How to Manage AI Tool Costs Before They Get Out of Control

Uber burned through its entire 2026 AI coding budget in four months. Not the whole company. The engineers, who had been encouraged to use AI tools as much as possible and were ranked on internal leaderboards by usage. By spring, individual engineers were running up $500 to $2,000 a month each in AI token charges. The company capped spending at $1,500 a month per person per tool in June and started having a real conversation about whether any of it was producing measurable productivity gains.

A few weeks later, Microsoft told engineers in one product division to stop using Claude Code by June 30 after hitting nearly the same number: about $2,000 per engineer per month. The division had burned through its annual AI budget in months.

Both are massive companies with finance teams watching every cost center. They still got surprised.

The businesses this matters for right now are the ones with 40, 70, or 120 employees where nobody has built any guardrails yet.

Why AI Costs Are Getting Harder to Predict

Most businesses started with flat subscriptions. ChatGPT Plus is $20 a month. Claude Pro is $20 a month. Microsoft Copilot is $30 a seat. That model is predictable.

The shift is happening fast. More capable AI tools, especially the ones that actually complete tasks instead of just answering questions, charge by usage. Claude Code and tools like Cursor in agentic mode can run $500 to $2,000 a month per heavy user. Microsoft moved GitHub Copilot to usage-based billing on June 1, 2026. More vendors are moving in the same direction.

A five-person team running ChatGPT, Claude Team, Copilot, and Notion AI pays around $445 a month. That's predictable. One person spinning up an agentic coding tool can match that total in a week without anyone noticing.

A Federal Reserve Bank of Atlanta survey put average AI spend at $2,068 per employee for all of 2026, up 50 percent from 2025. That average hides a lot. Employees running heavy agentic workflows consume a disproportionate share. The 49 percent of employees using at least one AI tool their employer has not officially approved are not reflected in those numbers at all.

Amazon ran into a related version of this problem. After building internal leaderboards ranking employees by AI usage, the gaming of those metrics drove compute costs up sharply enough that the leaderboard got scrapped.

The common thread across Uber, Microsoft, and Amazon is not that AI tools are too expensive. It is that per-seat budget logic broke down the moment usage-based pricing entered the mix.

Three Steps to Get Ahead of AI Spending

Build an AI tool inventory. Most businesses do not have one. Not just the tools IT approved. Every tool anyone is paying for, whether or not it went through a formal process. Any account a team member set up with a company card or work email. Anything running in a browser extension or connected to a company system.

Separate the flat-fee subscriptions from the consumption-based tools. These are two different budget categories. A $20-a-month subscription is a predictable line item. A consumption-based agentic tool is a variable that needs its own monthly cap.

Set monthly limits by tool, not annual pools. Uber had an annual AI budget. It was gone in four months. An annual number is easy to check quarterly. It is not easy to stop mid-quarter.

Monthly per-tool caps are the actual control mechanism. A 30-person professional services firm that sets a $300-per-person monthly limit on agentic AI tools will not end up with a $30,000 bill in Q1. That limit needs to get set before the tools are deployed, not after the invoice closes. Which means someone with visibility into how each tool actually bills needs to be in that conversation early.

Write two separate documents. Most AI policies try to collapse everything into one. A governance policy and a usage policy serve different audiences and do different jobs.

A governance policy defines which tools are approved, who can request new ones, what data is never allowed into outside AI systems, and how the company tracks spending. A usage policy covers what employees can and cannot do with approved tools day-to-day. Cyber insurers are asking about both at renewal and adjusting premiums accordingly. EU AI Act high-risk obligations take effect in August 2026.

Why This Is an IT Problem

Finance sees the bill after the fact. By the time a charge shows up on a monthly invoice, the usage has already happened.

Knowing which tools are consumption-based before deployment, tracking spend in real time, catching when a single account triples its monthly token usage before it becomes a quarterly surprise. That is infrastructure work. The same reason you would not find out a server was running at 100 percent by waiting for the electricity bill.

The companies handling AI costs well tend to have one thing in common: IT has visibility before finance does. Grant Thornton found organizations with integrated AI governance are four times more likely to report revenue growth. The causality runs both directions. Good governance makes AI tools usable at scale. And companies that scale AI well tend to have put governance in place first, not after the problem showed up.

If you have already read about the risks of unsanctioned AI tools spreading across your business or the governance challenge that comes with AI agents multiplying across departments, the cost question is the same problem at a different layer of the stack.

Frequently Asked Questions

How much should a growing business budget for AI tools per month?

At current per-seat rates, a 50-person business paying for standard productivity AI tools might spend $1,000 to $2,500 a month. That number changes significantly if anyone is using consumption-based or agentic AI tools, which can run $500 to $2,000 a month per heavy user. Build the inventory first, then set the budget.

What is the difference between an AI governance policy and an AI usage policy?

A governance policy defines which tools the company approves, who controls access, how new tools get added, and how spending is tracked. A usage policy tells employees what they can and cannot do with those tools day-to-day. Both are necessary and they serve different readers. One is an IT and finance document. The other is an employee document.

Can a business avoid AI spending surprises by using only free tiers?

Free tiers of tools like ChatGPT and Claude exist and handle basic tasks. They come with lower context limits, no enterprise data agreements, and limited administrative controls. For anything involving client data or business systems, free tiers are not the right answer from a data governance standpoint.

Does cyber insurance cover incidents involving unsanctioned AI tools?

Insurers are increasingly asking about AI governance at renewal. Using unsanctioned tools that expose sensitive data can create coverage gaps. Policies are evolving quickly. The safest position is a documented inventory of approved tools and a governance framework you can demonstrate to an underwriter.

What is the easiest first step for a business that has never tracked AI spending?

Ask every department head to list every AI tool their team pays for, approved or not, and what it costs per month. That inventory is the foundation. Most businesses are surprised by what turns up.