A tool that produces an answer in thirty seconds can still create twenty minutes of work. Someone has to prepare the input, check the result, correct it and move it to the place it belongs.
That is why “cost per useful outcome” is a more revealing measure than subscription cost alone. It also makes a modest improvement easier to recognise without turning it into an inflated savings claim.
Count the whole attempt
For one repeatable task, record:
- Tool and infrastructure charges attributable to the work.
- Setup and maintenance time.
- Preparation, review and correction time.
- Failed attempts and the manual work needed afterwards.
Keep the result simple enough to update. A small team does not need an elaborate reporting system to learn that checking a generated spreadsheet takes longer than using its existing template.
A worked example in Australian dollars
Imagine a business tries AI-assisted product descriptions. These are illustrative assumptions, not observed results.
It spends A$60 on the month’s tool use and two hours on setup. At an assumed internal cost of A$50 per hour, setup adds A$100. Drafting and reviewing fifty descriptions takes another four hours, or A$200.
Forty descriptions are accepted. The measured cost is A$360 divided by forty: A$9 per accepted description. The ten rejected attempts remain in the cost.
Compare that with the manual process for similar descriptions at the same quality standard. Repeat the calculation once setup is no longer a first-month expense, while keeping maintenance costs.
Separate first-month setup from ongoing operation
In the example, the A$100 setup cost is not automatically repeated every month. Report the first month and the ongoing estimate separately.
If the same A$60 tool allocation and A$200 staff effort produced forty accepted descriptions in a later month with no additional setup, the illustrative cost would be A$260 ÷ 40, or A$6.50 each. Any maintenance or retraining time would need to be added.
That comparison is useful only while the volume, quality standard and effort assumptions still hold. An optimistic second-month estimate is not evidence of an observed saving.
Time saved is not automatically cash saved
If the owner spends less time writing, that may create room for sales calls or an earlier finish. It does not automatically reduce the wage bill.
Keep cash expenditure, staff capacity and quality benefits separate. Also note new work created by the assistant, such as maintaining a reference library or investigating errors.
The National AI Centre’s implementation guidance supports ongoing oversight and evaluation. Our suggested cost measure brings that discipline into an everyday decision: continue because the complete result is useful, adjust the workflow, or stop paying for an experiment that has not earned its place.