Copilot billed on usage from 1 December: how an SME keeps control of its AI budget
It is now confirmed: from 1 December 2026, new Microsoft 365 Copilot Business licences sold through resellers will default to usage-based billing. Part of your office AI becomes a variable cost.
For a Swiss SME this is neither bad news nor good news. It is a change of nature: AI leaves the logic of the fixed subscription for the logic of the meter. And a meter is something you set before you open it.
Here we detail what changes, what an agent really costs, and the four settings to configure before switching anything on.
What Microsoft changes on 1 December
The mechanism is called usage-based billing. Microsoft announced it to partners on 16 September, then moved the date from 2 November to 1 December, according to the update published in the Partner Center on 14 October 2026.
In practice, every new Copilot Business licence bought through a CSP reseller now comes with pay-as-you-go switched on. Direct purchases from Microsoft have been affected since 19 October.
A ceiling is set by default: 4,000 Copilot Credits per user per month. This ceiling is a limit, not an invoice. You only pay for what is consumed, and the administrator can change it.
Eleven markets are excluded at first, including France, Germany, Belgium and the Netherlands. Switzerland does not appear in the published lists we consulted. Check your own situation with your reseller before placing any order.
The classic Copilot features in Word, Excel, Outlook or Teams remain covered by the fixed licence. What moves to the meter are the advanced uses: Copilot Cowork, the Work IQ APIs and the agents built in Copilot Studio.
One credit, one cent: what an agent really consumes
The Copilot Credit is the common unit of account for every Microsoft service billed on usage. At list price, one credit is worth 0.01 dollar, billed monthly in arrears. The 4,000 credits of the default ceiling therefore represent at most 40 dollars per user per month.
The published grid for Copilot Studio agents gives the scale:
a predefined answer: 1 credit;
a generative answer: 2 credits;
an agent action, meaning a call to a tool or a system: 5 credits;
tenant graph grounding, meaning a search across your documents and emails: 10 credits;
document processing: 8 credits per page.
A complex question put to an agent that digs through your SharePoint files therefore costs 12 credits: 10 for grounding, 2 for the answer. That is 12 cents. It is not expensive. It is above all unpredictable if nobody is watching.
Microsoft itself recommends a buffer of 10 to 20% above the estimate. The problem is not the unit price, it is the silent multiplication. A useful agent is an agent people call on, and an agent people call on consumes.
The real risk is not the invoice, it is the absence of a rule
We see the same pattern in most of the SMEs we work with: AI comes in through one subscription, then a second, then through uses nobody has arbitrated. Moving to the meter makes this blur visible, and that is a good thing.
Three situations come up often.
The agent launched as a test, never switched off, running for three people while twenty have access to it.
The automated flow that processes documents every night whose output nobody reads.
The enthusiastic user who connects the agent to the whole SharePoint and multiplies the cost of every question by five.
In all three cases the invoice is only the symptom. The cause is the absence of an owner per use, a measurable objective and an end rule. This is what we call usage governance, and it comes before any technical setting.
One side effect to anticipate: when a user reaches their limit, they lose access to the covered agents until the first day of the following month. A badly calibrated ceiling can therefore interrupt a business process at the worst moment.
Example: a Geneva accounting firm with 40 people
Take an accounting firm of 40 employees, equipped with Copilot Business, that wants to deploy two agents: an assistant answering client questions grounded in its files, and a flow that pre-checks accounting documents.
Realistic assumption: 15 employees use the assistant, 20 questions a day each, every one grounded in the tenant graph. That is 15 times 20 times 12 credits, or 3,600 credits a day and roughly 75,000 credits per working month.
The accounting flow processes 2,000 pages a month at 8 credits a page, or 16,000 credits. Total: about 91,000 credits, so 910 dollars a month at list price, before any reseller discount.
With the default ceiling, these 15 users share 60,000 credits between them. The ceiling would be hit around the 20th of the month and the assistant would stop dead. Conversely, the other 25 employees would carry a reserve of 100,000 credits they never use.
The right answer is neither 4,000 for everyone nor unlimited. It is an overall budget of around 110,000 credits, a limit per group and an alert at 70%. With one prior question: do those 20 questions a day replace 30 minutes of searching, or only five?
Four settings to configure before switching anything on
Everything is set in the Microsoft 365 admin center, under Copilot then Cost Management, and in the Power Platform admin center for Copilot Studio agents. Here is the order we recommend.
Set an overall monthly limit for the organisation, then a limit per security group, rather than a uniform limit per user.
Restrict Cowork and each agent to a named group, and turn off the option that auto-applies new services.
Set an alert at 50% and at 80% of the budget, and name the person who receives it and decides.
Export the credits report every two weeks, reconcile it with actual use, and cut what serves no purpose.
Add an estimator to this: Microsoft provides one for Cowork and one for Studio agents. A ten-minute calculation before activation avoids the end-of-month debate.
What we take away
Usage-based billing is not specific to Microsoft. Mistral has just published per-minute prices for its voice agents, and most vendors are heading the same way. AI in business is becoming a variable cost, and a variable cost has to be managed.
For an SME this means three things: an owner per use, a budget per group, and a short monthly review. This is not IT, it is management.
Want to know where your organisation stands on the governance of its AI uses? The AI Barometer gives you a first reading in a few minutes, and we can then talk it through.
Comments