Unused Seats, Unread Meters: SaaS Cost Basics for AI
SaaS management has a well-worn playbook: find every subscription, give it an owner, measure who uses it, take back the seats nobody uses and negotiate before the renewal date. It works, and it still works on AI tools. It just covers half of the bill.
The other half is a meter. Claude Enterprise is a seat plus usage at API rates. Microsoft 365 Copilot is a seat, and agents beyond the licensed user draw Copilot Credits. GitHub Copilot is a seat plus AI Credits. Four Vendors, Four Meters: AI Pricing Compared sets the published prices side by side. This post is about what a SaaS cost program has to add when the subscription comes with a meter attached.
The numbers behind the problem
Zylo's 2026 SaaS Management Index, published on 29 January 2026 and based on more than 40 million SaaS licenses and $75 billion of spend under management (accessed 23 October 2026), puts both halves in one report:
- Organizations leave an average of 36% of their SaaS licenses unused, measured against recommended utilization levels.
- Spend on AI-native applications rose 108% year over year, and 393% in organizations with more than 10,000 employees.
- Expense-based SaaS spend rose 267%, which Zylo ties to AI tools bought outside procurement. ChatGPT is the most expensed application.
- In a separate survey of 218 IT leaders in the same release, 78% reported unexpected charges tied to consumption-based or AI pricing models in the previous 12 months.
The first figure is the old problem: seats bought and not used. The last is the new one: usage nobody forecast. Shadow AI Is a SaaS Problem First covers how to find the AI subscriptions and keys in the first place. Here we assume you have the list.
Reclaiming seats still works
An AI seat that nobody opens is the same waste as any other unused license, and the same lever applies. Microsoft publishes a per-user Copilot usage report through Microsoft Graph with each licensed user's last activity in each app. Anthropic and OpenAI expose usage through their admin consoles and APIs. Pull last activity, agree a threshold with the business, warn, then reassign.
Two cautions specific to AI:
- Low use is not always idle. A seat used twice a month to draft a board paper may be worth keeping. Agree the rule with the owner before removing anyone.
- A cheaper seat can cost more. Moving a heavy user from a premium tier to a standard one can push their work onto a metered route. Claude's Team plan, for example, lists a Premium seat with five times the usage of a Standard one (Claude pricing, accessed 23 October 2026). Check where the work goes before you downgrade.
Meters need caps, not reclaims
You cannot reclaim a meter. You can only decide in advance how far it is allowed to run. Seat utilization does not predict the invoice: one engineer running an agent overnight can consume more than a hundred occasional chat users. Three controls do the work:
- Budgets and limits where the meter is. Per-project and per-key limits in the API consoles, and per-agent monthly caps in the platforms that offer them. Microsoft's Copilot Studio billing page lets administrators cap credit use for individual agents in the Power Platform admin center (accessed 23 October 2026).
- Alerts before the cap. A cap that stops a business process at month end without warning creates a different problem. Alert the owner at 50% and 80%.
- A named owner per meter. Every key, project, workspace and agent has a person who approves its cost, exactly as every subscription has one.
Pools, commitments and expiry
Many AI contracts now sell usage in advance. The details decide whether a commitment saves money or wastes it:
| Commitment | Published terms | Question to ask |
|---|---|---|
| Copilot Credit capacity packs | 25,000 credits for $200 a pack per month (Microsoft); custom agents are disabled at 125% of prepaid capacity (Learn) | Is consumption close to the packs bought, and is pay-as-you-go set up for the overflow? |
| Claude Team pooled API credits | Up to $500 a month in pooled API credits (Anthropic) | Who draws on the pool, and what happens when it runs out? |
| Claude Enterprise | $20 per seat per month annual, 20-seat minimum, usage billed on top (Anthropic) | Is usage forecast separately from the seat count? |
All terms accessed 23 October 2026. Microsoft's Learn page also says agent flow enforcement resets monthly when prepaid credits renew, so plan on a monthly allowance and ask in writing whether anything unused carries over. A prepaid pool that resets every month is a use-it-or-lose-it commitment, and it should be sized to the low months, not the peaks.
Renewals and auto-renew
AI contracts are young, and many were signed as pilots on monthly or one-year terms with automatic renewal. The renewal is the one moment when seat count, tier, pooled credits and commitments can all be reset together. Treat it that way:
- Put every AI subscription's renewal and notice date in the same calendar as the rest of the estate.
- Start 90 days out for anything material, with 12 months of seat activity and meter consumption in hand.
- Watch promotional prices. Copilot Business is $18 per user per month under a promotion running to 31 December 2026, against a list price of $21 (Microsoft, accessed 23 October 2026). The renewal after a promotion is a price increase unless someone negotiates it.
Not Every Software Renewal Deserves Equal Attention describes how to triage a renewal calendar by risk so the AI contracts get attention in proportion to what they can cost.
Showback and chargeback by team
The FinOps Foundation's FinOps for AI overview (updated 17 February 2026) recommends starting with showback: showing each team its AI cost without billing it, before moving to chargeback. For AI that means two lines per team, the seats assigned to it and the meters its keys, projects and agents consumed, side by side.
Showback changes behavior because the team that turned on an overnight agent sees what it cost. It needs tagging that already exists in cloud cost management: project, team, cost center and purpose on every key and workspace. The benefits of implementing a FinOps strategy covers why that shared language between engineering and finance matters, and Five Ways to Reduce Your Cloud Costs covers the cloud levers that now apply to model APIs too.
A short checklist
- Every AI subscription, key, project and agent has an owner and a cost center.
- Seat activity is pulled monthly and inactive seats are reassigned after notice.
- Every metered route has a budget, an alert at 50% and 80%, and a cap where the platform allows one.
- Every prepaid pool or credit pack is compared monthly with consumption, and its reset and overage rules are written down.
- Every renewal and notice date is in the renewal calendar, with promotional end dates flagged.
- Each team sees its seats and its meters together, every month.
The question this leaves open
The checklist assumes that seats, meters, pools and renewal dates can be read side by side. In most organizations they sit in different consoles, invoices and spreadsheets. Where do they meet, so that the unused seat and the unread meter show up in the same review? The next post in this series turns the meter into a unit cost that finance can use.
Since 3 November, SIM puts the seats and the meters in one place. SSO connectors for Microsoft Entra ID, Google Workspace and Okta show which SaaS apps people have access to, SaaS spend analysis adds what each one costs, and API connections to Anthropic, OpenAI, Google Gemini, Microsoft 365 Copilot and GitHub Copilot add detailed token analysis and usage per user. It is included in the Core plan, with no separate price.
To see your own seats and meters per person, Software Inventory Manager (SIM) now covers SaaS spend and AI cost management in one place. For the token mechanics behind the meter, read Tokens, Explained for the People Who Pay the AI Bill.
