How to Calculate the ROI of AI Automation (Before You Spend a Dollar)

“Massive ROI” is the easiest line in enterprise sales. It is also the hardest line to defend unless someone has written down assumptions you can audit. For Australian owners and operators, the useful version of ROI is boring: hours or dollars removed from a process, minus what it costs to build and run the system, minus the hidden tax of your own team’s time during rollout.

This article sets out a simple annual framing you can run on one page before you sign. It is not a substitute for a detailed business case on large spend—but it is enough to separate a serious initiative from a science project.

The formula (in plain English)

Start with the labour you expect to remove or reallocate. Pick a weekly number you believe is achievable in production, not in a demo.

Annual labour impact (rough cut)
Hours saved per week × fully loaded cost of that hour × 52 weeks.

“Fully loaded” means more than the wage line. It includes super, payroll overhead, software seats, and—honestly—the opportunity cost of that person doing higher-value work. If you do not have a precise number, use a conservative band and label it. Precision can come later; direction matters now.

Subtract one-off implementation
Build, integration, configuration, data cleanup, training materials—whatever the vendor quotes as the project.

Subtract ongoing costs
API or model usage, hosting, licences, support retainers, and periodic re-tuning. If the vendor waves this away, that is a warning sign, not a compliment.

Subtract internal time
Workshops, UAT, fixing data, chasing exceptions. If this line is zero, the model is wrong.

What is left is a first-pass annual benefit. Divide implementation (and optionally first-year run costs) by monthly net benefit if you want a payback horizon.

A small sensitivity table beats a single optimistic cell. In a spreadsheet, run base / upside / downside on hours saved (for example −20% / base / +20%) and on run-rate costs. If the downside case still clears your hurdle, you have a robust decision. If only the upside case works, you are gambling.

A worked example (illustrative, not a promise)

Suppose you automate first-pass handling of routine inbound enquiries and internal routing. After a baseline fortnight, you believe you can reliably give back twelve hours a week of blended admin time. You model that time at $45 per hour fully loaded—again, illustrative.

Annual labour value ≈ 12 × 45 × 52 = $28,080.

Your implementation quote is $12,000 fixed fee. You budget $200/month for API usage and $500/month for a light support retainer—$8,400/year combined. You add 40 hours of internal time at the same $45 rate for UAT and process cleanup—$1,800.

First-year net ≈ 28,080 − 12,000 − 8,400 − 1,800 = $5,880. That is not headline-grabbing, but it is legible—and you can improve it if API costs fall with optimisation or if the time saved is closer to fifteen hours than twelve.

The point of the example is not the numbers. It is the structure: every line has a name, and nothing magical sits in the footer.

Hidden costs people forget

Model and API usage scales with volume. Document-heavy workflows and long email threads burn tokens. Ask for an estimate at your expected monthly volume, not at “typical.”

Maintenance is real. Integrations break when vendors change APIs. Accuracy drifts when your products, pricing, or seasonality shift. Someone must own monitoring—either in-house or via a support agreement.

Optimisation has a cadence. The system you launch in September is not the system you run in March without adjustments. Budget time or money for that work.

Payback as a scope test

If your rough payback stretches well beyond what your business can tolerate on cash flow, the problem is usually scope—not ambition. Shrink the workflow, prove the metric in production for ninety days, then expand. Large programmes that try to “do AI everywhere” routinely absorb cash before they prove a single number.

Also separate cash timing from annual benefit. A healthy first-year cash story might still be negative if implementation is front-loaded—that can be fine if the balance sheet allows—but you should see it explicitly. Boards and banks forgive honesty; they punish surprises.

A red flag worth taking seriously

If a vendor cannot help you build this sketch before you engage, one of three things is true: they do not know your process well enough yet (fixable with discovery), they do not want a number on record (concerning), or they sell demos rather than operations (walk away).

Riverstone Labs projects ROI before we start engagements—assumptions explicit, costs inclusive, oversight designed in. If you want the same structure applied to a workflow you already have in mind, book a free assessment and we will pressure-test the maths with you.

Related guides

Service capability:

Want this implemented in your business? Book a Diagnose call — free 30-minute consultation, no pitch.