Your Subscription Became a Meter
The way your clients pay for AI changed shape, and most of them haven’t noticed — because it still looks like the thing it replaced.
Start with Forrester, which surveyed more than twenty-six hundred business and technology decision-makers about their software and AI budgets. Eighty percent said those budgets are going up. But the finding underneath that number matters more. Forrester points at three vendors — Anthropic, OpenAI, and GitHub — that have already shifted services off flat-rate subscriptions and onto usage-based billing. Then it adds Microsoft to the list for a different reason: a new premium license tier that bolts Copilot and security tools on top of what customers already buy. Three moving to a meter, one moving the price. Same direction. And Forrester’s advice to customers is to build a financial operations practice around token spend, fund controls like model routing, and set guardrails to stop runaway consumption.
Read that plainly. The subscription — a fixed number a business could write into a budget and forget about — is being replaced by a meter.
Except at the scale your clients operate, it doesn’t arrive looking like a meter. Look at how Zoom sells its new AI assistant: twenty dollars per user per month, and that price is described as including AI credits. Microsoft prices Copilot on the same pattern. The seat survived. The per-user line on the invoice survived. There is simply a consumption meter running inside it now, and nothing on that invoice tells a small business where the seat ends and the meter begins.
Now look at how many people that meter is running for. Gallup found fifty-two percent of American workers using AI tools on the job — thirty percent frequently, fifteen percent every single day — and forty-seven percent saying their organization has integrated AI into how it operates, up from forty-one. This is not a pilot in a lab somewhere. It’s half the workforce.
Then the part that doesn’t fit. The UK’s Office for National Statistics — a government statistics agency, not a vendor with something to sell — measured AI adoption across British business and found it had roughly tripled since 2023. Enormous growth. Then it measured the depth of that adoption, and found it had barely moved. Most firms are still running surface-level pilots rather than pushing AI into the processes their business actually runs on.
Hold onto the shape of that. Adoption tripled. Depth didn’t.
So: the price of AI quietly turned into a meter, half the workforce is running it, and the usage is a mile wide and an inch deep. Those three arrived together — and the reason has almost nothing to do with software. To see why, you have to stop looking at the software and start looking at the building it runs in.
Compute Ran Out of Room
Cloud computing was sold on a promise, and the promise was abundance. Capacity is effectively infinite, you pay for what you use, and there is always more where that came from. That promise is breaking, and everything else follows from it.
Computer Weekly examined whether Microsoft has overstretched its own cloud elasticity, and found the answer showing up in operations rather than in press releases — real capacity limits, service rollbacks in some regions, customers discovering that the resource they’d been told was bottomless has a floor. The response from those customers is to stop depending on one provider for it. That’s not a software problem. That’s a shortage.
And the shortage has a second source that has nothing to do with engineering. New York’s governor signed an executive order pausing state permits for new large-scale data centers for up to a year — the first statewide moratorium of its kind in the country. The politics underneath it matter more than the ban itself. Candidates have been winning local races by running against data centers, and Gallup has confidence in big technology companies down to one in five Americans — the lowest since it started measuring, and the only institution this year where the share with almost no confidence actually surged.
Sit with what that does to supply. Compute now requires land, power, and permission. The first two are getting more expensive. The third can be voted away.
Here’s the fair objection: scarce things get more expensive, and vendors raise prices all the time. Why would that change the shape of the bill instead of just the size of it?
Because of what the buildout is doing to the companies funding it. The Next Web reported that Big Tech’s AI capital spending has accelerated to the point where it’s straining cash flow and compressing margins — and that the pressure could push those vendors to cut other investments, raise prices on AI services, or go looking for outside financing.
That’s the pivot. A company selling an abundant good can price it flat, because one more customer costs it almost nothing. A company selling a scarce good it is borrowing money to produce cannot do that. It has to charge by the unit, because the unit is what costs it money.
So the meter isn’t a pricing strategy. It’s what a scarce good does on its way to market. And a bill attached to something physical only stays under control if somebody is watching it.
Which raises the question of who, exactly, is doing that watching inside a twelve-person business.
Nine Percent Ever Finish
So bring this down to the client you actually have — because the hidden meter is only half of their problem.
Research from IDC, conducted on behalf of the software firm SAS — so weigh the framing accordingly — puts a number on where small and midsize businesses genuinely stand with AI. Sixteen hundred SMB leaders across twenty-eight countries. Only nine percent worldwide have fully embedded it into daily operations. Roughly seventy percent are still in experimental or opportunistic use — trying things, not running on them. And nearly half report their data ownership is fragmented enough that scaling AI stalls on it.
Sit with the nine percent. It isn’t that small businesses haven’t started. Most of them have. It’s that almost none of them finished. And the distance between starting and finishing is exactly where the money for the next phase is supposed to come from — because nobody funds a second project when they can’t say what the first one produced.
Now watch a vendor act on precisely that. Microsoft is launching a program called Copilot in 30 — a thirty-day trial, twenty-five users, aimed at businesses under three hundred employees, and delivered through partners rather than sold direct. Look at what Microsoft packed into it. Alongside the setup guides and the adoption content, Microsoft is shipping a success planner: a tool to help the customer work out who should be using it, what for, and what success would even look like. Microsoft built the measurement scaffolding into the trial itself. Read why. A thirty-day trial converts to paid only if somebody can say at the end of it whether the thing worked — and Microsoft is betting that on their own, the customer can’t.
That’s the whole opportunity, handed to you by a vendor, inside a motion you already run.
So here’s the choice. You can become the provider who can state plainly what a client’s AI consumed and what it returned — and then sell the second deployment that proof unlocks, because the budget for phase two is sitting behind a question nobody has answered. Or you can keep competing to sell first deployments into a market where nine clients in ten are already stuck inside one, and never find out whether any of it worked.
That choice sounds like a strategy question. It arrives as a pricing one.
Because the first billable thing here isn’t the measurement — it’s the finding, and you can quote that next week. Your client cannot tell you what they spend on AI, because it isn’t a line item; it’s scattered across seat prices with credits buried inside them. Sell the discovery as a scoped, fixed-price engagement first, then price the ongoing reporting against the number that discovery uncovers. A fee anchored to the spend it governs is the one that survives the spend going up.
What to Consider
- Inventory the hidden meters before you quote anything. Go through the client’s stack and flag every per-seat subscription whose description mentions credits, included AI usage, or AI capacity — the Zooms, the Copilots, the tools that quietly added an assistant last year. For each one, write down what’s included, what happens when they hit the ceiling, and what the overage costs. You can’t put a price on governing a spend that nobody has totaled, and in most shops that total has never been assembled once.
- Price discovery and ongoing measurement as two separate things. The one-time work of finding and totaling AI spend across the stack is fixed-scope and finite, which makes it easy to quote and easy for a client to say yes to. The ongoing work of reporting consumption against what it returned is recurring, and it should be priced that way. Blend them into one number and you’ll make the recurring fee look enormous while giving away the discovery work that proves you’re worth it.
- Anchor the recurring fee to spend under management, and baseline it at the start. A percentage of the AI spend you govern scales as the meter runs, which means you aren’t renegotiating every time consumption grows. But set the baseline when the engagement begins — otherwise the first time you eliminate waste, you cut your own fee, and you’ve built a service that pays you less the better you do it.
If this trend continues: Within a year, “what does your AI actually cost you” becomes a question no small business can answer without outside help — and the provider who put a price on answering it first is the one every renewal conversation after that runs through.