The unit is an accepted deliverable
For a specification or a report, faster drafting is only one part of the work. Count preparation, prompting, checking, corrections, failed attempts and final review. Use comparable tasks and the same acceptance standard. Record tool and version, task complexity, user experience and the observation period.
A worked example, not a measured result
Illustrative example, not client work. The following figures are invented to demonstrate the calculation. They are not Elvaris client results or experimental observations.
Without AI: six hours of preparation and drafting, plus two hours of review, equals eight human hours. With AI: three hours of preparation and drafting, plus four hours of review and correction, equals seven human hours. The arithmetic difference is one hour, or 12.5% of the eight-hour baseline, not the three hours suggested by looking at drafting alone.
That example does not establish causal improvement or comparable quality. A real assessment needs acceptance results, adequate observations and a comparison design that addresses differences in task difficulty, learning and other changes.
Keep three kinds of value separate
An hour released is potential capacity, not automatically cash saved. Evidence is needed to establish what happened to that capacity.
- Capacity: total human effort per accepted output and how released time was used.
- Financial effect: realized or substantiated avoided expenditure, with all incremental AI costs counted for the same period.
- Service and quality: acceptance, corrections, errors, turnaround and material incidents.
Before implementation: forecast scenarios
Use conservative, central and optimistic scenarios, stating assumptions for volume, adoption, review effort, quality and costs. Show the conditions required to break even. A forecast must stay labeled as a forecast; it is not a measured return.
After implementation: test the claim
Collect actual records and compare them with an appropriate baseline or counterfactual. Include licenses, integration, training, maintenance, oversight and rework as applicable; do not count the same cost or benefit twice. Keep capacity value separate from cash effects.
For a defined period, financial ROI can be expressed as (incremental monetary benefits minus incremental AI costs) divided by incremental AI costs. The measure is undefined when the denominator is zero. If benefits remain estimates, label the result as estimated rather than realized ROI.
What a defensible conclusion sounds like
In an actual assessment, the available evidence might support lower effort while leaving quality or financial effects unresolved. Report those separately instead of converting incomplete evidence into a single success score. A comparison may also show no improvement or a deterioration.
GAO's August 2026 AI Strategy provides an example of an organization planning to track costs, performance and outcomes and to define evaluation criteria when pilots are approved. That is GAO's own strategy, not a government-wide requirement to buy an external ROI review.
Scope and interpretation
The calculation illustrates a measurement approach, not a measured return or a guarantee of savings. Conclusions depend on the quality of the baseline, comparison design and available records. Financial, statistical or technical conclusions beyond the engagement's competence require appropriately qualified support.
Sources and their limits
- GAO AI Strategy, August 2026
GAO's own approach to tracking AI costs, performance and outcomes; not a universal mandate or a source of customer baseline values.
- GAO-21-519SP: AI Accountability Framework
Accountability context. The illustrative calculation and measurement procedure are Elvaris's, not a GAO-issued ROI formula.