The Bill Nobody Can Check
The way you are charged for artificial intelligence is changing, and the change arrives as a line on an invoice rather than an announcement.
Start with Gartner, which projects that by 2028, consumption-based pricing will account for more than thirty-five percent of net new corporate legal technology spending. You pay for what you use. Gartner’s Shannon Nakamoto puts the cause plainly: as legal AI adoption grows, the tools consume ever more computing power, and that puts vendors under pressure to align pricing more closely with usage.
Now the accounting, and CIO Dive carried two findings worth putting side by side. The first is Accenture’s, from a report it calls the CIO’s Guide to AI Tokenomics, built on a survey of seven hundred and fifty senior executives — and that is a consultancy selling advisory work on this exact problem, so weigh it accordingly. Accenture’s finding is that only one dollar in five of enterprise token spend shows up as a quantified financial outcome.
The second is the scale sitting on top of that. Global AI spending reaches two point seven trillion dollars in 2026, up almost fifty percent year over year, most of it infrastructure. And Gartner’s John-David Lovelock names how a lot of it happens: CIOs are frequently funding AI without intending to, because it arrives embedded inside products and services they already bought.
Third, where the rest of the usage comes from. Deloitte surveyed workers in the United Kingdom and found thirty-one percent using generative AI at work without their employer’s knowledge — and seventeen percent paying for at least one tool out of their own pockets, nine hundred and fifty-eight million pounds between them.
Three counts. The pricing model is moving to metered. The spend does not tie to an outcome, and a large share of it was never a decision anybody made. And the usage nobody in the company bought is measured in the hundreds of millions.
So that is what the counting shows. The harder question is why nobody has a number of their own to check any of it against.
You Pay For Every Retry
A seat license does not care what you do with it. You pay the same whether the person logs in every day or never, which means nobody in the chain ever needed to count what happened after the login. Not the vendor, who had already been paid. Not the buyer, who had nothing to gain from the number. The absence of a meter is not an oversight. It is what the pricing model made unnecessary.
So look at what shows up once you do start counting. BambooHR put a breakdown on it — that is an HR software company researching the workplace it sells into, so weigh it accordingly. Across sixteen hundred American desk workers, about thirty-five percent of AI usage time goes to productive work. Forty-two percent goes to troubleshooting errors and iterating on prompts. That second number works out to roughly twenty full working days a year, per person, spent getting the tool to do the thing.
Hold onto the forty-two. Under a seat license, all of that is free. You rephrase the question six times, you run it again after it gets something wrong, and the bill does not move. Under consumption pricing, every one of those is billable.
The unit of payment stops being the person and becomes the attempt.
Now, you might reasonably say that iteration is not waste — that refining a prompt is how the tool is meant to be used, and somebody on their fourth try is working, not failing. That is fair, and it does not help. The meter cannot tell skillful iteration from flailing. Both render as tokens.
And the attempts land less often than the marketing suggests. The Real-SWE benchmark tested the top coding agents against private, production-grade code and found the best of them failing more than sixty percent of the time, with the failures traced to integration errors and missing requirements rather than anything exotic.
So the bill arrives per attempt, against work where most attempts do not land. To argue with that bill you would need a count of your own. And the UK’s National Cyber Security Centre has named the reason: when an organization tries to prevent employees from using AI, employees turn to tools the organization never approved — introducing risks that, in its words, can be hard to identify.
And notice who is on the other side of that. Under a seat license, a vendor whose tool needed six tries instead of two had a retention problem. Under a meter, that vendor has a better month.
In plain terms: the only complete record of what was consumed sits with the party sending the invoice.
The Meter Isn’t Yours
Everybody in this chain is building a way to see this except you.
Addigy launched an Intelligence Suite inside its Apple device management platform — that is Addigy describing its own product — and one piece of it is shadow AI monitoring: prebuilt compliance rules that let an Apple administrator see, monitor, control and block AI tools across a Mac fleet. Read what that actually installs. A vendor whose software already sits on every endpoint you manage now has a view of which AI tools are running there.
It is arriving from the other direction as well. OpenRouter, which routes requests across a catalog of AI models, now guarantees that anything sent to its United States endpoint is decrypted, processed and served entirely inside the country — or rejected if that cannot be done. It was built for data residency, and prompted by OpenRouter’s own numbers: open-weight models account for about sixty percent of the tokens consumed by US-originating requests, and Chinese models are the majority of those.Whatever routes, counts. A layer that can promise where a request was processed is a layer that knows precisely how many there were.
Above you and beneath you, the count is being built. In the tools you already sell, and in the plumbing your clients reach through. The MSP sits between them as the one party in the arrangement without a number.
So here is the fork, and it is a question about where you sit rather than what you believe.
You can be the reseller of AI consumption. You buy it, you mark it up, it goes on your invoice — which puts the variable inside your fixed-price agreement, with every retry on your side of the line.
Or you can be the governor of it. The client holds the vendor contract and the vendor bill. You get paid for the instrumentation, the policy, and the count. And you become the party who can dispute an invoice, because you are the only one who measured independently.
Somebody will call that a governance decision, and it is one. Governance is also too broad a word to act on. This is the specific version: pick one, per client, in writing, before the first consumption invoice arrives. Or leave it, and the renewal calendar picks for you.
Which is a decision you can start making in a conversation you have not had yet.
Because the conversation that decides this is one you can have with a client before anything changes, and it is a short one. Ask which AI tools their people are already paying for out of their own pockets, and tell them plainly that you want that list before a vendor moves that spending onto a meter. The client who hears it from you first is the client who calls you when the invoice jumps, instead of calling the vendor.
What to Consider
Run the question as an amnesty, not an audit. When you ask a client’s staff what AI tools they are expensing or paying for personally, the answer is only useful if nobody believes they are in trouble for answering. Say that out loud before you ask. An incomplete list gathered under suspicion is worse than no list, because you will build on it as though it were complete.
Say out loud, per client, which side of the invoice you are on. The sentence is short — either “we buy this and resell it to you” or “you buy this and we manage it for you” — and most providers have never actually said either one to a client. In the absence of that sentence, the client has assumed whichever answer is more favorable to them, and you will find out which one at the worst possible moment.
Start keeping a count you can defend, even a crude one. This does not require a platform. A monthly tally of which AI tools are running in each client environment, who pays for them, and roughly what they consume puts you ahead of every provider who has nothing. The value is not precision — it is being a second source on the day a client asks whether a bill is right.
Picture the provider who did this early. A client calls, angry about a bill that doubled. That provider opens their own record, walks the client through what actually got consumed and where, and finds two tools nobody needed. They did not win that call by arguing. They won it by being the only one in the room with a number.
If this trend continues, by the 2027 renewal cycle the first consumption-priced AI invoice lands on an SMB that never signed a consumption agreement — it arrived embedded in a product they already owned — and the provider who cannot produce an independent number for it becomes the one explaining somebody else’s bill.