Renzo DemartiniIn short
- Test volume, error cost, control risk, payback and the cost of doing nothing, and require every test to pass.
- Show hours apart from dollars, with software and running costs beside the payback.
- A good base case does not outweigh a failed control: defer, and name the missing evidence.
A sample proposal promises a six-month payback. It still fails, because one control is untested.
That is the decision a board should be able to make on one page: the evidence, the downside and the controls, not just the forecast.
Require a decision, not a forecast
Four of five tests pass, and one failed control still stops the approval
| Test | Evidence required | Sample pass |
|---|---|---|
| Volume | Four-week baseline | 1 |
| Error cost | Documented avoidable losses | 1 |
| Control risk | Stop and recovery demonstrated | 0 |
| Payback | Cash case with running costs | 1 |
| Doing nothing | Costed alternative | 1 |
A proposed board test, applied to a sample order-to-cash improvement across sales, operations and finance.
Illustrative figures, not client data.
Four passes do not override one failed control. This sample board needs every test to pass before rollout.
Each test has one pass or fail line, so the board can read the result in seconds. Put the alternatives and the cost of doing nothing beside the proposal, not in an appendix nobody reads.
Establish the baseline
Handling falls from 80 to 30 hours a month and avoidable cost from A$3,200 to A$200
| Measure | Current | Proposed |
|---|---|---|
| Completed orders | 600 | 600 |
| Handling hours | 80 | 30 |
| Paid overflow support | A$2,000 | A$0 |
| Correction losses | A$1,200 | A$200 |
Sample monthly figures for 600 orders, excl. GST. Hours are net of review and maintenance and are not converted to cash.
Illustrative figures, not client data.
Hours released: 50 a month. Cash avoided: A$3,000 a month, from the overflow contract ending and fewer corrections. Neither follows automatically from faster processing, so neither is booked until the contract ends and the loss figures fall.
Keep refund losses apart from staff rework, or the same error is counted twice. Ask management for source records, exceptions and the measurement period.
Test the cash case and the downside
Halve the cash benefit and payback stretches from 6 months to 24
| Measure | Base case | Downside |
|---|---|---|
| Upfront cash | A$12,000 | A$12,000 |
| Gross cash benefit a month | A$3,000 | A$1,500 |
| Software a month | A$300 | A$300 |
| Other running costs a month | A$700 | A$700 |
| Net cash benefit a month | A$2,000 | A$500 |
| Simple payback after go-live | 6 months | 24 months |
Sample, excl. GST. Upfront cash covers implementation and training. Implementation also takes 40 hours of existing staff time, with no extra cash assumed. Simple payback ignores discounting, financing and tax.
Illustrative figures, not client data.
Exhibit 5 · Try itThree years of net cash, discountedStart with the sample above. Change the start-up cost, the cash benefit, the running cost and the rate. See the net present value year by year, and with the benefit halved. Hours freed stay apart.Open the calculator ↓Close ↑
Illustrative figures from your own inputs. Not a quote and not a promise of savings. Cash falls at the end of each year, and the rate is your own choice: 10% is a sample, not a recommended hurdle. No tax, depreciation or financing is included. For year one alone, use the cost-to-own calculator in the automation cost article.
Payback is upfront cash divided by the positive monthly net benefit. Both cases assume steady benefits from go-live. Any ramp-up or delay extends recovery.
The downside halves the gross benefit and leaves costs fixed. Ask which evidence would make that outcome more likely. The costs of owning an automation are set out in what automation costs, what it saves and who owns it.
Payback stops at the break-even month. It does not say what a dollar next year is worth today. For a decision that runs three years or more, discount each year’s net cash at a rate you choose. The calculator below does this on the same sample and keeps the hours freed apart.
More detail
Where this comes from, checked on 3 November 2026. Microsoft’s NPV function takes a rate and a series of values that must be ‘equally spaced in time and occur at the end of each period’. A first cash flow at the start of the first period must be added to the result, not put in the list. The calculator follows that timing: the start-up cost sits at time zero and each year’s net cash at the year’s end. The rate is yours. The 10% in the sample is not a recommended hurdle.
Compare doing nothing with a smaller change
Automating costs A$1,200 a month; doing nothing costs A$3,200 and a manual fix A$2,700
| Option | Upfront | Overflow and corrections | Software and running | Total a month |
|---|---|---|---|---|
| Continue | A$0 | A$3,200 | A$0 | A$3,200 |
| Improve manual checks | A$1,000 | A$2,700 | A$0 | A$2,700 |
| Automate | A$12,000 | A$200 | A$1,000 | A$1,200 |
Sample monthly outflows, excl. GST. The manual option is a once-only A$1,000 for training and checklists.
Illustrative figures, not client data.
Doing nothing keeps today’s cost. It does not prove every dollar is recoverable.
A manual fix may buy time while controls are tested. Compare service quality and team capacity as well as cash.
Record the outcome
A board can record decline, defer or approve against a written trigger
| Outcome | Trigger |
|---|---|
| Decline | No credible benefit, or unacceptable residual risk |
| Defer | Missing evidence that a bounded test can resolve |
| Approve | Evidence, controls and funding conditions met |
The paper carries the sample to a decision: defer, and 30 days to fix the control
| Part of the paper | What it records | In the sample |
|---|---|---|
| Decision sought | The question put to the board, who decides and the review date | Approve rollout? The outcome below is defer |
| Options | Doing nothing, a smaller change and the build | A$3,200, A$2,700 and A$1,200 a month |
| Hours and cash | Kept apart, with software and running costs | 50 hours released; A$2,000 net cash a month |
| Controls | One line for each test, with its evidence | Stop and recovery not yet shown |
| Resolution | Decline, defer or approve, with conditions and a deadline | Defer; decide again in 30 days |
Sample figures from the exhibits above, excl. GST.
Illustrative figures, not client data.
Sample outcome: defer rollout. The stop-and-recovery demonstration is incomplete.
Set a 30-day evidence deadline. Require a tested recovery path and confirmation that overflow spending can end. Keep the present process running meanwhile. The minute should record the evidence the board saw and the date it expects the rest.
ASIC says, ‘A directorship is not just a title.’ This test supports informed oversight. It is not legal advice, so get professional advice on your director obligations.
Put it on one paper: download the AI investment decision paper (Word), filled with this sample and blank.
More detail
The paper holds the decision sought, the options including doing nothing, the evidence behind each claim, the hours and the dollars apart, the failed control and the resolution. Have your adviser or company secretary review the resolution wording. business.gov.au also has a free digital strategy template for goals and actions across the business.
What to do first
Four weeks of baseline and five tests turn a proposal into a board decision
- 01Measure four weeks
Orders, handling hours, overflow spend and correction losses.
- 02Add every running cost
Software, monitoring, maintenance and usage, beside the payback.
- 03Run the five tests
Volume, error cost, control risk, payback and the cost of doing nothing.
- 04Record the outcome
Decline, defer or approve, with a deadline for any missing evidence.
Start with one proposal and a four-week baseline.
Common questions
Are released hours a cash saving?
Only where spending actually changes. Otherwise record the hours as capacity and give that capacity useful work.
Can a good payback justify weak controls?
Not under this test. The National AI Centre’s guidance includes ‘Test and monitor’ and ‘Maintain human control’, so a failed control stops the approval.
How long should a pilot run?
Long enough to compare representative results with the baseline. A daily process gives comparable data sooner than a monthly one.
Chartered Accountant (CA ANZ) and finance executive, based in Canberra. Practical AI and automation for SMEs, and adviser to owners, CEOs and boards on integration and finance.