Tokenmaxxing and the AI Bill Nobody Budgeted For

If you’ve ever opened a cloud bill and thought “how on earth did we use that much?” get ready. AI is about to give you that feeling too.

For the last couple of years, most workplace AI has come bundled into a per-seat licence. Predictable, easy to budget, easy to forget about. That’s changing. From 3rd December, Atlassian moves Rovo to consumption-based billing, and it’s not alone. ClickUp already meters its agents in AI credits, and plenty of other platforms are heading the same way with credits, tokens and usage meters.

Which means the question is no longer “do we have AI?”. It’s “what is it costing us, and what are we getting for it?”.

Let’s look at what happens when organisations get this wrong, why agents make it harder, and what we’ve seen work.

Meet Tokenmaxxing

Tokenmaxxing is what happens when AI usage becomes the scoreboard. Some companies started publishing leaderboards ranking staff by how many tokens they burned, on the logic that more AI use must mean more productivity.

It didn’t. As ClickUp’s write-up on tokenmaxxing lays out, the results were pretty sobering:

• Uber ran through its annual AI budget in four months, with per-engineer spend sitting between US$500 and US$2,000 a month before caps came in.
• Meta reportedly consumed around 60 trillion tokens in 30 days, with estimates of the bill running well into the hundreds of millions.
• DX research found 92.6% of developers using AI, yet time savings plateaued at around four hours a week for over a year.

Amazon SVP Dave Treadwell summed it up nicely: “Please don’t use AI just for the sake of using AI.”

Why Agents Make It Harder

A person asking a chatbot a question uses a predictable amount of compute. An agent doesn’t.

Agents re-read context, retry when something fails, and call other tools along the way. A single loop can chew through tens of thousands of tokens before anyone sees an output. Multiply that by every agent running in every tool you own and the variance gets wild.

The core problem is a budgeting mismatch. Finance teams planned for AI like seat licences. It behaves like cloud compute.

The Real Cost Isn’t on One Invoice

Here’s the bit most organisations haven’t mapped yet. You’re probably not paying for AI once. You’re paying for it in your ITSM platform, your work management tool, your office suite, and whatever chat assistant your team signed up for on a credit card.

Each one has its own meter. Each one sees a slice of your work. None of them talk to each other about spend.

That’s the heart of our panel question. Do you let AI live inside every tool, or do you pick one layer to sit across the lot? There are good arguments both ways, and the billing model is now part of that decision.

Messy Context Costs Money

One thing that doesn’t get talked about enough: poorly organised information is expensive to feed an AI.

Atlassian’s content team restructured 140 of their design standards so their Rovo agents could use them properly. The result was roughly 21% lower token cost, around 60% fewer tool calls and retrieval about twice as fast.

To be fair, it’s not always a saving. In a separate test, restructuring support content improved accuracy by 30% but cost about 9% more per query. Better answers sometimes cost more. That’s fine, as long as you’re measuring the outcome and not just the bill.

What We’ve Seen Work

Every organisation is different and a blog post can’t see your setup. But here are the patterns that keep coming up:

• Measure outcomes, not usage. Shopify’s Farhan Thawar put it well: the real signal is “not who spent the most but those whose tokens generated the most impact”
• Count your meters. List every tool where you’re paying for AI, who uses it and how it’s billed. Most teams are surprised by the length of that list
• Set guardrails before you scale. Usage alerts and soft caps are much easier to put in at the start than after the first nasty invoice
• Clean up the context first. Tidying the knowledge your AI reads is often the cheapest cost saving available

Where it doesn’t hold up: caps that are too tight kill experimentation, and that’s where the genuine wins come from. The aim is visibility, not lockdown.

Come and Talk About It

We’re digging into exactly this at our breakfast panel, Where Should Your AI Actually Live?, on Thursday 29 October at Icehouse in Parnell, Auckland. AI strategist Tim Warren, AI New Zealand founder Justin Flitter and Barfoot & Thompson CIO Simon Casey will talk through what consumption billing means for your budget, and whether AI belongs in every tool or across them.

It’s breakfast and a conversation, not a pitch. Grab your seat here.