Updated May 2026. Refreshed with current AUD costs, real measurement methods we use with clients, and the ROI traps that quietly bury automation business cases.

Most automation business cases we see are wrong. Not by a small margin. Wrong by half, sometimes wrong by a multiple. The number that ends up in the slide deck is the number someone wanted to see, not the number that fell out of an honest calculation.

We are Osher Digital, an automation and AI consultancy in Brisbane. We have built and run automation programs for healthcare, recruitment, finance, and professional services clients for years. The pattern is consistent. The projects that fail are not the ones with weak technology. They are the ones where the ROI was inflated up front, the costs were underestimated, and nobody set up the measurement to catch the gap until it was too late.

This guide walks through how we actually calculate the return on investment of business process automation. The cost categories that get missed. The benefit categories that are real but not financial. The formulas that work, and the ones that flatter you into trouble. If you are weighing up an automation initiative or trying to retroactively prove one paid off, this is the framework we use. For broader context on automation programs see our automation roadmap guide and our n8n consulting page.

Why most automation ROI numbers are wrong

Three patterns produce most of the bad ROI numbers we see in client engagements.

The first is treating freed-up time as cash. A team of ten saves two hours a week each. Twenty hours a week. At an internal cost of $80/hour that is $1,600/week, $83,200/year. The slide says $83,200 in benefits. None of that is real unless you actually reduce headcount, redirect those hours into measurable revenue work, or cancel a planned hire. We see “soft” hours-saved benefits inflate the ROI by 3x to 5x without anyone catching it.

The second is omitting maintenance. The build cost is the easy number. The harder one is the cost of keeping that automation alive across vendor API changes, internal system upgrades, edge cases that surface in production, the analyst who has to reverse-engineer it three years later, and the inevitable rebuild when the underlying tooling shifts. A reasonable rule of thumb is that the first year of maintenance runs at 20 to 30 percent of the build cost. Many internal business cases assume zero.

The third is comparing automation to a perfectly executed manual baseline. The benefits side gets the steady-state number. The costs side gets a flawless implementation. Reality is messier. Implementation runs over, the first version misses an edge case, the team takes a quarter to actually trust the new process. If you do not pad the cost side for these realities you are comparing two different worlds.

None of this means automation does not pay off. It often pays off spectacularly. It just means the calculation has to be honest, or the project will get killed in year two when leadership notices the promised numbers never showed up.

The real cost side of automation

Before any benefit calculation, write down what the project actually costs. Not what it costs in a vendor quote. What it costs in your business once you account for the parts that vendors and internal champions both downplay.

Build costs

The headline number is implementation effort. For an n8n or workflow automation pattern that touches three to four systems, our typical client engagement runs $10,000 to $40,000 AUD all in. For a custom AI agent with retrieval and tool calls, $30,000 to $120,000. For an enterprise RPA program with UiPath or Blue Prism replacing a back-office function, well into six figures and often seven once you add license costs. There is no useful single number here. The honest version of this line in a business case is a range, not a point estimate.

Ongoing costs

This is the line that is most often missed. A representative ongoing cost stack for a self-hosted n8n workflow we run for clients looks like this:

Hosting on a small Sydney VPS: $30 to $80/month AUD
LLM API spend (typical mid-volume agent): $200 to $1,500/month USD depending on model and traffic
Third-party data tools (vector store, observability, document OCR): $100 to $500/month USD
Maintenance retainer or in-house engineering time: 4 to 16 hours/month at $150 to $300/hour AUD
An annual rebuild reserve. We pencil in 15 percent of the original build cost as a sinking fund. Things break, vendors change pricing, model identifiers get deprecated, and the cleanest answer is sometimes a partial rewrite.

None of these line items individually breaks the case. All of them together can quietly turn a $50,000 project into a $30,000-a-year operating commitment.

Change and adoption costs

The cost most often left off the spreadsheet. Training time, the productivity dip while people learn the new flow, the documentation that has to be written and kept current, the support tickets when something behaves unexpectedly, and the recurring conversations about whether to roll back. Budget two weeks of partial productivity on a team-wide rollout. Bigger than that for anything touching customer-facing work where staff confidence matters.

The benefit side, and what actually counts

We sort benefits into three buckets. Cash benefits go straight in the ROI calculation at face value. Cash-equivalent benefits go in if there is a credible, time-bound plan to convert them. Strategic benefits go in a separate section and get talked about, not summed.

Cash benefits

Direct cost reductions you can point to on a P&L. Cancelled software licenses. Reduced headcount through attrition. Lower vendor invoices because volumes are processed in-house. Refund or chargeback reductions because errors went down. Penalty avoidance with a documented baseline. These are the benefits that survive a finance team review without a fight.

Cash-equivalent benefits

Time savings that have a real plan attached. If automating invoice processing frees up 25 hours a week and the AP manager has a documented plan to take on supplier rationalisation work that would otherwise need a contractor at $120/hour, that is a credible cash equivalent. If those 25 hours are going to be redistributed across the team with no specific deliverable, do not put a dollar figure on them. Track it as capacity in your benefits register and revisit at six months to see whether the team actually moved on to higher-value work.

Strategic benefits

Faster turnaround for customers, lower error rates, improved staff retention because nobody has to copy-paste between three systems all day, audit trails that make compliance work cheaper. These matter. They are often the actual reason a project gets approved. Do not invent dollar figures for them. Describe them. Set targets. Measure them.

In our experience, business cases that separate these three categories survive scrutiny far better than ones that try to monetise everything. The CFO does not believe the soft numbers. They do believe the hard ones. Putting the soft numbers in their own clearly-labelled section is the difference between a believable case and one that gets sent back.

Formulas that work

The calculations themselves are not complicated. The discipline is in deciding what goes into the inputs.

Simple ROI

ROI % = (Net Benefits / Total Costs) × 100

Net Benefits = Cash Benefits + Cash-Equivalent Benefits − Total Costs
Total Costs  = Build + Ongoing (annualised) + Change

Useful for a single-year sanity check. Do not use it as the primary number for anything that lasts more than 18 months.

Payback period

Payback (months) = Build Cost / Monthly Net Benefit

The number we use most often when talking to clients. Easy to explain to a non-finance audience and surprisingly hard to fudge. Anything under 12 months is excellent. Twelve to 24 months is normal. Over 24 months and the project competes with too many alternatives to be obvious.

Net present value

NPV = −Build + Σ (Annual Net Cash Flow / (1 + r)^year)

Where r is your discount rate. Most Australian businesses we work with use 8 to 12 percent. Use NPV when the project will run more than two years and you want to compare it against other capital allocations on a like-for-like basis. NPV positive at year three is the bar we set with most clients.

A worked example: AI-assisted invoice processing

To make this concrete, here is a sanitised version of a real business case we built with a mid-market client last year. Volumes and dollar figures changed, structure preserved.

The problem. AP team of four people processing about 4,000 supplier invoices a month. Average handling time per invoice 4.5 minutes. About 60 percent of invoices match cleanly to a PO, 40 percent need investigation. Error rate around 2 percent, with errors averaging $400 each in supplier credits, late fees, or chargebacks.

The proposed automation. Inbound supplier emails parsed by an LLM-based extraction agent (Claude Sonnet 4.5 in this design) into structured invoice data. PO matching done in n8n with a confidence score. High-confidence matches auto-coded and routed to the ERP. Anything under threshold goes to the AP team with the extracted data pre-filled.

Cost side

Build: $55,000 AUD (development, integration, testing across 12 weeks)
Hosting: $80/month for the n8n instance and Postgres on a Sydney VPS
LLM API spend: roughly $0.05 USD per invoice in extraction tokens, so $200/month at 4,000 invoices
Document OCR fallback for scanned PDFs: $90/month
Maintenance: 6 hours/month at $250/hour = $1,500/month
Sinking fund for rebuild: $8,250/year (15 percent of build)

Total ongoing: about $32,000 AUD/year all in.

Benefit side

Auto-processing rate after stabilisation: 72 percent of invoices flow through with no human touch
Average handling time on the remaining 28 percent dropped from 4.5 to 1.6 minutes because the data was pre-filled
Hours saved per month: about 240
Cash benefit, year one: one team member moved into a new procurement-analyst role rather than being replaced when she went on maternity leave. Salary differential and avoided contractor cost: $78,000 AUD.
Error reduction: errors dropped from 2 percent to 0.4 percent. Avoided supplier credits and late fees: about $30,000/year

Total cash benefits: $108,000 AUD/year.

The numbers

Year-one net benefit: $108,000 − $32,000 = $76,000
Payback period: $55,000 / ($76,000/12) = 8.7 months
Three-year NPV at 10 percent discount: about $134,000

Notice what this case does not include. We did not put a dollar figure on improved supplier relationships from faster payments. We did not monetise the AP team’s reduction in repetitive work. We did not add a line for “future-proofing.” The case stood up without those numbers, which meant nobody on the finance side had to fight over the soft ones.

AI-specific ROI considerations

AI-driven automation has its own cost shape that traditional RPA business cases do not capture. Three things to watch.

Token cost is volume-linear. RPA scales beautifully because the marginal cost of a bot run is essentially zero once licensed. LLM-driven workflows have a per-call cost that grows with volume. At 100 documents a day this is a rounding error. At 50,000 a day it is a six-figure annual line item. Always model token spend at projected steady-state volume, not pilot volume.

Model deprecation forces rework. We have already had to migrate clients off two model identifiers in the last 18 months. Each migration was 2 to 5 days of regression testing. Build a model-version variable into your sinking fund.

Eval cost is real and ongoing. AI workflows that touch anything important need a small eval suite that gets run before any prompt or model change. Building this is part of the project, maintaining it is part of the ongoing cost. Skipping it is the most reliable way we have seen to convert a working AI workflow into a silent failure six months later.

For a deeper look at where AI agents pay back compared with classical RPA, see our piece on n8n for RPA.

How to measure ROI after launch

The business case is a forecast. Measurement is what tells you whether the forecast was right. The teams that run this well have three things in common.

They captured a baseline before launch. Volume, handling time, error rate, cycle time. Without a pre-launch baseline you cannot prove anything. We insist on at least four weeks of baseline data before go-live.

They review at fixed intervals. Month one, month three, month six, then annually. Each review compares actuals to the business case and updates the forecast. The most common pattern: month one is rough because edge cases are still being smoothed out. Month three is the first honest read. Month six tells you whether the project is on track.

They reset benefits when reality changes. If the team that was meant to absorb the freed-up capacity got reorganised, the cash-equivalent line evaporates. Update the case. Do not pretend the benefit is still there.

If you would like help structuring a baseline and review cadence for an active project, book a call and we can talk through the measurement design.

When ROI is the wrong question

Some automation work does not justify itself on ROI and trying to force it through that lens kills good projects. Three cases.

Compliance-driven automation. If a regulator has told you to log every customer interaction, the ROI is “you keep your licence to operate.” Calculate cost-effectiveness across alternatives, not return.

Customer-experience automation where the value is qualitative. Faster responses, fewer dropped handoffs, better self-service. These show up in retention and NPS over quarters, not in a year-one P&L. Measure them. Report them. Do not hammer them into a fake dollar number.

Foundational platform work. Building a workflow engine, a data integration layer, a shared LLM gateway. These are infrastructure investments that pay off across many future projects. Trying to attribute ROI to the platform itself is a category error. Justify it as a platform, then track ROI at the application layer.

The honest answer for some projects is “this is the right thing to do, but it does not pay back inside two years.” That answer is fine. It is more useful than a cooked ROI number that gets forgotten the moment the project goes live.

Ready to build an honest business case?

If you are sizing up an automation initiative and want a second pair of eyes on the business case before it goes to the executive, that is exactly the work we do most weeks. We will help you separate the cash benefits from the soft ones, pressure-test the cost side, and end up with a number you can defend. Get in touch or look through our AI agent development work for examples of where this analysis has paid off.

Frequently Asked Questions

What is a good ROI for business process automation?

A payback period under 18 months is the bar most boards we work with care about. Year-three NPV positive against an 8 to 12 percent discount rate is the second test. Headline percentage ROI is less useful in isolation because it is so easy to inflate with soft benefits. Look at payback first.

How is AI automation ROI different from traditional RPA ROI?

The big differences are token cost (variable with volume rather than fixed), eval and prompt maintenance overhead (a real ongoing cost that classical RPA does not have), and faster scope expansion (AI agents tend to take on adjacent tasks once they are running, which moves the benefit number up but also the maintenance number). Model these explicitly.

How much does business process automation cost in Australia?

For workflow-style automation on n8n or similar platforms, build cost typically runs $10,000 to $40,000 AUD with $300 to $1,500/month ongoing. AI agent builds run $30,000 to $120,000 AUD with $500 to $5,000/month ongoing depending on volume. Enterprise RPA platforms with seat licences and bot orchestration are six figures up and well above that for sustained programs.

Should I include time savings in my ROI calculation?

Only if you have a credible plan to convert those hours into something measurable: redeploying staff to revenue work, cancelling a planned hire, or absorbing growth without adding headcount. Hours saved with no plan attached is not a real benefit. Track it, do not bank it.

What payback period should I aim for?

Under 12 months is excellent and rare. Twelve to 24 months is the realistic target for most automation work. Over 24 months you are competing for capital with everything else the business could do, and the case has to lean on strategic factors that survive scrutiny.

What is the ROI formula for automation projects?

Simple ROI is (Net Benefits / Total Costs) × 100, where Total Costs include build, ongoing, and change management. For multi-year projects use Net Present Value with your business’s discount rate. Use payback period when communicating with non-finance stakeholders. The formula is the easy part. The discipline is in deciding what counts as a benefit.

How do I track ROI after the project goes live?

Capture a baseline before launch (at least four weeks of volume, handling time, error rate, cycle time). Review at month one, month three, month six, then annually. Compare actuals against the business case. Reset benefits if the assumptions underneath them change. Without measurement, the business case is just a story.

What is the most common ROI mistake in automation projects?

Treating freed-up time as cash without a plan to convert it. The second most common is omitting maintenance and rebuild costs from the ongoing line. Both inflate the calculated return. We see business cases come in 3x to 5x optimistic when both errors are present.

Calculating the ROI of Business Process Automation