Start with an economic hypothesis
Return is not the number of users or documents generated. It is the relationship between attributable benefits and the costs of a decision over a defined period. First write down the expected change, who benefits and the alternative against which it will be compared.
An experiment on professional writing tasks observed changes in time and quality under specific conditions. It is useful evidence of potential for those tasks, not a promise of organization-wide ROI.
Build a baseline and count full costs
Record volume, time per case, error rate, cost, quality and variation before the change. Include licenses, integration, data preparation, security, training, human review, support and maintenance. Also define costs that would arise without AI: compare against the best feasible alternative.
A simple formula is net benefit = realizable benefits − incremental costs. For period ROI, divide that net benefit by incremental costs; state the period and do not combine cash flows at different times without an appropriate financial method.
Released capacity is not always cash
Imagine a team releases 1,000 hours a year. If work is neither reduced nor reassigned to valuable activity, capacity is not automatically a payroll saving. It might absorb demand, improve service or reduce external hours; each hypothesis needs evidence and an owner.
Quality and risk benefits deserve monitoring even where assigning a precise monetary value would be misleading. Do not count both time saved and the margin from using those same hours as separate benefits.
Use scenarios and continuation decisions
Prepare conservative, base and favorable scenarios with explicit adoption, cost per use, quality and demand. Set warning measures and a follow-up review; adjust or stop if assumptions fail. The OECD highlights how data, skills and organizational changes influence adoption and outcomes.
Jevons can develop the economic case and agree how to test it. Delivery and outcomes depend on evidence and project-specific responsibilities.
Frequently asked questions
What if reliable data is not yet available?
Use a range of assumptions, identify the most sensitive variables and design an initial measurement. Do not present a point estimate as a proven outcome.
How often should the case be reviewed?
Agree a review at each relevant decision: before investing, after the test and during operation. Frequency depends on cost, risk and how quickly assumptions change.
Sources and further reading
Let’s discuss your next decision.
Tell us what is at stake, what remains uncertain and where a clearer strategic direction would help.
Start a conversation