The big AI cost blowout: You don't need a Ferrari for every AI task

For the past two years, the AI conversation in almost every business has been dominated by access.

Which model should we use? Which vendor should we choose? Which teams should get licences? How quickly can we get staff experimenting?

That made sense during experimentation. At scale, it's harder to justify.

Especially when the bill arrives.

79% of enterprises overspent their AI budgets in 2026, with no single team owning accountability for the bill. Uber's CTO burned through the company's entire annual AI budget in four months.

One company spent half a billion dollars in a single month after failing to set usage limits on its licences.

The reason costs spiral so fast comes down to how agentic AI works. A standard chatbot query triggers one model call. An agentic task, planning, tool calls, verification, self-correction, triggers 10 to 20.

Per-token prices have actually been falling. The volume has more than made up for it.

Not every task needs a Ferrari

Not every business task needs to be sent to the most powerful frontier model available. Some tasks require advanced reasoning.

Some can be handled by a private model trained on the company’s own knowledge. Some should remain with people. Some can be handled by deterministic automation.

That distinction matters because AI is quickly becoming a cost issue, a control issue and a knowledge ownership issue.

A frontier model may be the right choice for complex reasoning, high-value analysis or work that requires broad contextual understanding.

But many everyday workflows don’t need the AI equivalent of a Ferrari to get to the corner store (or in Australian-speak, the ‘milkbar’).

Classification, extraction, document summarisation, routine approvals: these are structured, repetitive and organisation-specific. For tasks like these, lower-cost models deliver five-to-ten times the efficiency without meaningful quality trade-off.

The problem is that most organisations haven't built the architecture to make that distinction. Every task goes to the Ferrari. Not because it needs to. Because nobody decided otherwise.

The model isn't the problem. The context is.

Frontier models like ChatGPT and Claude have an important role. The point isn’t that businesses should avoid them.

It’s that many workflows don't need a general model with a broad understanding of the world. They need a model that understands your world: your customers, your approvals, your exceptions, your tone and your operating history.

That distinction matters because most enterprise AI is operating without that context. It's working from a generic understanding of business, not your business.

The AI bill you don’t see

The rising token bill is the cost you see. There's another one that doesn't show up in any invoice.

Every task your AI completes, every exception it handles, every decision it makes contains signal about how your business works, what your customers need, where your processes break down.

If that signal flows into an external model, it stays there and compounds. That’s how an estimated $1.3 trillion in institutional knowledge flows out of enterprises and into foundation models every year, and most companies have no idea it's happening.

You get the output. They get the asset.

Microsoft CEO drew an explicit parallel to globalisation: where GDP looked fine in the aggregate while entire industries were hollowed out through offshoring. In the process, they lost the knowledge and capability that defined them. The cost showed up later, and unevenly.

He warned that the same dynamic is possible with AI. Companies can run a tighter token budget and still end up hollowed out if the learning from their workflows is building someone else's system rather than their own.

How DecidrOS solves this

DecidrOS maps how work happens, breaks it into executable tasks and makes explicit decisions about which form of intelligence handles each one: a frontier model, a person, a contractor or automation.

Expensive model usage gets reserved for work that justifies it. The learning from each completed task stays inside the business.

For that to work, AI needs to understand how your business actually operates.

Last year, we acquired Sugarwork to capture the tacit knowledge, informal processes and decision logic that rarely appears in documentation. This is the ‘archival dig’ your business needs to be able to implement AI effectively.

This year, we acquired Rumi, which adds the continuous layer: picking up signals from meetings and live working environments so the knowledge base updates as the business changes.

Together, they make routing decisions meaningful. AI that knows your context makes better choices about which intelligence to use — and at what cost.

Then, the system improves with use. Every time the workflow is run, the result goes back “into the brain”, so your business is “constantly flywheeling and improving”.

It's the same compounding effect Satya Nadella argues will define which enterprises pull ahead. Your company isn't just spending on AI usage. You’re building an internal asset.

The AI question your leadership team needs to ask

Most businesses are asking which AI model to use, and how to reduce the bill. Those are reasonable questions. They're also not the most important ones.

The question worth asking is: does our AI strategy protect our company knowledge?

And does it make sure the value we create is owned by us, compounds over time, and builds an asset we control?

That reframes what good AI governance actually looks like. Not just spend visibility or model selection. But a deliberate decision about where your company’s intelligence goes, who benefits from it, who owns it and whether it gets smarter every time it's used.

If you're looking at your AI spend and want to understand how to make it compound rather than just accumulate, see how DecidrOS approaches intelligent AI decision-making.