How to Budget AI Spend as Pricing Shifts From Seats to Meters

You used to be able to forecast your AI line item the same way you forecast any other software cost: seats times price times twelve months. Add a buffer, done.

That math is breaking. More of your AI spend now depends on what your team actually asks the software to do, not how many people have a login. A quiet month could cost you nothing extra. A busy one could blow past what you budgeted without anyone noticing until the invoice lands.

This isn't a one-vendor problem, and it isn't going away. The question worth answering isn't "what will AI cost us next year." It's "how do we build a budget process that still works when the pricing model itself keeps changing."

What the Latest Data Shows (as of September 2026)

The clearest signal came from Microsoft. On September 25, 2026, the company restructured Copilot into two distinct pricing tracks instead of one flat subscription. Everyday AI work, chat, drafting, summarizing inside Word, Excel, and Teams, stays on a per-seat license. Agentic work, Autopilot, Cowork, Code, and access to frontier models, runs on usage-based billing instead.

Key Facts

• Microsoft split Copilot into a per-seat license for everyday AI and usage-based billing for agentic work like Autopilot, Cowork, Code, and frontier models.
• Those metered services stay off by default. Futurum Group reports admins must set a spending policy in the Microsoft 365 admin center before usage-based billing turns on, and even then, finance teams "can cap spend but cannot easily forecast what a given agentic task will cost before it runs."
• Salesforce made a similar move earlier with Agentforce in 2025, replacing flat per-conversation pricing with Flex Credits billed at $0.10 per action, alongside a separate unmetered per-user license for teams that want to avoid metering entirely.
• Agentic workloads don't just shift the pricing model, they also multiply consumption under it. Our companion piece on why token prices keep falling while enterprise AI bills keep rising covers that mechanism in depth.

The pattern matters more than any single vendor's number. Microsoft and Salesforce arrived at the same structure from different starting points: a predictable per-seat tier for routine use, and a separate metered tier for anything that acts autonomously on a company's behalf. Expect most of your AI vendors to land somewhere similar over the next year, even if the unit changes from credits to tokens to something else entirely.

Microsoft's own admin controls, Agent 365's budgets, limits, and alerts, exist precisely because the company expects this to be a problem. Spending policies are off by default. Nothing metered runs until someone turns it on. That default tells you something: even the vendor shipping the metered product doesn't expect you to budget for it the old way.

Why "One AI Line Item" Stops Working

Most finance teams still track AI spend as a single budget line, the same way they'd track a project-management tool or an email platform. That worked when every AI feature was bundled into one flat per-seat price.

It stops working once part of your AI spend is metered. A single line item can't tell you whether a cost spike came from adding ten new seats or from one team running an agent in a loop overnight. You need the spend split at the source, not reconciled after the fact in a spreadsheet.

The fix isn't a bigger budget. It's a different shape of budget, one with two distinct categories that get managed differently because they behave differently.

A Two-Bucket Framework for AI Budgeting

Split every AI expense into one of two buckets before you try to forecast it. Treat each bucket with its own rules, its own owner, and its own review cadence.

A few rules make this framework hold up in practice rather than just on paper.

Nobody gets metered access by default. Everyday seat AI can roll out broadly because the cost per person is fixed and small. Metered agent AI should require an explicit decision: which team, which workflow, and who owns the number if it runs high. Microsoft built this gate into Copilot itself; you should build the equivalent gate into your own approval process even for vendors that don't. Our AI approval gates and vendor review framework is a reasonable starting structure if you don't already run one.

Set a monthly cap and alert thresholds before anyone starts using a metered feature, not after. Pick a number you're comfortable losing if the workflow turns out to be a dud, then set alerts at 50%, 75%, and 100% of it. Uber learned this the hard way: the company burned through its entire annual AI budget in four months after turning adoption into a leaderboard with no spending ceiling attached. A cap decided in advance costs you nothing. A cap decided after the invoice arrives costs you a very uncomfortable finance meeting.

Measure cost per completed outcome, not cost per token or cost per seat. A credit count tells you how much an agent consumed. It doesn't tell you whether the task it completed was worth that cost. Before you scale a metered workflow past a pilot, know what one completed task costs end to end, and compare that number to what the task cost before AI touched it, whether that's a person's time or a different tool's fee. Our guide to measuring AI adoption ROI walks through building that comparison without over-engineering it.

Review quarterly, and assume the unit of measurement will move. Vendors are still experimenting with how to meter agentic work. One might charge per action, another per token, another per compute-second. A quarterly review isn't just about whether you're over or under budget, it's about whether the vendor changed what a "unit" even means since your last review. If a contract renewal is coming up, pull your actual usage data first. For a structured way to compare vendors on this axis specifically, see our vendor evaluation framework for AI tools.

If you want the fuller cost picture before you set next year's number, the honest cost of AI transformation and the CFO conversation on AI budget both go deeper into what belongs in the number beyond the software line itself.

One side note: this split is easier to manage when the tools underneath your AI layer aren't also creeping toward metered pricing. A core operations platform priced in predictable packaged tiers, the way Rework structures its CRM, lead management, and work management products, leaves the AI layer as the only line item you need to actively watch each month.

What to Do in the Next 30 Days

1. Split your current AI spend into the two buckets. Go through every AI line item on your books right now and tag it seat-based or usage-based. If you can't tell which one a charge is, that's the first problem to fix with the vendor.
2. Name an owner for every metered workflow. Not a team, a person. If a metered agent's cost doubles next month, there should be one name finance calls first.
3. Set a cap and three alert thresholds on every metered feature you've already turned on. If a vendor doesn't support spending alerts, treat that as a red flag in your next contract review, not a minor gap.
4. Pick one metered workflow and calculate its real cost per completed task. Use actual data from the last 30 days, not a vendor's estimate. This becomes your baseline for every future "should we scale this" conversation.
5. Put "pricing model" on your vendor renewal checklist. Ask directly whether a flat-fee product is planning to introduce metered tiers. You want to know before you sign, not after.

The seat-based era of AI pricing made budgeting easy because the math was simple. That era is ending for the parts of AI that actually act on your behalf. The companies that handle this well won't be the ones who find the cheapest metered rate. They'll be the ones who decided, before the first invoice arrived, who gets to turn the meter on, how high it's allowed to run, and what they expect to get back for every dollar it spends.