Ask a CFO what kept them awake in 2024, and you’d hear about cloud spend, headcount, or the next funding round. Ask that same CFO today, and the answer has changed: AI token costs.

The examples are no longer hypothetical. Uber rolled Claude Code out to 5,000 engineers. Within four months, the company had burned through its entire annual AI budget, with no clear link yet between that spend and consumer outcomes. Microsoft revoked internal Claude Code licenses after token expenditures blew past annual projections, just six months into the rollout. One enterprise ran up a roughly $500 million AI bill in a single month, because nobody had set a spending limit on an agentic loop.

“The problem was never the price. It’s the volume – and almost no organization built its governance model for the scale AI can now consume at.”

The Falling-Price Trap

Here’s what catches most finance leaders off guard: per-token prices have fallen 98% since 2022, yet enterprise AI bills have risen 320% in the same period. Cheaper tokens didn’t produce a cheaper bill. They produced more ambitious use cases, and those use cases consumed far more tokens than the price drop ever saved.

Economists have a name for this pattern. It’s called the Jevons Paradox, first observed in 19th-century coal markets: as steam engines grew more efficient, coal consumption rose rather than fell, because cheaper, more efficient engines made more use cases viable. AI tokens are following the same curve.

The Agentic Multiplier

Conversational AI, the question-and-answer kind most people picture, typically consumes 200 to 2,000 tokens per interaction. That’s manageable and predictable. Agentic AI is a different animal entirely.

An AI agent doesn’t just answer a question. It plans, searches, drafts, checks its own work, revises, and loops until the task is done. Each step re-processes the growing context from every step before it. A complex agentic task can consume 500,000 to 1,000,000 tokens or more, a 500x jump from a simple chat exchange. Multiply that across hundreds of concurrent agents, and you get the invoice surprises making headlines in 2026.

Goldman Sachs projects that global AI agent token use will multiply 24 times by 2030. If your governance model was built for conversational AI, it wasn’t built for what’s coming.

The AI You Don’t Know Is Running

Before you can govern AI spend, you have to know what’s actually deployed. Most organizations can’t answer that question with confidence. Research across 52 enterprise organizations found that, on average, IT leadership underestimated their active AI agent count by 340%.

This is shadow AI: agents built by individual teams, connected to production data, running with API keys nobody formally reviewed. Unlike shadow IT, which wasted budget on unused software licenses, shadow AI burns budget in real time, every hour it runs, whether anyone is watching or not.

A Governance Framework That Actually Works

The organizations getting this right aren’t spending less on AI. They’re spending it deliberately, using a three-layer control structure:

• Layer One – An organizational budget, the total AI spend envelope authorized by the board for a defined period, modeled on realistic consumption scenarios rather than vendor list prices.
• Layer Two – A team and function sub-budget, allocating tokens by business unit based on historical usage and planned use case expansion.
• Layer Three – An individual and agent runtime quota, where consumption is controlled in real time and AI agents carry hard spending limits, not soft alerts.

Layered with role-based access tiering, this structure has a track record: organizations that implemented it reduced total AI spend by an average of 34% within six months, without any drop in user satisfaction or productivity.

Eight Ways to Cut Waste Without Cutting Capability

None of the following require slowing down your AI program. They require cutting waste, not ambition.

• Prompt compression – Strip prompts of accumulated redundancy, typically a 20 to 40% token reduction with zero quality loss.
• Context pruning – Remove stale conversation history that no longer serves the task, recovering 30 to 60% of consumption in long-running sessions.
• Prompt caching – Reuse stable system prompts and reference documents at 60 to 90% below standard rates. One legal team cut its monthly bill by 71% overnight simply by enabling this.
• Model routing – Send each query to the right-sized model instead of defaulting to the most expensive one, cutting inference costs 40 to 70%.
• RAG over full-context feeding – Retrieve only the relevant document sections instead of loading full documents into context, cutting consumption 70 to 90%.
• Output length control – Ask explicitly for concise output. It’s a zero-cost optimization every prompt should use.
• Batching – Move non-real-time workloads to batch APIs for a 40 to 60% cost reduction.
• Agentic session hard limits – Give every agent a hard token ceiling. Organizations with these limits in place reported zero catastrophic budget overruns, versus a 23% incidence rate among those without.

The Board Is Already Asking

AI spend is now a board question. The only choice leaders have is whether they answer it before the invoice does, or after. Five questions consistently separate the leaders who own that conversation from the ones who get ambushed by it:

• What is our total AI spend this month, broken down by platform, business unit, and use case?
• What is our cost per completed task for our highest-volume AI use cases?
• What are our hard spending limits for AI agents, and what happens when they’re reached?
• Where do we have shadow AI deployments, and what’s the plan to bring them into the governed estate?
• What is our AI ROI baseline, and how are we tracking spend against outcomes?

“Boards don’t reject AI investment. They reject AI investment without accountability.”

Get the Full Playbook

Everything above is drawn from TOKENOMICS: The How-to-Avoid-Token-Burn Playbook, co-authored by Atmoz Co-Founder & CEO Yael Shatzky, Atmoz Co-Founder & CTO Benjamin Bondi, The Executive Board’s Global Managing Partner Subrato Basu, and IHR Insights CEO K R Sreenivasan.

It’s grounded in original research spanning 147+ enterprise technology and finance leaders across six global markets.

Cloud and AI costs get created in real time. By the time they show up in a billing dashboard, the money is already spent and the engineer who made the decision has moved on. That’s the exact problem Atmoz exists to solve for cloud infrastructure, and it’s why token economics is a natural extension of our work. This book isn’t a product pitch. It’s the operating reality our customers live in, written down with the data and frameworks to act on it.

As a thank-you to our community, the book is free on Amazon through August 10th. Get your copy now: https://www.amazon.com/dp/B0GYT51CH5