Why most AI ROI figures do not hold.
Three errors recur. The value counted cannot be audited: productivity gains are declared rather than measured. The cost counted is incomplete: the infrastructure is priced, while talent, governance, integration and change are left out. And the attribution is loose: gains that market trends or other projects would have produced are credited to AI. Each error inflates the figure; together, they make it impossible to defend.
The value side: Total Economic Value.
The Odyssey AI ROI Framework (TEV-AI) establishes value across four dimensions, each on its own evidence.
Direct value — gains measurably attributable to the initiative: cost removed, revenue added, working capital released.
Indirect value — effects on adjacent processes and decisions: quality, cycle time, error rates, risk reduced.
Option value — the right, without the obligation, to pursue future uses the initiative makes possible: data assets, reusable components, skills acquired.
Non-use value — the value of a capability held without being exercised: readiness, resilience, competitive position defended.
The cost side: TCO-AI.
The total cost of ownership of AI covers three blocks: infrastructure (models, inference, platforms), talent and governance (the people who build, run, supervise and audit the system), and indirect costs (integration, change management, process redesign, compliance). Infrastructure is usually the smaller share, which is why cases built on the cloud bill alone overstate the return.
Hard ROI, kept apart from Strategic Upside.
The return is reported in two separate figures. Hard ROI counts only direct and measurable value against the full cost: it alone founds an investment decision. Strategic Upside records option and non-use value: it informs the decision without ever inflating the stated return. A board can bank on the first; it chooses whether to pay for the second.
The method, in five steps.
01 — Map the workflow the initiative changes, end to end.
02 — Define the outcome metrics that matter to the business: cost, revenue, cycle time, error rate, risk.
03 — Fix the pre-AI baseline, before deployment or reconstructed from records.
04 — Measure after deployment, over at least one full operating cycle.
05 — Attribute the value, separating what AI produced from what other factors would have produced anyway.
The method is applied in depth to two or three high-value workflows, then extended to the rest of the portfolio as a quarterly routine.
Frequently asked questions.
How long before the ROI of AI can be measured?
As soon as a baseline exists and one full operating cycle has passed after deployment. For two or three workflows with clean data, a defensible figure takes a few weeks; wider scopes or thin data take longer.
Should productivity gains count as ROI?
Only once they are converted. Time saved becomes Hard ROI when it is removed from cost or redeployed into revenue. Until then, it is capacity: recorded, but not banked.
Who should measure the ROI of AI?
A party with no interest in the result, under the authority of the finance function. The vendor or integrator that delivered the initiative cannot credibly certify its return.
What if the measured ROI is negative?
It is a valid finding. The initiative is then redesigned, rescoped or stopped, and the capacity redeployed. A method that cannot produce a negative answer is not a measurement.
Can ROI be measured across a whole AI portfolio?
Yes, by applying the same method workflow by workflow. Most organisations start with the two or three highest-value workflows, then run the measurement every quarter across the rest.