To measure AI ROI, tie every AI project to one number in revenue, cost or cash, record a baseline before you start, compare results against that baseline or a control group, and review the value actually realised every quarter. If a project cannot name its number up front, it cannot prove a return later.
- Every AI project must move revenue, cost or cash — name which before starting.
- Without a baseline captured beforehand, any result is opinion rather than evidence.
- Use a control group or side-by-side run to separate AI impact from luck.
- Count the full cost: software, data work, change management and running costs.
Why AI ROI is hard to prove
Many AI projects report activity — models built, users trained, dashboards launched — instead of business results. Others claim credit for improvements that would have happened anyway, such as a seasonal sales lift. Both make the board sceptical.
The fix is discipline, not a cleverer formula: one number per project, measured the same way before and after.
Tie each AI project to revenue, cost or cash
Translate every use case into money. Better forecasts mean fewer lost sales (revenue) and less excess stock (cash). Faster support means fewer hours per ticket (cost) and better retention (revenue). Predictive maintenance means less downtime and lower repair bills.
Pick the one effect you will be judged on, and agree its definition with finance before the project starts.
Baseline and attribution: proving AI made the difference
Record a baseline over a normal period — typically three to six months of history. Then compare like with like: AI-planned products against similar products planned the old way, or one region against another. Where a split is impossible, run AI and people side by side and log which recommendation was followed.
Adjust for known events such as price changes, promotions or a large new customer.
Value-realisation reviews after go-live
ROI is not a one-off calculation at launch. Hold a quarterly review with finance: what was promised, what was realised and why any gap exists. Track adoption too — a recommendation nobody follows delivers nothing.
This is why Owllys puts the target outcome in writing before a project starts.
Example: measuring a forecasting project against its baseline
Illustrative example.
How to measure the return on an AI project
- 01Name the numberChoose one revenue, cost or cash measure and agree its exact definition with finance.
- 02Capture the baselineRecord three to six months of performance before the AI goes live.
- 03Set up a comparisonUse a control group, a pilot region or side-by-side runs to isolate the impact.
- 04Count the full costInclude licences, data work, integration, training and ongoing running costs.
- 05Review value quarterlyCompare promised with realised value, track adoption, then decide to scale, fix or stop.
Questions leaders ask
What is a good ROI for an AI project?
There is no universal figure. A sensible test is whether realised value covers the full cost — software, data work and change — within a period your finance team accepts, often 12 to 18 months. Projects that free cash tied up in stock can pay back sooner, because the effect shows on the balance sheet.
How do you measure the value of AI that only makes recommendations?
Log every recommendation and whether it was accepted, changed or rejected. Then compare outcomes for accepted recommendations against the decisions people made instead. Over a few months this shows both the value of the advice and how much is lost when it is ignored.
Should soft benefits like time saved count towards AI ROI?
Count them, but separately. Hours saved only become money if the time is redeployed or hiring is avoided, so finance will rightly discount them. Lead with hard effects on revenue, cost or cash, and present time saved and better decisions as supporting evidence.