AI Bill Shock Is Here. Hand the Meter to Your CFO

Uber, a company that runs some of the most sophisticated technology cost management on the planet, exhausted its annual AI budget in four months and responded by capping staff AI spend at US$1,500 a month. That detail comes from Paul Smith's reporting on corporate "AI bill shock" for the Australian Financial Review's technology desk this week, which described a growing revolt among corporates against unpredictable model costs.

The AFR piece cites FinOps Foundation figures showing 73 per cent of enterprises exceeded their AI budget projections. And the ACS's Information Age reported that enterprise AI spend has grown 497 per cent even as per-token prices fell, with flagship model output now costing up to US$180 per million tokens at the top end.

Sit with that middle number for a second. Prices per unit went down. Total spend went up five-fold. If your business is anywhere on the AI adoption curve, understanding why that happens is worth ten minutes, because the same mechanics are heading for your P&L at your scale.

Why cheaper tokens produce bigger bills

Three forces compound.

Success multiplies usage. An AI workflow that works gets used more. The pilot that cost pocket money at 50 queries a day costs real money at 5,000. This is the good problem, and it still wrecks budgets that were set when nobody believed the pilot would stick.

Agents multiply tokens per task. The industry's shift toward agentic AI, systems that plan, use tools, check their own work and retry, means one business task can consume many model calls where a simple chat request consumed one. You are no longer buying answers by the each. You are buying reasoning by the metre.

Premium models carry premium rates. The gap between top-tier and mid-tier pricing is wide, and defaulting everything to the best model, which is exactly what busy teams do when nobody sets routing rules, means paying tender-response rates for meeting-note work.

None of this is a scandal. Vendors publish their prices. It is simply what happens when a consumption-priced utility meets an organisation that has not yet built the muscle to manage consumption. Most businesses learned this lesson once before with cloud computing, and an entire discipline, FinOps, grew out of it. AI is now re-running that curriculum at speed.

Five controls worth installing before the bill teaches you

1. Give the number an owner, and make it a finance owner. AI spend scattered across personal subscriptions, team tools and API keys, with nobody consolidating it, is how four-month budget burns happen. One person, ideally in finance, should see the whole number monthly. If your CFO cannot tell you this month's total AI spend within a day of being asked, start there.

2. Measure cost per task, not cost per month. "We spent $4,000 on AI in June" is noise. "Each AI-drafted proposal costs about $3 against roughly two hours of writing time saved" is a management signal. Unit economics turn the bill from a worry into a decision. Some tasks will look spectacular. Others will not survive the arithmetic, and finding out is the point.

3. Set caps and alerts before you need them. Every major provider offers spend limits, usage alerts and per-key budgets. They take an hour to configure. Uber's US$1,500 cap made headlines because it was retrofitted after the blowout; yours can be boring and pre-emptive.

4. Route work to the cheapest model that passes. Not the cheapest model, the cheapest one that meets the quality bar for that task, which you establish by testing, not by vibe. Reserve the frontier tier for work where the quality difference is worth the rate. A sensible routing policy commonly cuts spend materially without anyone noticing a quality change, because most volume work never needed the premium model.

5. Review the portfolio quarterly like any other spend category. Usage drifts. Teams adopt tools sideways. Models get cheaper and rate cards change under you. A quarterly hour where the owner from control one walks the unit costs from control two keeps the whole thing honest.

Spend going up is not the failure

A closing distinction, because the bill-shock framing invites the wrong conclusion. The problem in these stories was never that companies spent a lot on AI. Rising AI spend attached to measured returns is what winning looks like. The problem is spend that nobody owns, nobody prices per task, and nobody can connect to an outcome, and that problem is entirely self-inflicted.

The cure is not caution. The cure is the same unglamorous financial discipline you already apply to fleet, freight and electricity, applied to a new meter. Businesses that install it now get to scale their AI usage confidently while their competitors oscillate between binge and panic. Hand the meter to your CFO this month, while the numbers are still small enough to be a spreadsheet rather than a headline.

Where does your business stand? The free AI Scorecard takes three minutes and shows you. If you want a straight steer from a person, book an AI Opportunity Call.