How to measure the ROI of an AI project

Why most AI ROI claims cannot be checked

Most AI ROI figures cannot be checked because they have no baseline: nobody recorded how long the work took, or how often it went wrong, before the AI arrived. Without that number, “saves 10 hours a week” is an estimate from the people who bought the tool.

The macro picture is genuinely positive, which is part of the problem. The Productivity Commission’s final report on data and digital technology, released in December 2025, considers labour productivity growth of about 4.3% over the next decade likely from AI, roughly $116 billion of extra GDP, while stressing considerable uncertainty about the size. The National AI Centre’s AI Adoption Tracker reported in June 2025 that 41% of small and medium enterprises were adopting AI. Neither figure tells you whether your project will pay. That needs a measurement you run yourself.

Regulators see the same gap from the governance side. ASIC’s REP 798 reviewed 624 AI use cases across 23 financial services licensees and found that 57% were less than two years old or still in development, and that in some cases governance lagged adoption. Projects that move fast without measurement are hard to defend later, to a board or to a regulator.

Step 1: pick one workflow and measure it before you change it

Choose a single, repeated task with a clear start and end, and measure it for two to four weeks before the pilot begins. Good candidates have high volume, consistent inputs and an existing quality check: invoice coding, enquiry triage, first drafts of quotes, document summaries, claims intake.

Record five numbers for the baseline:

1. Volume: items per week or month.
2. Handling time: minutes of staff time per item, from a sample or system timestamps, not from memory.
3. Cycle time: elapsed time from arrival to completion, which is what customers feel.
4. Error rate: items that needed rework, were wrong when checked, or caused a complaint.
5. Cost per hour: a loaded hourly cost for the people doing the work, covering wages, super, payroll tax and overheads. Write the assumption down.

If you can’t measure the baseline, you have learned something useful: the workflow is not yet defined well enough for AI, and the first job is to define it.

Step 2: decide which benefits count

Count only benefits you can measure against the baseline, and keep them in separate columns. Mixing them hides which one is doing the work.

Benefit type	How to measure it	When it becomes dollars
Staff hours	Handling minutes per item, before and after	When hours are redeployed or a cost (overtime, temp staff, a hire) is avoided
Errors and rework	Errors per 100 items; rework minutes	When rework hours or error costs (credits, penalties) fall
Speed	Cycle time, response time	When faster turnaround wins work or reduces churn you can trace
Revenue	Conversion rate on the affected step	Only with a comparison group; otherwise report as unpriced
Risk	Exceptions, near misses, audit findings	Rarely priced; report as a measured change

Speed and revenue benefits are real but the easiest to overclaim. If quotes go out in one hour instead of one day, track the win rate for the AI period against the same months last year, and say plainly if the sample is too small to be sure.

Step 3: count all the costs

A full cost line includes the build, every monthly running cost, and the staff time spent checking the AI’s work. Review time is the cost most business cases leave out, and it is often the largest.

• Build or setup: development, integration with your systems, testing. See AI development cost in Australia for typical ranges.
• Model usage: tokens or per call fees, which scale with volume and prompt size. LLM running costs explained shows how to estimate them.
• Licences and hosting: seats, vector databases, logging, cloud hosting.
• Review time: minutes per item a person spends checking or correcting output. Measure it in the pilot; don’t assume it.
• Change management: training, updated procedures, the productivity dip in the first weeks.
• Maintenance: prompt and model updates, evaluation reruns, fixes when an upstream system changes. Budget it monthly.

Step 4: separate time saved from money saved

Time saved becomes money only when the freed hours are used for paid work or a cost actually disappears. If a team saves 100 hours a month and spends them on slightly longer lunches, the business saved nothing.

So add one more assumption to the model: the redeployment rate, the share of saved hours that turns into value. Base it on a named plan. For example, the accounts team will absorb 20% more volume without the planned hire, or the two hours a day saved on triage go to outbound follow up with a target attached. Without a plan, use a low rate, and report the full time saving separately as capacity.

Step 5: run the pilot as an experiment

Treat the pilot as a test with a hypothesis, a measurement period and a decision rule set in advance. The Guidance for AI Adoption makes the same point from the governance side: define acceptance criteria before deployment, test against them, and monitor performance afterwards.

• Write the hypothesis: “AI drafting cuts handling time per enquiry from 9 minutes to 4, with no rise in error rate.”
• Use the same kind of work as the baseline, ideally with a comparison group still doing it the old way.
• Exclude the first one or two weeks from the steady state numbers, while still counting their cost.
• Check quality on a sample with someone who didn’t do the work, using an evaluation set where the output can be scored.
• Decide beforehand what result means scale, adjust or stop.

A worked business case

The example below is hypothetical, with round numbers chosen to show the arithmetic. It is not a client result or a quote. All figures are AUD, ex GST.

A business processes 1,500 supplier invoices a month. The baseline shows 6 minutes of handling per invoice, or 150 hours a month, at a loaded cost of $55 an hour. In the pilot, AI extracts and codes each invoice, and staff review it in 2 minutes. That is 50 hours, a saving of 100 hours a month, worth $5,500 if all of it is redeployed.

Line	Year 1	Year 2
Hours saved (months 1 and 2 at half rate)	1,100	1,200
Benefit at $55/hour, 100% redeployed	$60,500	$66,000
Build and integration	$40,000	$0
Model usage, hosting and maintenance ($1,200/month)	$14,400	$14,400
Training and procedure updates	$3,000	$0
Total cost	$57,400	$14,400
Net benefit, 100% redeployed	$3,100	$51,600
Net benefit, 60% redeployed	($21,100)	$25,200

At full redeployment, the project roughly pays back in the first year and returns about 76% on its two year cost of $71,800. At 60% redeployment, year 1 is negative and the two year net is about $4,100: still positive, but thin. The redeployment assumption, not the AI, decides whether this is a good investment. That is why it belongs in front of the board, not in a footnote.

Two further checks make the case honest. Run a sensitivity line on review time: if checking takes 3 minutes instead of 2, the saving falls by a quarter. And if the error rate also fell, show the rework hours saved as a separate line rather than folding them into the headline.

The R&D Tax Incentive belongs in a separate line

Show any R&D Tax Incentive benefit as a conditional line below the main case, never inside it. The incentive offsets some costs of eligible research and development, but routine integration of an existing AI service is unlikely to qualify, and eligibility depends on the activities and the records kept. The government also announced changes in the 2026 to 27 Budget that take effect from 1 July 2028, including removing supporting R&D activities and raising the minimum expenditure from $20,000 to $50,000. Our guide to the R&D Tax Incentive for AI projects explains the test; a registered R&D tax agent should confirm any claim.

Reporting results to a board

Give the board one page: the baseline, the measured change, the full cost, the redeployment assumption and a range. A single headline multiple invites the question you can’t answer.

1. The workflow, its baseline numbers and the measurement period.
2. Measured results: handling time, review time, error rate, cycle time.
3. Total cost to date and the monthly run rate.
4. Net benefit at a conservative and an expected redeployment rate.
5. Quality and risk observations, including incidents and complaints.
6. The decision requested: scale, adjust or stop, with the next measurement date.

How AWLabs approaches this

When a business asks us to build an AI feature, the Discovery Audit for AI work, our AI readiness assessment, starts by choosing the workflow and agreeing how the baseline will be measured. That baseline then shapes the Schedule of Work, which sets a fixed price with three options (The Minimum, The Recommended and The Complete), so the cost side of the business case is known before the build starts. We build logging into the feature so handling and review time can be measured in production rather than estimated. If the numbers don’t support a build, we say so. For the overall cost of a project, see AI development cost in Australia, and for a quick start without a custom build, AI for small business in Australia.