Your AI Bill Is About to Look Like a Phone Bill
Business AI is now a seat plus a meter, and here is how a 20–50 person company keeps the meter from surprising it.
By John Martines · September 27, 2026 · 7 min
Microsoft, OpenAI and Anthropic now all sell business AI as a seat plus usage. The seat is predictable. The meter isn't. Here's what changed as of late September 2026, and the four things to do before anyone switches on an agent: caps first, read the first bill, track by workflow, keep contracts short.
This isn't a price increase. It's a new shape of bill.
For most of the time small businesses have been buying it, business AI was priced like any other software: a flat monthly fee per person. You counted heads, multiplied, and knew the number. As of late September 2026, the three vendors most small businesses use all sell it a different way — a seat for the everyday work, and a meter for the heavy work.
The seat still behaves like a subscription. The meter behaves like the phone bill your kid ran up on a trip overseas.
Microsoft. On September 25, Microsoft announced a rebuilt Copilot app with three parts: Home (chat plus the Cowork agent, with Word, Excel and PowerPoint built in), Code (build small apps by describing them) and Autopilot (an always-on agent that keeps working when you're not). Microsoft said the per-user license covers everyday use — chat and Copilot inside Word, Excel, PowerPoint, Outlook and Teams. Cowork, Code, Autopilot, other long-running agent features and the frontier models run on usage-based billing, paid in Copilot Credits. Home and Code were starting in Microsoft's Frontier early-access program and Autopilot was headed to private preview at month's end, so not every feature is in every tenant yet. The billing model is.
The controls live in the Microsoft 365 admin center, under Copilot, then Cost management. According to Microsoft's admin documentation, updated September 25, an admin can create spending policies, set a monthly limit for the whole company and a separate monthly limit per person, scope policies to security groups, and set email alerts at a threshold. When someone hits their limit, they lose access to the metered services until credits reset on the first of the month. At the time of writing, the services managed there were Cowork, apps built with Cowork, and the Work IQ API.
Two details matter more than the rest. First, the setup screen offers "Don't limit monthly spending" as a choice, and the per-person limit is optional. Second, every spending policy has a setting called Auto-apply new services, and it is on by default. Leave it on and any new metered Copilot service Microsoft adds later is automatically covered by that policy — and billable under it. Turn it off and you review each new service before it's added.
OpenAI. ChatGPT Business now runs on what OpenAI calls flexible pricing. Each seat has included usage. Past that, eligible work — deep research, the thinking models, image generation, Codex — draws from workspace credits, if the workspace owner allows it. OpenAI's help center says the owner manages credit purchases, automatic reload, alerts and workspace spend limits.
Anthropic. On Claude's Team plan, an owner can turn on usage credits so people keep working after they reach their seat's limit. Anthropic's help center says those credits are billed at standard API rates, that owners can set an organization-wide monthly limit and a per-person monthly limit, and that the feature is off by default. On Claude Enterprise, billing has two parts: a fixed seat fee and separate usage charges at API rates. The Enterprise seats don't come with an individual token allowance at all.
One more date, if you run Power Automate. Microsoft is removing the AI Builder credits that came included with licenses like Power Automate Premium on November 1, 2026. After that, the AI steps in your flows — reading an invoice, running a prompt — draw on purchased Copilot Credits or pay-as-you-go. A flow that ran "free" in October can start costing money in November without anyone touching it.
A pilot looks cheap. Five people drafting emails barely moves the meter, because chat and drafting mostly sit inside the seat.
Agents are different. An agent that works through the inbox every morning runs whether anyone is watching or not, and it can run a lot. The site's September 21 brief carried a Bloomberg report on Harvey, the legal AI company: one agent update drove a twentyfold rise in its token use, and its gross margin swung from about 50% to -50%. That's a software company whose whole business is AI, and the meter still got away from it.
A free account opens the rest of this article and the whole library.