All articles
Founder Guides

How to Stop Your Team From Burning the AI Budget (Without Banning It)

The durable fix is not rationing tokens after the overspend, it is modeling cost per task up front so every team gets a budget tied to real unit economics.

Jun 24, 2026 · 4 min read
How to Stop Your Team From Burning the AI Budget (Without Banning It)

Key takeaways

  • The "tokenmaxxing" phase is ending. Companies that once encouraged liberal AI use are now rationing it after employees burned budgets on small, low-value tasks.
  • The root cause is missing unit math: most teams shipped AI access without ever pricing a single task.
  • A workable AI budget builds up from cost per task, to cost per seat, to cost per team.
  • Caps land better when they map to value, not to raw usage.
  • You can model all of this before you set a single limit.

Why are companies suddenly rationing AI?

Reporting this week describes a sharp turn from "use AI for everything" to deliberate token rationing, with consultancies including Accenture noting that employees were maxing out budgets on trivial tasks. The headline framing is behavioral: people overusing a shiny tool. The operator reality is structural: most companies handed out AI seats with no idea what a task costs, so there was never a benchmark to ration against.

If you do not know your cost per task, every cap you set is a guess.

What actually drives an AI bill?

Three numbers, multiplied: tokens per task (input plus output), price per token for the model you picked, and the number of times the task runs. That third number is where budgets quietly explode. A single trivial call is rounding-error cheap. The same habit, repeated thousands of times a day across a whole org, becomes a line item.

Say a "summarize this email" task uses about 2,500 tokens. At an illustrative blended rate of $5 per 1M tokens, that one call costs roughly $0.0125. Now run it 20 times a day across 500 employees: 10,000 calls a day, about $125 a day, or roughly $2,750 a month, for one low-value habit. Trim the prompt or swap the model and that number moves a lot. (Figures are illustrative, plug in your own.)

How do you set an AI budget that holds?

Build it bottom-up.

1. Price the task

Pick your 5 to 10 most common AI tasks and estimate tokens per run. Multiply by current model prices to get a cost per task. This is the unit everything else rests on.

2. Roll up to a seat

Estimate how often a typical user runs each task per day and sum it. Now you have a defensible cost per seat per month instead of a vague "AI is expensive."

3. Cap on value, not volume

A $0.40 task that drafts a customer proposal is cheap. A $0.01 task someone fires 200 times out of habit is not. Tie limits to outcomes: give revenue-facing workflows room, throttle the idle loops.

4. Route by task, not by default

The most common waste is running a frontier-tier model on work a budget-tier model would finish fine. Route trivial tasks to cheaper models and reserve premium models for jobs that need them. This single change often beats any usage cap.

Why do blanket token caps backfire?

A flat per-seat token cap treats a $50,000 deal proposal the same as a meme summary. People hit the ceiling on the work that matters and learn to avoid the tool exactly when it pays off. Value-based budgets avoid that trap, but they only exist if you have done the unit math first.

Takeaway: rationing is a symptom, missing unit economics is the disease, so price the task before you cap the team. You can model cost per task, per seat, and per provider in Calcaas before you roll out a single limit.

Frequently asked questions

How do I calculate AI cost per task?

Estimate the input plus output tokens a task consumes, then multiply by your model's per-token price. Run the same task across realistic daily volume to see the monthly figure. That cost per task is the foundation for any seat or team budget.

Should I ban AI tools to control spend?

No. Bans push usage into untracked tools and kill the productivity you paid for. A better lever is routing trivial tasks to cheaper models and tying budgets to the value of the work, not the raw token count.

What is tokenmaxxing?

It is the informal term for employees using AI liberally for everything, including tiny tasks, until token spend balloons. The reaction now underway is token rationing, which works far better when it is grounded in cost-per-task math.

Which is the biggest hidden driver of AI cost?

Repetition. A cheap task run thousands of times a day usually costs more than a handful of expensive ones, so frequency and model choice matter more than any single prompt. Place this JSON-LD inside a `<script type="application/ld+json">` tag in the page head. The schema mirrors the visible FAQ above. Source / topic signal (no hotlinking): TechCrunch, "Companies are scrambling to stop employees from maxing out AI budgets with small tasks."

More from the blog

The Margin Memo

Pricing math, in your inbox.

One short note a week on AI pricing, token economics, and margin. No spam, unsubscribe anytime.