AI spending has got to the point where it needs limits. The danger is letting those limits decide which work deserves attention while nobody makes that call out loud, writes Lucio Ribeiro.
In 2008, Melbourne was asked to live on 155 litres of water per person per day. Target 155 put conservation on the shower screen and on the bill.
I’d arrived from Brazil 3 years earlier, and the campaign stayed with me. People compared their usage the way they now compare step counts.
Nobody turned off your tap at litre 156. The number gave a large, complicated consumption problem a place inside an ordinary household.
Something similar is arriving in offices: the monthly AI allowance.
Melbourne knew what it wanted to conserve and why. With AI, a smaller bill might mean less waste. It might also mean worthwhile work was never attempted, but the AI invoice can’t tell you which.
The wallet arrives
Atlassian has put its research and development staff on AI wallets, with monthly allowances running from US$500 to US$2,000 depending on role. Staff get a warning as they approach the limit, spending pauses when the balance runs out, and they can ask for more.
That’s a defensible design. A wallet with a working release valve makes cost visible while leaving room to experiment. This is a very similar approach I have taken inside our own offices at TBWA Australia.
Uber is the harder case, and by now an AI dinner table story: the company spent its full-year budget for Claude Code and Cursor in roughly 4 months, then capped engineers at US$1,500 a month per tool in June.
A month before the cap, Uber’s chief operating officer Andrew Macdonald was asked whether all that token consumption was showing up in the product customers actually use. “That link is not there yet,” he said.
The money ran out in month 4. Nothing about the worth of the work changed on the way there, only the date.
When the number becomes the decision
Call it the metering reflex: the point where controlling AI consumption becomes a substitute for judging its contribution.
When you can’t price the output, you price the input, and the number you can see becomes the number you manage. And this is why CFOs now control over 90% of everything contracted in AI.
A budget makes a resource constraint explicit, which is useful. The metering reflex starts when staying inside the budget counts as success on its own, with nobody asking what got delivered or dropped.
Elastic surveyed more than 500 senior AI decision-makers at Australian organisations with at least 50 staff. One in 3 (33%) went over their AI budget in the last financial year.
But wait…just 8% measured AI against commercial outcomes like revenue or productivity. The rest were watching prompt volumes and token consumption.
That is the reflex, counted.
The spread in what firms spend tells you almost nothing on its own. Ramp’s August 2026 AI Index puts monthly AI spend per employee at US$11.95 for the median US business and around US$7,400 among the top 1%. Don’t miss that number if you are building your AI business case.
Some organisations do connect the two. Commonwealth Bank lifted technology spending from $2.3 billion to $2.4 billion this financial year and put roughly $200 million of gross AI benefits against it, with more expected next year. Whatever you make of that number, somebody had to define it and put their name to it.
The larger risk sits with work nobody has commissioned yet: a service nobody has tested, a customer group nobody has looked at properly. Existing work arrives with a client and a deadline. Possibility rarely arrives with either.
If exploration runs on leftover allowance, an organisation funds what it already does and calls the remainder innovation. A cap can become self-justifying: the experiments it prevents never produce the evidence that justified a larger budget.
I have been on the spending side
I’ve spent years encouraging businesses to experiment with AI. That puts an obligation on people like me to say what the experimentation is for.
In 2018, consulting for a large food retailer, we used IBM’s Watson to match creative headlines to customer segments drawn from loyalty data. A$30,000 of investment produced an estimated A$400,000 lift in in-store spending. Those were sales, not profit, but the money had an outcome attached to it. You can see the business case here.
That same year I invested in another AI business, called Social Pulse, after almost $1M in investment, it failed. We’d built too much of it on a dependency on Facebook.
One result doesn’t cancel the other. Together they make me wary of blanket enthusiasm and blanket restraint. Spending more is not a strategy. Neither is spending less. The difficult part is deciding what deserves the next dollar before the result is known.
What the meter doesn’t tell you
I run an AI film festival, DISRUPT. Deakin University and Swinburne University analysed the 371 films submitted to it.
Several entrants made substantial films for under A$1,000, and what a film cost predicted nothing about whether it was any good. What separated the strongest work was intention, taste and critical judgement. As production gets cheaper, the scarce skill becomes knowing what is worth making.
That holds for a token budget as squarely as it holds for a film. Tokens measure the material a model processes and generates. They say nothing about whether the result was worth having. Anthropic’s own cost guidance points the same way: match the model to the task, and stop carrying irrelevant material through a session.
The question is whether a given limit removes waste or prevents useful exploration. Answering it takes judgement about the work, shared between the people funding it and the people doing it.
Professional services should recognise the shape of this. The billable hour already measures human effort by the clock. An AI wallet adds a second meter. If neither connects to a useful result, we have two meters and no scoreboard.
Spending also moves rather than disappears. A cheaper AI workflow that needs more human checking can cost more overall. A blocked tool can send someone elsewhere.
PagerDuty surveyed 1,250 office professionals at large companies across Australia, Japan, the UK and the US. Two-thirds said they had used AI tools they believed weren’t permitted at work.
Nobody can draw a straight line from spending caps to that number. It still shows that a lower bill from approved tools is an incomplete account of what is happening.
Four questions before you set a cap
Leaders need a way to make sensible decisions about different kinds of work, short of a return-on-investment calculation on every prompt.
1. What are we funding? Separate routine production from exploration, and give exploration its own bounded allocation so it doesn’t live on whatever survives the routine jobs.
2. What counts as better? For established tasks, compare cost, turnaround, error rates and human review time against the previous workflow. For experiments, agree in advance what would justify carrying on or stopping. More output is not automatically better output.
3. Who can release more, and how fast? Put the decision close to the work, with clear authority and reasons recorded. A 48-hour approval is a deadly if the deadline is tomorrow.
4. What happened to the work we stopped funding? Check whether it disappeared, shifted to manual hours, moved to an unapproved tool, or genuinely wasn’t worth doing. Review those cases beside the successful ones, or the savings report tells half the story.
All of this works inside a budget. What it needs is someone who stays responsible for the decision after the allowance is set.
Keep the meter. Keep the judgement.
I wrote here last month about deployment debt, the gap between the capability an organisation switches on and the accountable work it can absorb. The fourth condition I set out was Measure. The metering reflex is what happens when Measure quietly collapses into cost.
Melbourne’s water target worked because the number stood for something people understood: a reservoir dropping through a drought.
An AI allowance needs that same connection. Name what the organisation is trying to achieve, and name who decides when a piece of work warrants more.
A cap can tell a team when to stop spending. It can’t tell them when to stop thinking. That decision still needs a name against it.
Lucio Ribeiro is an artificial intelligence keynote speaker, educator and entrepreneur based in Melbourne, Australia. He writes on AI for Forbes Australia, lectures in AI at RMIT University, and holds executive certifications in AI and innovation from MIT Sloan. He has led AI and innovation at Optus, Seven West Media and Nine, and is Chief AI & Innovation Officer at Omnicom TBWA Australia
Want to see more Forbes articles on your feed? Tap here to make Forbes Australia a preferred source on Google.
Look back on the week that was with hand-picked articles from Australia and around the world. Sign up to the Forbes Australia newsletter here or become a member here.
