||
| Agent/platform cost | |||
| Model/API cost | |||
| Review cost | |||
| Failure/recovery cost |
Three numbers that matter most
Cost per accepted result = all agent, API, review and rework cost divided by accepted outputs. Human hours released = baseline human time minus review/recovery time. Quality retention = accepted results after automation divided by accepted results before automation.
Decision rule
Expand the agent only when cost per accepted result improves without unacceptable quality or risk. If the agent is cheap but creates large review or recovery overhead, the workflow is not genuinely automated.
A practical method for estimating AI agent ROI
ROI starts with the current workflow baseline. Record monthly volume, average handling time, loaded labor cost, correction rate, delays and any paid tools already used. If the baseline is unknown, a precise automation ROI percentage is false precision. Measure a representative period first.
Define what counts as an accepted output. For recruiting it might be a sourced candidate that meets explicit criteria and includes verifiable evidence. For support it may be a correctly resolved ticket. For research it may be a completed brief that passes review. The acceptance definition prevents teams from counting low-quality automated activity as value.
Estimate the future-state cost in layers: platform execution, model/API usage, third-party tools, human review and expected retries. Add one-time implementation work separately. If setup is a project cost, amortize it over a chosen period rather than hiding it inside the monthly run rate.
Time savings do not automatically equal cash savings. Some teams use saved hours to absorb more volume, improve response time or move specialists to higher-value work rather than reduce headcount. Record capacity value separately from direct budget reduction so stakeholders can see what type of return is expected.
Use confidence ranges instead of a single ROI number. Create conservative, expected and optimistic cases for volume, success rate and review time. A strong pilot should narrow those ranges with measured data. If the economic case only works when every assumption is optimistic, deployment risk is high.
After launch, replace assumptions with observed values. Track accepted outputs, total execution cost, review minutes, retries, failure recovery and turnaround time. Recalculate monthly so changes to models, tools or workflow design do not quietly erode the original business case.
Frequently asked questions
What is the best ROI metric for an AI agent?
Cost per accepted outcome is usually more decision-useful than raw execution cost because it includes quality and review. Pair it with turnaround time and the business value of increased capacity.
Should time saved be valued at the employee hourly rate?
Only if that reflects how the saved capacity will actually be used. Distinguish direct budget savings from productivity or capacity gains.
How long should an AI agent pilot run?
Long enough to include representative normal and edge cases. Define the sample size and success criteria before starting rather than stopping as soon as early results look favorable.