AI ROI
How to calculate AI ROI
A complete method for calculating AI ROI: value drivers, the full cost model, a worked example, payback period and the measurement discipline that keeps the number honest.
Miguel Torres, Founder, Merjora · Published 12 August 2026 · Updated 3 September 2026
In short
Calculate AI ROI as (annual gross value − annual running cost) ÷ total first-year investment, where gross value is built bottom-up from volume × handling time × loaded hourly cost plus avoided error and rework, and investment includes build, integration, data preparation, evaluation, change management and the ongoing model, review and maintenance cost. Express the result as a range with a stated confidence level, and fix the baseline metric before go-live so realized value can be measured rather than asserted.
Key takeaways
- Bottom-up value only. Percentage improvements borrowed from vendor decks are not a baseline.
- Running cost is not just inference — human review is often the largest line.
- Report a range and a payback period, not a single ROI percentage.
- If you cannot measure the baseline today, the first deliverable is measurement.
The value side
Time released is not automatically cash. State explicitly whether hours convert into avoided hiring, redeployed capacity or absorbed growth — a finance director will ask, and 'productivity' is not an answer.
- Labour time released: volume × handling time × loaded hourly cost (salary plus employer cost plus overhead)
- Error and rework avoided: incident rate × cost per incident × expected reduction
- Cycle-time value: only count it where speed demonstrably converts into revenue or retention
- Capacity value: growth absorbed without adding headcount, counted as avoided hiring cost
- Revenue value: count only with an agreed attribution method, otherwise leave it out
The cost side people forget
| Cost line | Typically underestimated because |
|---|---|
| Data preparation and access | Integration work is scoped after the business case is approved |
| Evaluation and testing | Quality thresholds are agreed late, then require rework |
| Human review | Review time per item persists long after go-live |
| Change management and training | Adoption is assumed rather than planned |
| Model and platform running cost | Usage grows with adoption, not with the original volume estimate |
| Maintenance and drift monitoring | Treated as a project rather than an operating cost |
Worked example (illustrative)
A services company processes 1,800 inbound document-based requests per month. Each takes an average of 14 minutes to read, classify and re-key, at a loaded cost of €32 per hour.
Baseline: 1,800 × 14 ÷ 60 = 420 hours per month, or €13,440 per month — €161,280 per year.
With extraction plus review, handling time falls to an estimated 4-6 minutes. Value released is therefore roughly €92,000-€115,000 per year gross. First-year investment of €120,000 and annual running cost of €38,000 gives a first-year net of around €54,000-€77,000 and a payback period of roughly 13-16 months, improving materially in year two once build cost is behind you.
These figures are an illustrative model, not a client result. The point is the structure: an explicit baseline, a range, and a payback period rather than a single percentage.
Keeping the number honest after go-live
- Record the baseline metric for at least four weeks before launch
- Track handling time and exception rate weekly for the first quarter
- Report realized value against the original range, including when it lands low
- Recalculate running cost quarterly as volumes grow
Frequently asked questions
- What is a good ROI for an AI project?
- There is no universal benchmark worth quoting. A defensible target for mid-market process automation is payback inside 18 months with a value range whose lower bound still clears the running cost. Anything promising a specific multiple without your baseline is guessing.
- How do you calculate ROI for generative AI?
- The same way, with two adjustments: usage-based inference cost scales with adoption, and quality variance means you must budget human review as a permanent line rather than a temporary one.
- Should time savings count as ROI?
- Only when you can say what happens to the time. Avoided hiring, absorbed growth or redeployment to revenue work are countable; unattributed 'hours saved' is not.
Get the baseline before the business case
Merjora quantifies each opportunity from your own volumes and handling times, and shows the confidence behind every range.
Quantify your AI opportunitiesRelated reading
Editorial standard. Merjora publishes analysis, frameworks and publicly documented examples. We do not publish invented statistics, unattributed benchmarks or unverified customer stories.