Australia's AI budgets are bleeding tokens - here's how to stop it

Ask any Australian or New Zealand CFO how their AI budget performed over the past year, and you're likely to hear the same story: costs that outran forecasts long before anyone could explain why. 

New research from Elastic found that one in three Australian businesses exceeded their AI budget in the last financial year, and that 32% have since paused, cancelled or scaled back AI deployments because the spend couldn't be justified. 

In conversations with customers and boards across the region, I'm hearing a version of this story almost every week.

The numbers globally tell a similar story. Goldman Sachs analysts have predicted a 24-fold increase in token consumption by 2030 as AI agents proliferate - a surge in demand steep enough to worsen an already tight chip supply over the next 12 to 18 months.

Deloitte's 2026 research into AI tokenomics found that half of leaders are now spending between 21% and 50% of their digital transformation budgets on AI, and profiled one healthcare enterprise where token usage grew 8–10% a month, adding more than US$6 million in unplanned annualised cost before finance even had visibility into what was driving it. 

There's also no shortage of stories about engineers gaming internal leaderboards by routing trivial tasks to agents just to inflate usage stats. 

None of that is value. It's noise.

As Manuel Haug, Celonis' Field CTO, puts it: "After years of unconstrained AI experimentation, organisations are entering an era of AI token austerity. Leaders are moving past asking what AI can do, to demanding real outcomes from their compute spend. Without strict governance, generalist frontier models are left to guess, burning through bloated prompts and silent retries."

That guessing is the real cost driver, and it's what I saw play out at our PI Day event in Melbourne last month. Patrick Thompson, our Global SVP of Customer Transformation, put it plainly: companies that have invested heavily in data layers and AI agents are often burning enormous volumes of tokens just trying to build a context model from scratch - one that goes stale the moment the underlying systems of record change. 

"People are realising I have this data layer, I have these agents, but I have no context, no knowledge model, no intelligence around this data. It's just rows and columns," he said. 

His point: that context model doesn't need to be hand-built at all. If it's generated automatically from the systems organisations already run, the token budget earmarked for building it can instead go toward the agentic AI work that actually drives ROI.

This is the shift I'd encourage every business leader wrestling with AI cost blowouts to make: stop treating every task as if it needs a frontier model and start treating operational context as infrastructure. 

When an AI agent understands how work actually flows through your business - not the internet's general idea of how a supply chain or invoice process works - it can send routine steps to smaller, cheaper, faster models and save the expensive reasoning for problems that genuinely warrant it. 

Better still, if that agent can call on a pre-built answer - a manufacturing lead time, an exception rate, a process variant frequency - rather than trawling through raw data and guessing its way to a result, you cut out both the wasted tokens and the retries that come from a wrong guess. 

Every saved step is a saved cost, and a reduced risk of error. It's the difference between a satnav routing you around traffic and simply driving until you stumble on the destination: the same size fuel tank, far less waste.

The Celonis Context Model is built on exactly that premise. It draws on process data pulled straight from ERP, CRM and ITSM systems and desktop-level activity, layers in the business rules, benchmarks and KPIs that define what "good" looks like for a given organisation and applies process and decision intelligence to explain why bottlenecks happen and what to do about them.

Because every agent interaction and every human override feeds back into it, the model keeps compounding: each new agent deployed is cheaper to run and more accurate than the one before it, because it inherits everything the previous ones learned.

For organisations under pressure to justify AI spend to the board, that's the opportunity in front of them. The answer to runaway token costs was never going to be spending less on AI: It's spending it on the right things, grounded in the operational reality of the business rather than a frontier model's best guess. Get that right, and AI stops being an open-ended experiment and becomes exactly what CFOs have been asking for all along: a line item they can actually budget against.