To calculate AI ROI, divide three-year realised benefit minus three-year total cost by total cost. Count net time saved after review, value it only as hires avoided, spend retired or signed gains, and confirm the case survives losing a third of its benefit.
Most AI business cases fail on one question from finance: when the tool saves an hour, where does that hour actually go?
A second checkout lane in a supermarket pays only if there is a queue. Without one, the new cashier is a cost that looks like progress. AI is similar: saved time is worth money only if it replaces work the business would otherwise pay for. This guide shows how to build an AI ROI calculation on realised value, and how to test it so that finance can sign it.
Why AI business cases fail finance review
The problem is rarely that AI does nothing. It is that the case measures the wrong thing. Most cases multiply a provider's pilot savings by headcount and salary, producing a figure that is gross, theoretical and unowned.
Gartner's survey of 644 organisations in late 2023 found that difficulty estimating and demonstrating AI value was the primary obstacle to adoption, cited by 49%. In July 2024 it predicted that at least 30% of generative AI projects would be abandoned after proof of concept by the end of 2025. Unclear business value and escalating costs were among the reasons.
- Gross time: minutes counted before anyone checks, corrects or redoes the output.
- Capacity as cash: freed hours counted as savings when no hire is avoided and no spend is cancelled.
- No baseline: nobody measured the task before the tool arrived.
- Pilot conditions: keen users and light review produce numbers a full rollout will not repeat.
- No owner: nobody is accountable when the benefit fails to appear.
The formula, and the figures to report beside it
ROI equals total benefit minus total cost, divided by total cost, over the same period. We recommend three years, because costs and adoption both change after year one. Take the cost from a separate total cost of ownership model, so that pricing detail does not cloud the benefit debate.
| Figure | How to calculate it | What it tells finance |
|---|---|---|
| Three-year ROI | (Benefit minus cost) ÷ cost, over three years | Whether the case is worth pursuing |
| Payback period | Months until cumulative benefit exceeds cumulative cost | How long the money is at risk |
| Break-even realisation | Three-year cost ÷ three-year modelled benefit | The share of benefit that must materialise |
| Margin of safety | 100% minus break-even realisation | How far the forecast can miss safely |
| Cost per successful task | Annual cost ÷ tasks completed without rework | Whether it beats today's unit cost |
Break-even realisation is the most useful figure and the least used. A case with 50% ROI and a 10% margin of safety is fragile. A case with 30% ROI and a 40% margin is robust, and finance teams fund robust cases.
How to measure the benefit honestly
Count net time, not gross. Net time saved is baseline time minus time with the tool, including checking, correcting and redoing discarded outputs. Measure it on your real tasks, scored by the people who do the work.
Then convert capacity into value with one test: identify what the business will stop paying for. Two answers count as hard savings: a planned hire avoided, or a named contract retired. Quality gains, such as fewer errors, count only when a leader signs them; otherwise they are upside.
A worked AI ROI calculation (hypothetical)
A 40-person support team handles 144,000 tickets a year at a measured 12 minutes each. In a pilot on real tickets, agents accept 85% of AI drafts, taking 8 minutes including review. The other 15% are discarded and take 13 minutes.
Net time is 0.85 × 8 plus 0.15 × 13, or 8.75 minutes. That is a saving of 3.25 minutes, not the 4 that accepted drafts suggest. It releases 7,800 hours a year, about 4.4 full-time roles.
The approved plan included 3 new hires, now avoided, worth USD 158,400 at a loaded USD 30 an hour. An overflow contract worth USD 30,000 a year is also cancelled. Annual benefit at full adoption is USD 188,400.
Year one realises 60%, so the three-year benefit is USD 489,840 against a modelled cost of USD 330,000. ROI is about 48% and payback arrives around month 16. Break-even realisation is 67%, a margin of safety of 33%.
Under stress, with 60% of benefit realised and costs 20% over, ROI falls to about minus 26%. The prudent answer is therefore a conditional yes: a price that holds at volume, a phased rollout and named kill criteria.
A go or no-go scorecard
Before the case reaches a budget committee, run five tests. A case that fails any of the first three is not ready, however attractive its headline ROI.
| Test | Pass | Fail |
|---|---|---|
| Baseline | Time, volume and errors measured on real work | Estimated from memory or provider figures |
| Ownership | Each hard saving has a named owner and date | Benefits belong to the project, not a budget |
| Robustness | Margin of safety near a third; survives a stress test | Positive only on base-case numbers |
| Payback | Inside the initial contract term | Depends on renewal at today's price |
| Kill criteria | Agreed thresholds that stop or shrink the rollout | No defined point to switch off |
Once the tool is live, re-measure the same figures each quarter: net minutes on a fresh sample, roles not filled, contracts cancelled and cost per successful task. The aim is not to defend the forecast. It is to learn early whether to expand, renegotiate or stop.
The bottom line
Conclusion
The second checkout lane is worth its cost only when there is a queue to serve. Hours saved are a possibility; money saved is a decision someone must own. A case built on net time, hard savings and a stress test is one finance can approve with confidence.
The same figures then become the quarterly test of whether to expand or stop. This week, choose the task you most want AI to help with and time twenty real examples. That baseline underpins every later calculation.
Frequently asked questions
What is a good ROI for an AI tool?
There is no universal figure; robustness matters more than size. Look for positive three-year ROI, a margin of safety of roughly a third and payback inside the first contract term.
How do I measure time saved by AI?
Time a sample of real tasks before and with the tool. Include reading, checking and correcting outputs, and the full time spent on outputs that are discarded.
Should freed-up employee time count as AI savings?
Only when it removes a cost or creates signed value, such as a planned hire no longer needed or an outsourced contract cancelled. Otherwise record it as capacity, not savings.
What is break-even realisation in an AI business case?
It is the share of modelled benefit that must actually materialise for the tool to cover its cost: total cost divided by total modelled benefit.
How often should AI ROI be reviewed after purchase?
Quarterly is a practical rhythm. Re-measure net time per task, confirm the savings actually realised and compare cost per successful task with the baseline.
References
Disclaimer: General information only, not legal, security, compliance or financial advice. Check the cited sources and seek professional advice. Names belong to their owners; no endorsement implied.