How to measure AI ROI: a framework for finance and IT.
A method finance and IT can share: what to count as cost, how to set a baseline, which benefits count, the formula, and when to decide.
- Topic
- AI ROI
- Product
- Novie Control
- Format
- Guide
- Reading time
- 3 min
- Published
- 5 November 2026
To measure AI ROI, take each AI workload on its own. Record a baseline before it starts. Count every cost of building and running it, agree with finance which benefits count and how they are valued, and divide net value by total cost. Review the result monthly alongside every other AI workload, and decide whether to scale, fix or stop each one.
Why one ROI number does not work for AI.
Gartner told CFOs in March 2026 that AI is a portfolio of investments with different cost and value profiles, and should be measured with a balanced set of indicators. A coding assistant, a customer support agent and a forecasting model cost different things and return different things. Deloitte found 85 percent of the organisations it calls AI ROI leaders use different frameworks or timeframes for generative and agentic AI.
TakeawayMeasure per workload and add the results up, because AI is many investments with different costs and returns.
Step one: count every cost.
The cost side is where most ROI calculations go wrong, because AI spend is spread across many budgets. BCG found that more than 80 percent of AI spending sits outside the IT budget. Gartner found about 11 percent of functions did not know what they spent on AI in 2025.
Usage-priced costs need the most care. A seat costs the same each month. Model usage moves with every task, every retry and every change of model, so it has to be read from the provider bills or a gateway, per workload, each month.
TakeawayCount licences, usage, cloud, build, people and compliance, per workload, from the bills.
Step two: set a baseline before anything changes.
Record how the work is done today: how long it takes, what it costs, its error rate and its volume. Without a baseline, any later number is an estimate of an estimate. If the work has already changed, use the nearest comparable team or period, and say so.
TakeawayA baseline recorded before the change is what makes the later number credible.
Step three: agree which benefits count.
| Benefit | How to value it | Counts when |
|---|---|---|
| Hours returned. | Hours times the loaded hourly cost. | The hours are redeployed to work with a value, or a planned hire is avoided. |
| Cost avoided. | The spend that would otherwise have happened. | There is a record of the alternative. |
| Revenue. | Margin on revenue the workload produced. | It can be attributed with a test or a control group. |
| Quality and risk. | Errors, rework or losses avoided, priced. | Finance agrees the price before measurement starts. |
Hours are the most common benefit and the easiest to overstate. Time saved by individuals becomes money only when it is used for something else. Agree a realisation rate with finance at the start, and keep to it.
TakeawayAgree with finance which benefits count and how they are priced before the first result comes in.
Step four: apply the formula.
AI ROI equals the value delivered minus the total cost, divided by the total cost, over the same period. Report payback period alongside it, since many AI workloads carry a large first-year build cost and a smaller running cost after that.
AI ROI = (value delivered - total cost) / total cost Payback period = one-off cost / (monthly value - monthly running cost)
TakeawayROI is net value over total cost, and payback tells finance when the first-year cost is earned back.
Step five: review monthly and decide.
Gartner found that organisations which track returns continuously, manage AI as a portfolio and cut underperformers saw positive returns on 81 percent of their AI initiatives. A monthly review of every workload's cost against its value, with an owner who answers for each, is what that looks like in practice.
TakeawayA monthly review that ends in a decision is what turns a measurement into a return.
How Novie Control does it
You can see what AI costs to build.Can you see what it costs to run?
Sources
- Gartner, CFOs Need to Rethink the ROI of AI Investments (24 March 2026)
- Gartner, Survey Finds Only 22 Percent of Organizations Have Successfully Scaled AI Across Multiple Business Units (1 September 2026)
- Deloitte, AI ROI: The paradox of rising investment and elusive returns (22 October 2025)
- BCG, AI Is Starting to Pay Off. Almost 50% of Companies Now Generate Value with It (30 September 2026)