How to calculate the ROI of an artificial intelligence project

How to calculate the ROI of an artificial intelligence project

Your team spends hours answering enquiries, preparing information or following up sales opportunities. Automating some of that work sounds sensible. But before commissioning a project, the question that matters is: what would need to improve to make the investment worthwhile?

Answering it means understanding the current process, adding up the full cost of the solution and distinguishing released capacity from money actually saved. An attractive percentage is of little use if it depends on assumptions nobody has checked.

This guide builds a straightforward calculation using a hypothetical example. The figures are not CIVIA prices, forecasts or client results.

1. Define your comparison and time horizon

Return on investment, or ROI, relates an investment's net benefit to its cost. See the definition in the business.gov.au financial glossary.

For this analysis, we will use:

ROI (%) = [(monetary benefits over the period − project costs over the period) / project costs over the period] × 100.

Compare the project with what would happen without it. Use the same horizon for costs and benefits; for example, twelve months of operation, including the initial investment required to launch.

Forecast ROI depends on assumptions. Measured ROI uses observed results and a credible comparison. The two are not interchangeable.

2. Measure what the process costs today

Choose a specific workflow: answering WhatsApp enquiries, classifying requests or preparing quotes. If you need help prioritising, start with our guide to business processes that AI can automate.

Record volume, human working time per case, hourly cost and error frequency. Separate waiting time from staff working time: a response taking two days does not mean it consumes two working days.

Include reviews, handovers and corrections. Use a representative period and note campaigns or seasonal peaks. Apply a consistent hourly cost basis: full employment cost to value capacity, and genuinely avoidable expenditure to calculate savings.

3. Separate capacity, savings and additional revenue

Released capacity: hours the team can spend on other tasks. Explain how they will be used. Australia's National AI Centre also connects the value of recovered time with its use for useful work.

Verifiable savings: expenditure reduced against the no-project scenario, such as paid overtime or external services that are no longer needed.

Additional margin: if sales follow-up improves, count the margin on incremental sales attributable to the change, after their variable costs; not total revenue.

Avoid counting the same effect twice. If released hours enable more sales, do not automatically add their labour value to the margin earned. Document other potential causes of improvement too: a campaign or price change may have an effect.

4. Include every project cost

Separate the initial investment from running costs:

• Set-up: analysis, configuration, integrations, information preparation, testing, training and internal launch time.
• Operation: licences, model and messaging usage, hosting, support, maintenance and content updates.
• Oversight: human review, exception handling, corrections and incidents.

If review time is already deducted from time saved, avoid charging the same cost again separately. Add an explicit contingency allowance where appropriate and check which items vary with volume.

Also compare with a manual process improvement or rules-based automation. Our article on mistakes when implementing AI explains why choosing the tool first can lead to unnecessary investment.

5. Hypothetical example: handling enquiries

Assume a process with these conditions:

• 2,000 monthly enquiries, requiring 6 minutes of human work each: 200 hours a month.
• Automation removes 60% of that work: 120 gross hours.
• Reviews and exceptions take 20 hours: 100 net hours released.
• At €25 an hour, their capacity value would be €2,500 a month.

However, only 40 of those 100 hours eliminate overtime previously paid at a total cost of €25 an hour. Monetary savings are €1,000 a month. The other 60 hours improve service and are excluded from this example's monetary ROI.

Now assume an initial investment of €6,000 and running costs of €400 a month. The initial investment includes launch work; the €400 covers additional operating expenses. Internal review is already deducted from net hours. We add no sales or other benefits.

For twelve full months of operation at the assumed level:

• Monetary benefits: 1,000 × 12 = €12,000.
• Project costs: 6,000 + (400 × 12) = €10,800.
• Net benefit: 12,000 − 10,800 = €1,200.
• ROI: (1,200 / 10,800) × 100 = 11.1%.

This is an estimate under specific conditions, not a promise. If rollout is gradual, calculate each month separately: do not assume twelve months of full performance from contract signing.

6. Calculate when the investment pays back

With constant cash flows and no financial discounting, the simple payback period would be:

Initial investment / (monthly savings − monthly running cost).

In the example: 6,000 / (1,000 − 400) = 10 months of operation. Add implementation time when counting from the start of the project.

If net monthly cash flow is zero or negative, this formula does not yield a payback. For decisions spanning several years, extend the analysis with cash flows and net present value; simple ROI does not capture when each receipt or payment occurs.

7. Check what happens if results are lower

Keeping the example's costs unchanged, consider three assumptions for monthly monetary savings, not market benchmarks:

• Conservative: €700 a month. Annual benefits of €8,400 and ROI of −22.2%.
• Base: €1,000 a month. Annual benefits of €12,000 and ROI of 11.1%.
• Favourable: €1,300 a month. Annual benefits of €15,600 and ROI of 44.4%.

If the project only makes sense in the favourable scenario, review its scope or test a pilot. In this example, first-year break-even requires €900 a month in monetary savings: 10,800 / 12. Different costs or timelines change that threshold.

8. Turn the pilot into evidence

Before starting, agree on a baseline, who will measure results and the conditions for expanding, adjusting or stopping the project.

Compare similar cases and equivalent volumes, using a comparison group where feasible. Measure actual savings, human working time, correct resolution, reopened cases and handovers. A fast answer that forces the customer to contact you again may shift work instead of removing it.

For a local council, availability and accessibility can also justify public value. Report them through their own indicators without turning them into fictitious budget savings.

Practical worksheet: gather these details before requesting a proposal

1. Process and owner: where the scope starts and ends.
2. Volume and period: monthly cases and months being compared.
3. Current situation: human minutes per case, errors and avoidable expenditure.
4. Objective and boundaries: what should improve and what a person must review.
5. Expected costs: set-up, operation, internal time and contingencies.
6. Separate benefits: savings, additional margin and released capacity.
7. Assumptions: adoption, share of work suitable for automation and launch time.
8. Decision: minimum threshold, quality indicators and review date.

If a figure is missing, mark it as awaiting measurement. Do not replace it with an optimistic guess.

Frequently asked questions

What ROI should I require from an AI project?

There is no universal percentage. It depends on risk, timing, alternatives and organisational goals. Define the criterion before assessing the proposal.

Do hours saved count as financial savings?

They can be valued as capacity. They should only be presented as reduced expenditure when a cost is actually avoided.

Can I calculate ROI without historical data?

You can develop an initial hypothesis, but you will need to measure a representative sample of the process before treating it as a robust forecast.

Does positive ROI guarantee that I should go ahead?

No. Check quality, adoption, dependencies and the ability to maintain the service too. A positive calculation based on weak assumptions does not establish viability.

Is your process worth automating?

At CIVIA, we can work with you to analyse workload, current costs and the conditions a solution would need to meet to justify the investment.

Bring one specific process and whatever data you have. We will identify what still needs measuring and whether moving towards a pilot makes sense. Prepare for that conversation with our guide to implementing AI in a business.

Analyse my process with CIVIA