You've received a proposal to automate recurring client reports. It promises to give your team hours back every month. What will you do with those hours?
If payroll stays the same, a shorter reporting process frees up staff time. You might use it for more client work, cut paid overtime or give an overloaded manager a manageable week. Those outcomes need different evidence before you can put a financial value on them.
To calculate AI automation ROI, you need the full cost of owning a specific workflow and evidence of the benefits your business can realise. If you're still deciding which workflow to tackle, start with choosing a measurable first workflow. For this calculation, assume you've chosen reporting and need to decide what you can afford to invest.
Measure the work through to acceptance
Decide what counts as finished before timing the work. For a client report, that might mean the right figures, current project context, checked recommendations and approval by the person responsible for sending it. A generated draft still needs work. Time both the baseline and the assisted workflow through to the same acceptance point.
Record monthly eligible volume and separate routine reports from difficult ones. Note who touches each report and count their combined human minutes across gathering inputs, preparation, checking, corrections, exceptions and manual fallback. Include failed attempts. If one person saves time and another spends longer fixing the result, count both.
Track elapsed turnaround separately. A report that arrives sooner might improve client service even when human labour barely changes. And removing ten minutes of work won't necessarily bring a deadline forward if approval still waits until Friday. Measure turnaround and labour separately.
Published productivity gains won't give you a forecast for this workflow. Brynjolfsson, Li and Raymond's Generative AI at Work studied 5,172 customer-support agents and found an average 15% increase in issues resolved per hour, with effects varying substantially across workers. That finding concerns a particular deployment and population. You'll need to measure your reporting process on its own terms.
In a different setting, METR's early-2025 experiment found experienced open-source developers took 19% longer with the tools studied, despite believing they had worked faster. Read that historical result alongside METR's February 2026 methodological update. The later experiment faced selection effects and unreliable task timing when developers used concurrent agents. Its raw estimates didn't establish a reliable current effect size.
For your own comparison, record the case mix, all human involvement and any exclusions. Use comparable conditions and count accepted output.
Separate the benefits you can actually realise
HM Treasury's Government Efficiency Framework distinguishes cash-releasing benefits, which reduce spending, from non-cash-releasing benefits, which can include staff time available for other work. This is a UK public-sector framework. We're borrowing its benefits distinction for SME planning. It sets no Australian accounting requirements.
For avoided spending, identify the expense you'll remove. You might cancel recurring contractor work or reduce paid overtime. Record how much spending will disappear and when. An avoided future hire needs a credible hiring baseline and evidence that the automation removes the need for that expenditure.
Recovered capacity means staff can do something else while payroll continues. It can be valuable even when spending doesn't fall. Multiplying hours by a loaded labour rate gives you a capacity valuation. You still need to establish a financial return. Our planning rule is to keep it outside the financial-benefit total until there's an evidenced spending change or productive use.
Incremental contribution comes from additional work the business can sell and deliver. You need customers for that work, staff available when they're needed and enough capacity elsewhere in delivery. Calculate the additional income after incremental delivery costs. Revenue alone overstates the benefit. Compare against the unchanged-payroll baseline so existing wages aren't deducted again simply because those staff now do different work.
Allocate each released hour once. Hours used for additional client work can't also be claimed as payroll savings. Improved quality, reduced stress or faster responses can justify investment on their own terms, but describe them separately unless you can defend their financial valuation.
Count the cost of owning the workflow
Start with the supplier's build price, then ask what else you'll need to pay for. What must your team supply? Who will be responsible after launch? Include the human attention a workflow still needs when defining ownership.
- Setup includes external implementation fees, data preparation, integration, internal implementation labour, training and rollout.
- Recurring costs include subscriptions, usage charges, support, maintenance, monitoring and additional ownership labour.
- Remediation includes expected investigation and correction work, with a clear allowance for incidents beyond ordinary handling.
- Exit costs include any required data export, handover or replacement work within the appraisal period.
Put routine per-report review, exceptions and rework in the assisted handling time. Add only oversight outside that measure to ownership overhead. Count that manager's checking time once. Use the same discipline for support fees and remediation allowances when a contract already covers the work.
State which appraisal you're making. An economic appraisal includes internal opportunity costs as well as external spending. A cash-flow appraisal follows actual changes in money paid and received. State whether costs include GST and use that treatment consistently. Ask for a written account of exclusions before treating an incomplete quote as a total cost.
Calculate the first year and the payback period
First-year ROI equals first-year realised benefit minus setup costs and first-year running costs, divided by setup costs plus first-year running costs, multiplied by 100. The denominator is the full cost of the first year. State whether the figures represent economic value or cash movements.
Monthly net benefit equals monthly realised benefit minus monthly running costs. Simple payback equals setup cost divided by positive monthly net benefit. If monthly net benefit is zero or negative, there is no simple payback under those assumptions.
The shortcut assumes steady monthly benefits and costs. Adoption starts immediately at the modelled level, and there's no discounting. It doesn't account for the timing of invoices or receipts. For a phased rollout, write a monthly schedule, subtract running costs from benefits in each month and accumulate the results. Payback occurs when that cumulative net benefit covers setup. Include implementation delays and training months in the schedule.
A worked example with every assumption exposed
The following example uses hypothetical Other Stuff planning assumptions in AUD, excluding recoverable GST. The figures aren't client results, market prices or productivity benchmarks, and none comes from the research percentages above.
Assume a business produces 400 eligible reports each month. Staff use the automation for 75% of them, so 300 reports follow the assisted workflow. Each currently takes 20 human minutes through acceptance. Assisted handling takes eight human minutes, including preparation, checking, exceptions and rework averaged across those reports. The remaining 100 reports continue unchanged.
Recovered capacity is 300 × (20 − 8) ÷ 60 = 60 staff hours a month. At an assumed loaded labour rate of AUD45 an hour, that's AUD2,700 of monthly capacity value. Payroll remains unchanged. Until the business establishes a spending reduction or additional contribution, recognised financial benefit is zero.
Assume setup costs AUD6,000, comprising AUD4,000 of external spending and AUD2,000 of internal implementation opportunity cost. Recurring economic cost is AUD900 a month. That comprises AUD600 for software, usage and support, plus AUD300 of additional ownership labour outside the per-report handling time.
First-year economic cost is AUD6,000 + (12 × AUD900) = AUD16,800. With no evidenced financial use of the time, the financial-benefit model shows zero benefit against that cost. Under our deliberately conservative recognition rule, first-year economic ROI is −100%, with no payback. The business has capacity available. It hasn't yet established a commercial use for it.
The calculation changes if the business can put that time to financial use. Suppose the business can demonstrate enough additional demand and usable scheduling for 40 of the 60 released hours to produce AUD60 of incremental contribution per hour. That contribution is after all incremental delivery costs, before automation setup and running costs, measured against unchanged payroll. The other 20 hours receive no financial value in this calculation.
Monthly benefit becomes 40 × AUD60 = AUD2,400. After AUD900 of running costs, monthly net economic benefit is AUD1,500. Assuming that level begins immediately and continues for 12 months, first-year benefit is 12 × AUD2,400 = AUD28,800. First-year net benefit is AUD28,800 − AUD16,800 = AUD12,000.
First-year economic ROI is AUD12,000 ÷ AUD16,800 × 100, or approximately 71.4%. Steady-state simple economic payback is AUD6,000 ÷ AUD1,500 = four months. Both results depend on the business actually delivering that conditional redeployment and meeting the quality threshold.
The AUD2,700 capacity valuation stays outside the ROI numerator. Payroll remains unchanged, and the incremental contribution uses that baseline. Cash payback needs its own schedule of actual receipts and payments, consistently excluding opportunity costs. You can't turn this economic payback into cash payback by changing only the setup figure.
Recovered time has no financial return until it has an evidenced use
- 60staff hours recovered each monthCapacity available in the hypothetical reporting workflow
- 2,700AUD of monthly capacity valueExcluded from the ROI numerator; payroll is unchanged
- 0AUD of recognised financial benefitBefore any evidenced spending reduction or extra contribution
Source: Hypothetical worked example in this article, before conditional redeployment; not client results.
Test the assumptions that decide the result
Even if the technology works, lower adoption can push payback further out. Hold all other costs and handling times fixed, but reduce adoption to 50%. Now 200 reports release 200 × 12 ÷ 60 = 40 hours a month. If demand and scheduling allow only 30 of those hours to earn AUD60 contribution each, monthly benefit is AUD1,800. Monthly net economic benefit is AUD900 and simple economic payback is AUD6,000 ÷ AUD900, or about 6.7 months.
Extra review time can push payback out too. Return to 75% adoption but let assisted handling rise to 14 minutes. The 300 reports now release 300 × (20 − 14) ÷ 60 = 30 hours. If all 30 can earn AUD60 contribution each, monthly benefit is again AUD1,800 and net economic benefit AUD900. With other inputs unchanged, simple economic payback is again about 6.7 months.
In the pilot, verify how the business will use the released time. Both scenarios depend on it. Test a higher recurring quote and a slower adoption ramp in your own monthly schedule. Check whether subscriptions or ownership labour change with volume instead of assuming fixed costs at every scale.
If faster reports contain material errors or clients reject them, the accepted-output assumption fails. The benefit depends on quality. Include correction effort and lost usable output before claiming a gain. For longer investment horizons, consider how competition can change the margin over time. A project that pays back under today's contribution assumptions may earn less if prices or demand change.
These are comparisons between assumptions in the same worked example. Longer bars mean a longer wait for economic payback; neither sensitivity case is an observed result.
Lower adoption or more review pushes payback beyond six months
Source: Hypothetical sensitivities calculated above. Setup AUD6,000; monthly net benefit AUD1,500 or AUD900; economic, not cash payback.
Give the pilot a decision rule
Agree a spending ceiling, minimum quality standard and maximum acceptable payback before starting. For frequent reporting work, a four-week pilot is one illustrative option. Low volumes, seasonal demand or unusual reporting cycles need a longer window to produce representative evidence. Keep enough evidence for someone else to reconstruct the decision.
- Compare representative baseline and pilot windows, with case difficulty and exclusions recorded.
- Count eligible reports, actual adoption, accepted outputs, rejected outputs and quality incidents.
- Capture complete human handling time, including exceptions and fallback, alongside separate turnaround measures.
- Reconcile invoices, usage and additional ownership work against the cost model, with a named ongoing owner.
- Show where recovered hours went and what spending changed or incremental contribution was delivered.
- Recalculate monthly net benefit and payback using measured results, including rollout timing and unresolved risks.
Continue if accepted output, ownership costs and evidenced benefits meet the agreed case. If one bounded assumption needs testing, such as a different review procedure or better staff adoption, revise the pilot with a limit on further spending and testing. Stop if the workflow repeatedly fails acceptance or the evidence shows no viable financial benefit. A pilot needs an end date.
An automation earns its place when the business can show a worthwhile benefit after paying for implementation, operation and the human work around it. Hours recovered are part of that case; the decision depends on what those hours let the business do and whether the output remains useful.
Before committing to a larger rollout, put your baseline, full costs and assumptions on one page. If integration or ownership costs are still unknown, use those gaps to scope the automation. Agree what a worthwhile result would look like, then use the pilot to replace assumptions with evidence. That leaves you with a decision you can explain and revisit as the workflow changes.