The most capable client I spoke to last week has bought the AI, has the licences, has the appetite — and has not switched it on fully for his team. His reason had nothing to do with risk, security or value. He could not work out where to watch the bill.

He said it plainly: he holds a personal Claude subscription, and on that he can open his organisation's page and see his consumption. In the Microsoft estate, he said, it is "kind of spread out." So the rollout waits. Not because anyone objected. Because a competent operator will not hand a metered resource to thirty people while the run rate is invisible.

In the same conversation one of my team walked the setup path for Cowork on another client's tenant and hit a wall that ought to be famous by now: the configuration screen tells the customer that if their billing is managed by an MSP, they need to contact that MSP. The client cannot turn the AI on without the provider — not because of what the AI costs, but because the provider owns the billing relationship. The consumption may be passed straight through at cost. The customer still cannot see it, cap it, or start it without asking.

The collision

For two decades the managed services proposition has been predictability. The whole pitch is that variable, unpredictable IT becomes one budget line you can plan around. Every commercial habit in this industry — per-seat pricing, all-inclusive support, annual uplift — exists to sell that promise.

AI is metered by construction. Copilot Studio consumes credits depending on how an agent is configured. Cowork draws on a credit balance. Agents run on consumption in the same way Azure does. These two facts cannot both hold in the same agreement, and the industry has so far resolved the contradiction by not mentioning it.

What surprised me is which side breaks first. The commentary that exists on this — and there is some, mostly from the United States — treats it as a margin problem for providers: metered AI meets fixed-fee support, and "we'll reconcile later" becomes "we'll eat it later." That framing does not quite survive the trip here. Providers pass consumption through; almost nobody is wrapping tokens into a fixed price. What lands in the fixed fee is not the tokens — it is the work around them: setting allowance profiles, fielding the requests for more, engineering agents to run cheaply, explaining a bill that moved. That effort scales with consumption. The fee does not. But it is still the second-order effect.

The first-order effect is that adoption stalls. Careful organisations refuse to deploy a metered capability they cannot forecast, and careless ones deploy it and get a surprise. Neither of those is a licensing problem. Both are governance problems that nobody wrote into the contract.

What it looks like on the ground

At a hundred-person engineering software firm, the IT lead described roughly twenty AI champions who periodically come to him saying they have run out of tokens and can they buy more. He has no policy and no strategy for answering, so each request is decided on the spot. That is not a technology gap. That is an organisation discovering it has created an uncapped internal utility with no meter reader.

We hit this ourselves early, and the answer was dull and effective. We set credit usage profiles, gave everyone a default allowance, and made anyone who exhausts it come and ask. The team also changed how they work: develop a skill in the chat interface, then take only the final pass into the metered environment, because that is where the credits burn. We engineer for cost inside the builds too — our document-creation agent uses AI to generate the content but a logic app to write the file, because letting the model produce the document is materially more expensive. It runs for about $150 a month and roughly a cent a document.

None of that is clever. It is just the ordinary discipline of managing a consumption resource, applied a year earlier than most organisations will get to it.

The conversation to have before renewal

If you run a Small to medium-sized organisation, three questions will tell you where you stand, and none of them require a technical answer.

1. Where do I see my AI consumption, in one place, without asking anyone? If the answer involves a request to your provider, you do not have visibility, you have a reporting dependency.
2. Who holds the spend policy — who sets the per-user allowance, who approves an increase, and what happens when someone exhausts theirs mid-month? If nobody owns this, your budget control is currently an honour system.
3. Does my agreement say anything at all about AI consumption — who pays for it, who caps it, and what happens if it triples?

The uncomfortable conclusion is that the all-inclusive agreement was always a bet that the effort of running your IT would stay flat, and AI is the first thing in twenty years to make that bet obviously wrong. The tokens will reach you at cost. The governance around them — the allowances, the approvals, the cost engineering — is work nobody has priced. The providers who reprice openly, with a spend policy attached, will look expensive for about one quarter. The ones who stay silent will look cheap right up until the renewal where they are not.

Set the spend policy before the pilot. It is a far easier conversation than the one that follows the first invoice nobody expected.