How to Calculate the ROI of AI Automation (With Worked Examples)

Calculate AI automation ROI as: (hours saved per year × loaded hourly cost of the staff time) minus (one-time build cost plus annual run cost), then divide the build cost by the monthly net savings to get the payback period in months. This treats AI automation like any other capital investment, with explicit, labelled assumptions rather than optimistic guesses.

Most ROI claims in this space skip the run costs or use unrealistic time-saved estimates. This guide walks through the formula and two fully worked examples with every assumption stated, so you can plug in your own numbers rather than relying on generic percentages from a vendor pitch.

What is the formula for AI automation ROI?

The formula is: Net annual value = (hours saved per year × loaded hourly cost) − annual run cost − amortised build cost; Payback period (months) = build cost ÷ (monthly savings before run costs, minus monthly run cost).

• Hours saved per year: realistic time freed up, not the theoretical maximum if the automation were perfect.
• Loaded hourly cost: salary plus benefits and overhead, typically 1.25–1.4x base salary, divided by annual working hours (usually ~2,000).
• Build cost: the one-time development cost, amortised over the useful life if comparing to ongoing savings.
• Annual run cost: LLM API usage, hosting, monitoring and maintenance, which recur every year the automation runs.

It helps to write these four numbers down explicitly before running the calculation, and to date-stamp the assumptions. Hours saved, in particular, is where most ROI estimates go wrong, either by assuming full automation of a task or by ignoring the time staff will still spend reviewing edge cases the system can't handle confidently.

It's also worth calculating a second, more conservative scenario alongside your main estimate, using a lower hours-saved figure (say, 60% of your first estimate) to see how sensitive the payback period is to that assumption. If the automation still pays back within a reasonable window under the conservative case, that's a much stronger basis for a go/no-go decision than a single optimistic number.

Before running the formula, it's worth spending a week or two actually timing the current manual process, rather than estimating from memory. Staff often under- or overestimate how long a task takes, and a short time-tracking exercise, even an informal one using a spreadsheet, produces a far more defensible hours-saved figure than a guess.

Worked example 1: AI invoice processing automation

Assumptions: a finance team of 2 people spends 15 hours a week combined manually entering and checking supplier invoices; average loaded hourly cost is $35; the automation reduces that to 3 hours a week combined for review and exceptions; build cost is $6,000; annual run cost (LLM API, hosting, maintenance) is $1,800.

This example is a strong candidate for automation precisely because the process is well-structured (invoices arrive in predictable formats) and high-frequency (weekly, ongoing), which is exactly the profile that tends to produce fast payback periods. Slower, less structured or infrequent processes rarely look this good on paper.

Worked example 2: AI customer support chatbot

Assumptions: a support team currently spends 25 hours a week on tickets the chatbot could resolve; average loaded hourly cost is $28; the chatbot resolves 60% of that volume without human involvement, so 15 hours a week are saved; build cost is $18,000; annual run cost is $4,200.

Why does example 2 have a longer payback period?

The chatbot costs three times more to build and run than the invoice automation but saves only a similar amount in year one, which is common: customer-facing conversational agents typically need more build effort and ongoing token spend than a narrow backend workflow, since they handle open-ended conversation rather than a fixed document format.

How should you weigh non-financial benefits?

Not every benefit shows up cleanly as hours saved. Faster response times, more consistent answers, and the ability to handle volume spikes without hiring temporary staff are real value but harder to price precisely. A reasonable approach is to calculate the core ROI on hours saved alone, as shown above, and list these secondary benefits separately as qualitative factors that strengthen the case without being folded into the headline payback number, which keeps the calculation honest and defensible.

Error reduction is one non-financial benefit worth quantifying separately where you can. If a manual process currently produces a known error rate that leads to rework, refunds, or customer complaints, and you can estimate the cost of fixing an average error, that figure can be added to the ROI calculation as its own line item, provided the error-rate improvement from automation is a realistic, stated assumption rather than an assumed elimination of all errors.

Revenue-side benefits, such as an AI agent responding to leads faster and improving conversion rates, are real in many businesses but even harder to isolate cleanly, since conversion rate is affected by many factors beyond response time alone. Where you have historical data linking faster response to higher conversion, it's reasonable to include a conservative estimate of that effect; where you don't, it's safer to treat it as a qualitative upside rather than a number in the payback calculation.

What mistakes make ROI estimates unreliable?

• Using base salary instead of loaded cost, which understates true savings by 25-40%.
• Assuming 100% of a task is automated when a realistic estimate is 50-80%, with humans still handling exceptions.
• Ignoring annual run costs entirely, which overstates net savings every year, not just in year one.
• Using best-case adoption rates instead of a phased, realistic rollout, especially in the first few months.
• Skipping the time cost of managing and correcting the automation itself in its first weeks of operation.

A related mistake is assuming the saved hours translate directly into cost savings, when in practice staff often redirect that freed-up time to other valuable work rather than being let go or having their hours reduced. That's not a reason to skip the calculation, but it's worth being explicit with stakeholders about whether the ROI case is framed as "cost reduction" or "capacity freed up for higher-value work," since the two are different claims with different implications for headcount planning.

Finally, be cautious with ROI figures presented by a vendor or agency without their underlying assumptions shown. A payback period is only as credible as the hours-saved, loaded-cost and run-cost numbers behind it, so ask any agency quoting an ROI figure to show their working, the same way this article does, before treating the number as a basis for a decision.

How does this compare to typical project cost ranges?

Check your own build cost estimate against typical ranges for your region and project type before running this calculation. See AI automation cost in India, AI automation cost in Dubai and the UAE, or AI agent development cost depending on your market.

Next step

We build ROI estimates with clients before scoping a project, using their real staff costs and volumes rather than generic figures. See our AI automation services or book a discovery call to work through your numbers.