For a couple of years, buying AI for a team looked like buying any other SaaS tool: pick a plan, multiply by headcount, renew once a year. That model is quietly disappearing. The plan pages of the major AI tools now talk about premium requests, usage pools, session limits and credits. The monthly seat price still exists, but it increasingly buys an allowance rather than unlimited use.
For an operations or IT leader who signs off on team licences, this changes the question. It is no longer only "how many seats do we need?" but "how much AI work will these people do, which models will they use for it, and what happens when they run out?"
This guide explains what is changing and gives a simple way to budget and govern metered AI without turning every prompt into a finance ticket.
What is actually changing
Vendors describe it in different words, but the pattern is the same: a fixed fee buys a base allowance, heavier or more capable usage draws it down faster, and anything beyond it is either blocked, slowed or billed separately.
A few examples from current vendor pages (check them yourself before you budget; terms change often):
- GitHub Copilot plans list premium requests alongside the seat, with base credits, pooled usage and additional allowances depending on the plan. Requests to more capable models count against that allowance.
- Cursor states in its pricing FAQ that every tier includes a pool of AI usage, and positions its business plans around pooled usage and invoicing.
- Claude (Anthropic) documents that its Pro and Max plans have a five-hour session limit and a weekly limit, shared between the chat interface and Claude Code, with options for what to do when a limit is reached.
None of this is a scandal. Running large models costs real compute, and agentic tools that read whole codebases or run long tasks consume far more than a quick chat question. Flat pricing was always going to be under pressure once usage became uneven. The practical consequence is that AI spend now behaves more like cloud spend than like a software licence: it varies with behaviour, and someone has to watch it.
Why the old budgeting approach breaks
Headcount-based budgeting assumes that every seat costs the same and that usage does not matter. With metered plans, three things break that assumption:
- Usage is highly uneven. In most teams, a small group of power users does most of the heavy work: long coding sessions, document analysis, agents running multi-step tasks. Everyone else asks occasional questions. The same seat price hides very different consumption.
- Model choice drives cost. The most capable models typically draw down allowances faster. A team that defaults to the top model for every small task spends its allowance on work a lighter model would handle.
- Limits create hidden downtime. When a developer or analyst hits a usage window mid-task, the cost is not on the invoice. It shows up as waiting, switching tools, or a workaround on a personal account, which is a governance problem in its own right.
A practical budgeting model
You do not need a full FinOps programme to get this under control. A simple four-step model is enough for most teams.
1. Split users into tiers
Look at who actually uses AI and how. Three tiers usually cover it:
| Tier | Typical use | Plan approach |
|---|---|---|
| Occasional | Questions, drafting, summaries | Base plan, no add-ons |
| Regular | Daily drafting, analysis, light coding help | Base plan, monitor monthly |
| Heavy | Agentic coding, long document work, automation | Higher tier or pooled allowance, with a named budget owner |
Assign people to tiers based on observed usage after a month or two, not on job title. Reviewing the tiers each quarter keeps you from overpaying for idle seats or starving the people who create the most value.
2. Budget per workflow, not per seat
Instead of "40 seats at plan X", write the budget as the work it supports: code review assistance for the platform team, contract summarisation for legal operations, support reply drafting. For each workflow, estimate the volume (tasks per week), which tier of model it really needs, and who owns it. This makes it obvious where spend is justified and where it is habit.
If you have not yet mapped where AI actually saves time in your organisation, start there; our guide to running an AI audit walks through it.
3. Set alerts, caps and an overage owner
Most business plans offer some combination of admin dashboards, spending limits and pooled allowances. Use them deliberately:
- Soft alert at around 70–80% of the monthly allowance, sent to the workflow owner, not just to IT.
- Hard cap for the occasional and regular tiers, so a forgotten agent loop cannot burn the month's budget.
- Named overage owner for the heavy tier: one person who decides whether to buy extra capacity, switch the task to a lighter model, or wait for the reset.
The percentages are a starting point, not a rule. What matters is that every overage is a decision someone makes, not a surprise on the invoice.
4. Measure cost against outcome
A monthly review should answer two questions per workflow: what did it cost, and what did it replace? Hours of manual work saved, faster turnaround, fewer errors. If a workflow consumes a large share of the allowance but nobody can point to an outcome, it is a candidate for a lighter model or a redesign. If it clearly pays back, it deserves a protected budget. Comparing AI spend to the true cost of the manual process it replaces is usually more persuasive to finance than any per-token figure.
Governance: keep usage inside the company
Metered plans create a specific risk: when people hit limits, they look for a way around them. Personal accounts, free tiers and browser extensions all appear quickly when the official tool stops working at 3 p.m. on a deadline day. That moves company data into tools you do not control.
Three simple rules help:
- Make the escalation path faster than the workaround. If someone hits a limit, they should know exactly who can raise it today.
- State clearly which tools are approved for which data, and that personal AI accounts are not a substitute.
- Review limit events, not just spend. Frequent limit hits in one team are a signal to adjust tiers before they become a shadow-IT problem.
For organisations running AI agents at scale, these rules belong in the same policy framework as other AI controls; we covered that approach in agentic AI governance through policy as code.
What to do this quarter
If you own AI licences for a team, a short checklist:
- Open each vendor's current plan page and write down what the seat actually includes: allowance, limits, reset period, overage rules.
- Export usage data from the admin console and sort users into the three tiers.
- Rewrite the budget line as workflows with owners.
- Turn on alerts and caps where the plan supports them, and name an overage owner.
- Put a quarterly review in the calendar, because the terms will change again.
The end of "unlimited" AI is not a reason to cut back. It is a reason to manage AI the way you already manage cloud infrastructure: visible usage, clear owners and spend tied to outcomes. Teams that do this early will keep their best users productive while everyone else is arguing about the invoice.
If you want help mapping AI usage to workflows and setting up a budget your finance team will accept, get in touch.
Sources
- GitHub, Copilot plans and premium requests
- Cursor, Pricing and usage FAQ
- Anthropic, Using Claude Code with your Pro or Max plan
- FinOps Foundation, FinOps for AI overview