Insights

AI Governance

The seat no longer caps your AI bill

AI vendors are adding a usage meter behind the seat, while many small and mid-size businesses still budget AI by seat count. What changed in AI pricing, why costs spiral, and a basic routine for tracking and capping spend.

October 12, 2026 · 6 min read

The seat no longer caps your AI bill

On June 1, 2026, GitHub moved every Copilot plan to usage-based billing, with each request drawing down credits by the input, output, and cached tokens it consumes. Seat prices did not change. What changed is that the seat no longer sets the ceiling on the bill.

That is the pattern across many of the AI tools a small or mid-size business is likely to buy this year. This article looks at how AI pricing is shifting from the seat to the meter, why costs tend to spiral at smaller companies, and what a basic cost-control routine looks like. Sources are linked throughout.

AI pricing is moving from the seat to the meter

It is tempting to say every vendor is moving to pure token pricing. That is not quite right. The accurate version is seat plus meter.

None of this is a scandal. Usage pricing ties cost to use, and GitHub was candid that a quick chat question and a multi-hour agent session had been costing users the same amount. The difficulty is that most smaller companies budget AI the way they budget software: a seat count times a price. That model stops working once a meter sits behind the seat.

Why the bill spirals at smaller companies

Six things tend to stack up. None of them needs a purchase order.

Seats and tokens are billed in different places. Seats show up in the productivity suite's admin center. Tokens show up on a cloud bill, a model provider's console, or a credit card. Nobody adds them together unless someone is assigned to.

Shadow AI is the starting inventory. In Microsoft's 2024 Work Trend Index, 78% of AI users said they bring their own AI tools to work, and the figure was 80% at small and medium-sized companies. That survey is from 2024. Every one of those tools is a possible account, key, or expense line that nobody inventoried.

Agents multiply calls. One task can trigger many model calls, and each call can re-send a growing history. When Cursor moved its Pro plan to usage at API prices in June 2025, it said the hardest requests cost an order of magnitude more than simple ones, and it offered refunds for unexpected charges from the rollout. In July 2025, Anthropic added weekly limits for subscribers running Claude Code "continuously in the background, 24/7". Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027, and escalating costs are the first reason it names.

Keys leak, and caps move. In February, a three-developer startup in Mexico had a Google Cloud API key stolen. The Register reported that the thief ran Gemini models on it for 48 hours, running up charges many times larger than the company's usual monthly spend. Truffle Security found 2,863 live Google API keys, originally used as project identifiers, that now also authenticate to Gemini. In May, The Register described another developer whose low spending tier was raised automatically as attacker-driven usage met Google's criteria for a higher tier. Google refunded both developers who spoke to The Register on the record after the story ran, and said it is keeping the automatic tier upgrades. A limit set once is not a limit that can be trusted without checking it.

No one owns it. AI spend lands in IT, finance, marketing, and engineering at once. Without a named owner per tool, an alert goes to a shared inbox and waits.

No one reports it. If leadership does not see AI cost once a month next to what it delivered, nobody can fund what works or stop what does not.

A basic cost-control routine

The fixes are mostly administrative. A routine that works for a small or mid-size business has five parts:

  1. Inventory every AI system and API in use, including seats, model provider accounts, cloud AI services, and anything paid for by card.
  2. Assign a named owner to each one. Any AI use without an owner gets one within 30 days, or it loses its key or its seats.
  3. Tag API usage to a team and a use case so spend can be traced.
  4. Set budgets and alerts at 50, 80, and 100 percent, add a forecast alert where the tool supports one, and send each alert to a named person. Confirm at the provider whether a spend limit can raise itself.
  5. Put step and token caps on agents, and review spend once a month in a single report that leadership sees.

A reasonable first step is to pull last month's AI charges from every place they can land: the productivity suite admin center, each cloud bill, each model provider console, and the company card statements. Put them on one sheet with an owner beside each line. If that takes more than an afternoon, the inventory and ownership gap is the first thing to close.

Resources

The free AI Cost Control Kit at kit.carbene.ai packages this routine as a Start Here guide, a 30-day playbook, a general spend tracker, a healthcare (HIPAA) edition tracker with PHI and BAA checks, and a one-page monthly report template. It is free for internal use, and every data row in the trackers is a labeled example with fictional figures rather than vendor prices or benchmarks. Consultants, MSPs, and vendors who want to use it in work they deliver to other organizations need a partner license (contact@carbene.ai).