How do you calculate AI project ROI?
Use the same period and currency for benefits and costs. The formula shows how much you gained relative to what you invested.
The Australian government's financial glossary defines return on investment, or ROI, as net profit divided by investment cost. It's expressed as a percentage.
ROI = (financial benefits − project costs) / project costs × 100. Include implementation and operating costs for the period in the denominator. Count expenses actually avoided and other attributable revenue as benefits.
The payback period answers a different question: when cumulative net savings cover the initial investment. If net cash flow is constant, divide the initial investment by monthly savings after operating costs.
Before implementation, ROI is an estimate based on assumptions. Afterward, replace those assumptions with observed costs and benefits. A positive result against the current process can still be worse than a simpler alternative.
What data do you need?
Assign someone to record each data point and keep the assumptions with the calculation. That way, you can review them once the system is in production.
- Describe the current process. Record case volume, time per case and costs you can avoid. Include errors and corrections. If volume varies by season, compare similar periods.
- Separate capacity from savings. The National AI Centre suggests valuing time saved using staff costs. It notes that the time creates value when it's used for useful work. Record that potential separately from money you'll actually stop spending.
- Gather the costs. Include implementation, data preparation, testing and training. Add licenses, usage, oversight and corrections during operation. If a fee already includes support, don't count it again.
- Build scenarios. Calculate a conservative, an expected and a favorable scenario. Change assumptions such as volume, usage and correction costs. Document which figures are observed and which are estimates.
Which alternatives should you compare?
Calculate each option's total cost over the same period and workload. Keep the quality each option delivers visible.
| Option | When it makes sense | What to review |
|---|---|---|
| Current tool | Its rules and features can handle the task. | Configuration, limits and remaining manual work. |
| Automation platform | The workflow connects tools and follows known rules. | Licenses, integrations and maintenance. |
| Custom AI development | The task requires interpreting content and the other options aren't enough. | Data, evaluation, oversight and operation. |
Cost and timeline depend on scope, integrations and data quality. They also depend on the level of human review needed.
How does the calculation work in an example?
This example is hypothetical. It doesn't describe a client or predict another project's results.
A team receives emails about outstanding payments. Each message includes a request and attachments. The AI system classifies the case and drafts a response; a person reviews it before sending.
For comparison, the company assumes the same volume and acceptable quality across the options. The current process costs COP 6 million per month. Automation without AI would cost COP 3 million to implement and reduce the monthly cost to COP 5 million.
The AI option would cost COP 12 million to implement. It would avoid COP 3 million in existing monthly expenses but add COP 1 million in monthly operating and oversight costs. Those expenses are actually avoided; they're not a valuation of time saved.
Over 12 months, AI's gross benefits would be COP 36 million. Its costs would be COP 24 million: COP 12 million for implementation and COP 12 million for operation. Estimated ROI against the current process would be (36 − 24) / 24 × 100 = 50%.
Net monthly savings would be COP 2 million. At that constant rate, the initial investment would pay for itself in six months. If benefits start after implementation, add that wait to the payback period.
The conclusion changes when compared with the alternative without AI. Over 12 months, automation would save a net COP 9 million and AI would save a net COP 12 million. Both figures are compared with the current process. AI would provide COP 3 million more, but require COP 9 million more upfront.
The incremental return would be about 33% on the additional COP 9 million. The additional investment would pay for itself in nine months because AI would save COP 1 million more per month. This comparison depends on both options meeting the agreed quality standard.
In a conservative scenario, lower usage or more corrections could reduce net savings. If cash flow changes each month, add the monthly net flows until they cover the initial investment. Don't divide by an average.
How do you verify returns and quality in production?
The financial calculation depends on reliable results. A fast response that needs many corrections can erase expected savings.
- Prepare evaluation cases. Gather common requests and difficult cases, with expected responses and protected data. The Google Cloud documentation recommends diverse examples aligned with the task. Compare the current process and the system using the same criteria.
- Agree on a threshold. Define what makes a response correct, which errors are unacceptable and when a person steps in. The Australian implementation guidance recommends agreeing on acceptance criteria and tests before deployment.
- Compare attributable results. Record cases resolved, total time, corrections and expenses avoided. When possible, compare results with a similar group using the previous process. Check for changes in volume or staffing before attributing the difference to AI.
- Review and correct. Sample responses in production and rerun the evaluation set when the data or system changes. Define who can stop the workflow and how cases will be handled if it fails.
