Picture the budget review

Someone from Finance asks what the AI line actually bought this quarter, and the marketing lead who approves the tools every month doesn’t have a number ready. That’s not evasiveness. It’s a symptom. Until fairly recently, most organisations venturing into AI approached adoption experimentally, with a test and learn mindset. The specific use case they were buying AI to solve wasn’t always clearly defined. Without that baseline, there’s nothing to check the number against, so when someone finally asks, silence is the only honest answer. That’s expensive in a way an ordinary budget line never is. Nobody defends a number they can’t explain, and a number nobody defends is the first one cut. 

Westpac and Commonwealth Bank show what closing that gap actually looks like. Westpac CEO Anthony Miller keeps a personal AI token tracker on his own laptop. His chief AI officer’s team matches tool to task deliberately, rather than defaulting to the most capable, and most expensive, model for everything. CBA runs the same discipline at scale, with usage tracking, forward cost forecasting, and caps applied where the numbers call for it.¹ It didn’t take a complicated system. It took someone deciding the number was worth watching before it was demanded of them. 

Agentic AI makes that discipline harder to skip, because it costs more than first appears. A single request now triggers a chain of hidden steps, including planning, checking, retrieving and refining, each burning tokens with nothing on the invoice but the total. Anthropic’s own engineering team cut one agent workflow from 150,000 tokens to 2,000, not by using a smaller model, but by changing how much surrounding information the agent had to read through just to complete one task. A mid-size SaaS deployment saw the same failure mode push cost-per-active-user from $0.40 to $1.10 within eight weeks.² The model is rarely the expensive part. What’s expensive is how much you’re asking it to read before it does anything, and that’s a decision your own team makes every time it designs a workflow, not something a vendor controls. 

The answer isn't simply to spend less

Australian retail brand T2 didn’t cut its AI budget. Instead, it moved the cost centre out of IT and into its own standalone P&L line, so spend could be tied to outcomes that marketing could actually explain and own.³ Small edtech company Turbo AI went the other way. After an unplanned $30,000 token spike in a single month, its founders judged the productivity gain worth the cost and kept building. They made targeted fixes afterward, switching to standard-speed processing and smaller models for routine tasks where it made sense.⁴ The difference isn’t whether to spend more or spend less. It’s whether the number has a story behind it. T2 needed structure because its spend was scaling faster than anyone could account for. Turbo AI didn’t, because its founders had already made a deliberate call that the extra cost was worth it, and could explain that call when asked. The right response depends on which of those two situations you’re actually in, not on picking a side in a spend-more-or-spend-less debate. 

Get confident enough to have that conversation 

Give AI spend its own line, owned by marketing, not IT. As long as token cost sits inside an IT or cloud budget, marketing has no reason to explain it and no real incentive to defend it. Take a leaf out of T2’s book. Move your AI cost centre out of the IT budget and into your own standalone P&L line, and make per-query token limits visible up to senior decision-making level.⁵ If you don’t yet have the authority to move the budget line itself, you can still separate AI spend into its own item in your own reporting and make it visible upward. The ownership question doesn’t have to wait for the org chart to catch up. 

Once the spend has an owner, the next question is what’s actually driving it. Use the right model based on complexity of the task, rather than defaulting to the most expensive model. Build that routing into the system rather than relying on people to choose well each time. Most day-to-day marketing AI use is genuinely low-stakes, including caption variants, first-pass social copy and campaign reporting summaries, and doesn’t need the same model doing strategic brief development or complex creative concepting. Leaving that judgement to individual discretion gets expensive at scale, because people default to whatever tool feels safest, which is usually the priciest one. It’s the same waste as shooting broadcast-quality footage for a placement that only ever runs as a six-second bumper ad. The extra production value doesn’t survive the format, and neither does the extra token spend. AT&T re-architected its own agent traffic on this logic. Routine work goes to smaller, domain-specific models, and frontier models are reserved for tasks that actually need them.⁶ The result was a 90% cost reduction and 3x throughput, without cutting how much AI the business used. Employment Hero takes a simpler approach. Every employee has a licence, but anything requiring more capability is gated and needs approval.⁷ 

Ownership and smart routing still won’t tell you which specific workflows are actually worthwhile. A monthly token bill can’t tell you which parts of a workflow are earning their keep and which are quietly burning budget on rework nobody’s tracking. Track cost at the workflow level, not just the monthly total. Tie that tracking to what the output actually achieves, not just what was spent. Reckitt applied AI selectively across marketing, product development and R&D, rather than everywhere at once, and cut content development time by up to 60% as a result.⁸ The gain came from being deliberate about where in the workflow AI actually shortened the path to a finished asset, not from broad deployment. As Elastic’s Jeremy Pell puts it, AI ROI is a measurement problem before it’s a budget problem.⁹ 

All of this still assumes the metric on the dashboard is measuring the right thing in the first place. Design the metric around outcome, not volume, before the team optimises for the wrong one. A dashboard that tracks how much AI got used will always look like progress, right up until it stops meaning anything. ServiceNow’s global marketing chief, Colin Fleming, re-engineered the company’s demand-generation process specifically to break that pattern.¹⁰ Lead volume dropped 63%, and conversion rose 133%, because the team had been optimising a pipeline number instead of a revenue one. Elastic’s Jeremy Pell frames where this is heading industry-wide, describing a shift from AI usage to strict AI accountability, with success no longer measured by how busy the usage dashboards look.¹¹ 

The size of the number was never really the issue. The ability to explain it is.

---