How to Calculate the ROI of AI Automation

Most AI ROI numbers you’ll see are made up. Vendor calculators assume every saved hour converts to cash, price staff time at the wrong rate, and quietly leave out the cost of maintaining the thing once it’s live.

I build automation for Australian businesses, so I have an obvious interest in the maths working out. That’s exactly why I want you to run the numbers properly before you spend anything. A client who automates something with a four-year payback period doesn’t come back. A client who automates the right workflow, sees payback inside a year, and can show their accountant the numbers becomes a client for a decade.

Here’s the framework we use, with worked Australian examples and honest numbers, including the cases where the answer is “don’t automate this.”

The basic formula

ROI = (annual benefit - annual cost) / total cost

Payback period = upfront cost / monthly net benefit

Simple. The hard part is filling in the inputs honestly. Almost every bad automation decision comes from getting one of three inputs wrong: pricing staff time at the award wage instead of the fully-loaded cost, counting hours “saved” that were never redeployed into anything, or forgetting that automation has running costs and maintenance on top of the build price.

Let’s fix each one.

Step 1: what an hour of staff time actually costs in Australia

This is where most calculations go wrong, in both directions. The base wage understates the cost. But so does dividing salary by 38 hours a week, because nobody works 52 weeks a year.

Take an office administrator on $68,000, which is close to the national average. Indeed’s salary data puts the role at roughly $69,000, and SEEK’s office administrator range runs $70,000 to $75,000.

On top of that salary you’re paying superannuation at 12% (the rate since 1 July 2025), which is $8,160. Workers compensation for an office role runs around 0.5% to 1%, call it $680. If your total wages are over your state’s payroll tax threshold (roughly $1 million to $1.5 million depending on the state) add another 4.85% to 6.85% on top.

For a small business under the payroll tax threshold, that’s about $76,800 a year in mandatory costs before you’ve bought a laptop, a desk or a software seat.

Now the part people miss. A full-time employee is paid for roughly 1,976 hours a year (38 hours across 52 weeks) but works far fewer. Subtract four weeks of annual leave, ten public holidays and ten days of personal leave and you’re at about 1,670 productive hours.

$76,800 divided by 1,670 hours is about $46 per productive hour.

The “$34 an hour” admin actually costs you around $46 for every hour of work you get, and closer to $50 once you allocate equipment, software and office overhead. Employer cost calculators from Australian payroll providers land in the same territory: a true cost of roughly 1.2 to 1.3 times base salary for office roles, higher once recruitment and ramp-up are included.

That $46 to $50 figure is the number to use when you value saved time. Use the award wage and you’ll undervalue automation by 30% or more. Use revenue per employee, as some vendor calculators do, and you’ll overvalue it wildly.

Step 2: measure the time saved honestly

Don’t estimate. Measure. Pick the task, and for two weeks have the person doing it track actual time in a spreadsheet or their calendar. Estimates are almost always wrong, and they’re wrong in both directions: people underestimate the death-by-a-thousand-cuts tasks like email triage and chasing documents, and overestimate the tasks they hate.

Then apply two discounts.

First, the automation won’t handle 100% of cases. A well-built workflow with human-in-the-loop review typically handles 70% to 85% of volume autonomously and routes the rest to a person. If the task takes six hours a week, budget on reclaiming four and a half.

Second, someone still has to supervise it. Reviewing exceptions, spot-checking outputs and handling the edge cases the automation flags takes real time. For document-heavy workflows we usually budget 10% to 15% of the original task time for oversight, at least in the first six months.

So a “six hours a week” task realistically returns about four hours a week of genuinely reclaimed time. Be suspicious of any proposal that doesn’t include these discounts.

Step 3: count every cost, not just the build

An automation quote has four cost layers, and all four belong in your spreadsheet.

The build cost is the one-off design and development. Single-workflow builds commonly land in the low five figures in Australia, and multi-system builds go up from there.

Running costs cover hosting, LLM API usage, third-party subscriptions and monitoring. For a typical single workflow this is usually $100 to $400 a month, depending on volume and how much AI inference is involved.

Maintenance is the layer everyone forgets. APIs change, software gets updated, edge cases surface. Budget 15% to 20% of the build cost annually. Anyone who tells you a custom automation is maintenance-free is lying to you or hasn’t run one for a full year.

And then there’s your own time. The two weeks of measurement, the process documentation, the testing, the exception reviews. Price this at your fully-loaded rate too. It’s not zero.

Worked example 1: invoice data entry

A Perth wholesale business has an admin spending seven hours a week keying supplier invoices into their accounting system, checking them against purchase orders and chasing discrepancies.

The benefit: measured time is seven hours a week. Apply the discounts. The automation handles about 80% of invoices straight through, and exception review takes about 45 minutes a week, so the net reclaimed time is roughly five hours a week. Across 46 working weeks that’s 230 hours a year, worth $10,580 at $46 a productive hour.

The cost: a custom build with OCR extraction, PO matching and accounting integration comes in at $9,000. Running costs (hosting, AI API usage, monitoring) are $180 a month, or $2,160 a year. Maintenance allowance: $1,500 a year.

The maths:

• Annual net benefit after running costs: $10,580 - $3,660 = $6,920
• Payback period: $9,000 / ($6,920 / 12) = about 16 months
• Year two ROI: ($10,580 - $3,660) / $3,660 = 189%

Payback in the 12-to-18-month band, so worth doing if the process is stable. And this example only counted time. It didn’t count the late payment fees avoided, the early-payment discounts captured, or the error rate. Manual data entry runs an error rate somewhere around 1% to 4%; automated extraction with validation rules runs far lower. For a business processing $2 million a year in supplier invoices, a handful of avoided duplicate payments can cover the running costs on its own. I leave these out of the base case deliberately. If the numbers only work when you include the soft benefits, the numbers don’t work.

Worked example 2: missed calls at a trades business

An electrician’s office misses around 20 calls a month, mostly while the team is on the tools. This one is a revenue example, and revenue examples are where vendor maths gets most dishonest, so watch the method.

Not every missed call is a lost job. Say 25% of those calls would have booked: five jobs a month. Then comes the step everyone skips. You don’t bank revenue, you bank margin. If the average job is $450 with a 45% gross margin, each captured job is worth about $200, not $450.

An AI phone agent that answers, qualifies and books into the calendar won’t capture all five either. Say it converts 60% of them. That’s three jobs a month, about $600 a month in gross margin, or $7,200 a year.

The cost: $2,500 for setup and integration with the job management system, plus $250 a month including call volume, or $3,000 a year.

The maths:

• Annual net benefit: $7,200 - $3,000 = $4,200
• Payback period: $2,500 / ($4,200 / 12) = about 7 months
• Year two ROI: $4,200 / $3,000 = 140%

If the vendor pitch had used revenue instead of margin and assumed 100% capture, this same scenario would have shown $27,000 a year in “benefit”. The real number is a quarter of that. Still worth doing, but you should walk in knowing the real number.

Worked example 3: when the maths doesn’t work

A professional services firm asks about automating a weekly status report an admin compiles. One hour a week.

The benefit is one hour a week, times 46 weeks, times $46. That’s $2,116 a year, before discounts.

The cost is the problem. The data lives in three systems with awkward APIs, so the quoted custom build is $6,500, plus about $60 a month running.

The payback period comes out over four years. In four years the firm will have changed at least one of those three systems, and the automation will need rework before it has ever paid for itself.

The answer we gave: don’t build it. Either live with the hour, solve it with a $30-a-month off-the-shelf reporting tool, or park it until there are four or five workflows worth bundling into one build so the fixed costs are shared. Turning that work away is cheap for us. Being the firm that told them the truth is not.

Other cases where you shouldn’t automate

The report example fails on payback period, but there are failure modes the formula alone won’t catch.

The freed time has nowhere to go

Saving five hours a week only creates value if those hours are redeployed into billable work, sales, service or an avoided hire. If the admin’s week just gets 12% more comfortable, you’ve bought capacity, not cash. Capacity is worth something, but be honest that it’s a softer return, and don’t pay a cash price for it.

The process is changing soon

New accounting platform next year? New CRM? Automations are built against specific systems. Automate after the migration, not before.

The process is broken

Automating a bad process gives you bad outcomes faster. If the quoting workflow produces wrong quotes 10% of the time, fix the workflow first. This is usually a process design job before it’s a software job.

Volume is too low

As a rough screen, a task has to consume at least three to four hours a week of someone’s time, or touch revenue directly, before a custom build is worth pricing up. Below that, off-the-shelf tools or simply leaving it alone usually wins.

What a good payback period looks like

For Australian small and mid-sized businesses, the bands we use are straightforward. Under 12 months is strong: proceed if the process is stable. Twelve to 24 months is reasonable, but only for processes that won’t change and systems you’ll keep, and this is where most well-scoped single-workflow builds land. Over 24 months is usually a no. Your systems, your team or the AI tooling itself will change before you break even. The exception is compliance-driven or error-driven automation, where the benefit was never time in the first place.

Notice these bands are shorter than what you’d accept for machinery. That’s deliberate. Software automation sits on top of other software, and other software changes.

Run the numbers before you talk to anyone

You can do this whole exercise in a spreadsheet in an afternoon. Pick the two or three most repetitive workflows in the business, measure them for a fortnight, price the time at your fully-loaded hourly cost, and see which ones clear a 12-to-24-month payback with honest discounts applied.

If one of them clears the bar and you want a real quote against real numbers, book an automation audit with us. We’ll map the workflow, tell you what a build costs, and give you the payback maths in writing. And if the maths doesn’t work, we’ll tell you that too. It’s the cheapest thing we ever do, and it’s why people come back.