On 21 August, the Australian Financial Review's Chanticleer column reported that Commonwealth Bank booked roughly $200 million of gross AI benefit in FY26, guided that figure to double to $400 million in FY27, and delivered it against an operating cost base of $13.6 billion while headcount went up rather than down. The column's verdict was that AI in corporate Australia is a mess. That reading is fair enough as market commentary. The more useful reading for a director is narrower and more uncomfortable: across an entire ASX profit season, one company produced a number a board could actually test, and the number was 2.9 per cent.
The figure itself is not the point. The denominator is. Almost every AI update that reaches an Australian board arrives as a count of something: pilots launched, licences deployed, staff trained, hours saved, use cases in flight. None of those divide by anything, which is precisely why they are comfortable to present and impossible to challenge. The moment benefit is expressed as a share of the cost base it stops being a story and becomes a ratio, and a ratio can be trended, benchmarked, and compared against what the board approved spending to get it. Directors have known this since long before AI arrived. We do not accept a revenue figure without knowing the base it grew from, and there is no reason to grant AI an exemption from arithmetic we would apply to a new product line.
Look at the wording CBA chose, because every word in it is doing work. Gross, not net, which means the spend required to produce the benefit has not been deducted. Benefit, not saving, which signals that a meaningful part of it is revenue or redeployed capacity rather than cash that has left the cost base. And headcount up, which is the quiet demolition of the business case most boards were shown in 2024 and 2025. The Centro principle applies here as squarely as it ever did to a balance sheet: read the numbers yourself, frame them against what you already know about the business, and when something does not sit right, ask. If the most resourced deployer in the country reports a number that modest and hedges it that carefully, a paper landing on your table with cleaner arithmetic and a confident net figure deserves a second read, not a faster approval.
None of this is a technology problem. It is a reporting-basis problem, and boards created it. AI spend has largely entered organisations through technology, change and operating budgets, approved in tranches, each one below the threshold that would have triggered the questions a board asks of a capital programme. The measurement basis was never set at the gate, and attribution retrofitted after the money is spent is the single most reliable way to lose an argument about value. What CBA has done, whatever one thinks of the result, is put a definition on the table ahead of a regulator, an auditor or an analyst insisting on one. That definition is now the benchmark against which your own executive team's next AI paper will be read.
- What is our AI benefit as a percentage of our operating cost base for the year just closed, gross and net of AI spend, and if we cannot state it today, what exactly have we been approving?
- Of the benefit management reports, how much is cash that has genuinely left the cost base, how much is new revenue, and how much is capacity that was redeployed rather than released?
- Which two or three lead indicators tell us this year's benefit is actually going to arrive, and do we see them at every meeting, or do we only ever see the lag number once a year?
- Has headcount moved in the direction our approved business case assumed, and if it has not, is the case wrong or is the execution behind?
- What happens to the benefit line if inference, licence or vendor pricing rises materially, who owns that input, and at what point does it reach this board?
- Who prepares and signs the AI benefit number before it comes to us, and would that person be comfortable if it were disclosed externally with their name attached?
- Benefit reported as hours saved, seats deployed or pilots launched, with no denominator anywhere in the pack. That is a measure of activity being presented as evidence of value.
- The words up to and potential carrying the weight of the claim, describing a range whose lower bound is zero and which nobody is asked to defend.
- AI spend spread across technology, change and business-unit budgets so that no one in the organisation can produce the net position on request.
- The benefit number is prepared by the team whose continued funding depends on it, and no independent party has reviewed the attribution method before it reaches the board.
- Every year's benefits are presented as new benefits, and last year's claims are never re-tested against what actually landed before this year's are approved.
- A business case forecasting net savings and headcount reduction materially cleaner than anything an ASX-20 institution has managed to report. Optimism is not a methodology.
The upside is that the fix is one ratio and a handful of agreed definitions, which makes it the cheapest governance improvement available in this area right now. A board that asks for AI benefit as a percentage of operating cost base, gross and net, changes the conversation inside a single meeting. Management stops presenting activity and starts presenting arithmetic, and the programme itself gets sharper, because a team that knows the denominator chooses different projects from a team measured on adoption. There is an external dividend too. Analysts have already begun sorting the ASX by whether AI shows up as monetisation, productivity or capital expenditure, and the boards that can state their own position with a straight face are the ones that will not be caught out when an investor, a lender or an auditor asks the question first.
The downside is that a number a board demands is a number a board will get, and it will not always be an honest one. Ask for a clean figure and management will produce one, complete with an attribution method nobody outside the programme can audit and a benefit definition generous enough to survive any outcome. The subtler risk is what a single annual ratio does to patience. Some of the work that matters most here is capability building with a payback beyond the current reporting period, and a board governing AI on one percentage will quietly starve exactly that work while rewarding the small automations that report well. The measure is a discipline, not a verdict, and a board that forgets the difference buys precision it can trust in place of value it cannot.
Do three things before your next AI paper is approved. First, ask management for AI benefit as a percentage of operating cost base for the year just closed, stated both gross and net of AI spend, with the definitions written down and tabled rather than explained verbally, because a definition that only exists in a presenter's answer will change the next time the number is inconvenient. Second, require the benefit to be split three ways every time it is reported: cash that has left the cost base, new revenue, and capacity redeployed rather than released, with headcount stated alongside it, since those three things are not interchangeable and only one of them shows up in next year's accounts. Third, make the measurement basis a condition of approval for anything new, name two or three lead indicators to be reviewed at every meeting rather than once a year, and re-test last year's claims before you fund this year's. Then hold your nerve when the honest number comes back small. Australia's most resourced deployer reported 2.9 per cent, gross, with headcount up, and disclosed it plainly. A board that can produce its own version of that figure is governing an investment. A board still being shown pilot counts is watching a project, and there is a real difference between the two.
Researched and drafted by Brad's agentic AI team. Edited and published by Brad Ferris.