Essay

When you pay for AI per task, checking the work is what counts

Microsoft and Atlassian are starting to bill AI that does work on its own by usage instead of per employee. That gives every task a price, and makes one question important: what does it cost you to check the result?

September 28, 2026 · Levy van Bremen · Lees in het Nederlands

On September 25, Microsoft presented a new version of Copilot, the AI assistant built into Word, Excel, Outlook and Teams. For business owners, the most important news was in the pricing. Chat and the help inside the Office apps stay under the fixed per-user license. Three new parts do not. Cowork takes a task from you and returns a finished result. Code builds an app or tracker from a description in plain language. Autopilot is a digital colleague: you give it a name, a role and a goal, and from then on it watches channels, follows up on threads and does recurring work by itself. For those three, you pay by usage.

On September 16, Microsoft had already told its resellers that new licenses for Copilot Business, the version for smaller companies, bought through a partner will come with usage-based billing by default from November 2. Pay-as-you-go, where you are billed afterwards for what you used, is then switched on, with a monthly limit Microsoft has set in advance. Atlassian, the maker of Jira and Confluence, the tools teams use to keep track of their tasks and knowledge, announced a similar shift on September 1. From December 3 it will meter three things: heavier AI features such as its chat and a coding assistant, automations per executed step, and requests that an AI agent handles completely without a person getting involved. By an AI agent I mean a program that carries out a task on its own, rather than just giving an answer. Plans come with an included amount; above that, you buy more or pay as you go.

What changes is the unit you pay for. Until now you bought access: a license per person, whether that person used it a lot or hardly at all. Now you increasingly buy work, and every task gets a price.

My argument: when you pay for AI per task, the time your people spend checking the result decides whether it pays off. The vendor's rate comes second. If you don't know that first number, you cannot judge any price per task.

Why the per-person license is disappearing

A fixed price per person worked as long as AI was a tool someone used while doing the work themselves. An agent that keeps working after that employee has gone home does not fit that model. One license can then cover a great deal of work, or almost none. The vendor is now shifting that difference to you.

Microsoft presents this as an advantage. According to the announcement, usage-based billing gives you full control to match cost to value. That is true, on one condition: you have to know what the value is. The fixed license never asked you that.

The same task, up to thirty times the cost

How unpredictable those costs per task are became clear on September 22 during an online session by the consulting firm McKinsey, which the business magazine Fortune wrote about the next day. Senior partner Lari Hämäläinen explained that an agent can take different routes to the same result. As a result, the same task can cost up to thirty times more from one run to the next. The models themselves are getting cheaper per unit, but according to him vendors are keeping part of those savings for themselves, through higher margins.

His advice: don't look at the price per chunk of text a model processes, look at the task. What does it cost to carry out, how often does it succeed, and how much time does it take a person to check the work? He added a rule of thumb. If a task takes an employee an hour and the agent's work can be checked in six minutes, an agent that succeeds more than ten percent of the time can already start to create value.

That sounds like a low bar, but do the math. Every attempt costs you six minutes of checking. If it fails, you still do the hour yourself. If the agent succeeds more than one time in ten, you save time. But turn those six minutes into half an hour, and the agent already has to succeed half the time. If checking takes almost as long as the work itself, it has to succeed nearly every time to deliver anything. The vendor's rate does not even appear in that calculation.

So checking is the big cost. And checking is something you can design.

I see this in what I build myself. Gemeenteregels, my service that answers questions about Dutch municipal rules, gives the article from the municipality's official rules with every answer, along with a link to the source. Anyone who wants to verify it clicks and reads. To me that link is the most important part of the answer: an answer an advisor cannot check quickly is one they end up looking up themselves. The second number, how often it gets things right, I measure separately. When I improved the part that puts municipalities side by side on September 24, I measured how often it chose correctly: first 76 percent, then 97 percent. Only with those two numbers, success rate and checking time, do you know what a task is worth.

The strongest objection: a fixed price is calmer

The best argument against paying by usage is that a fixed price is predictable and a meter is not. Business owners budget per year. A bill that moves with how often staff put an agent to work is hard to plan. On top of that, the vendor earns more when you use more. Microsoft says so openly to its resellers: according to the announcement, early usage creates a natural path to expanded consumption and upsell conversations. Since September 1, partners also get a bonus when their customers start using Cowork. The Dutch Microsoft partner Universal Cloud warned on September 17 that Cowork used to require a deliberate decision to switch it on, and that soon the door will be open by default, with a monthly limit you did not choose.

I take that argument seriously. The default setting favors the seller, and a bill after the fact is less calm than a fixed amount.

But the fixed price was less innocent than it looked. You also paid for licenses nobody used; McKinsey senior partner Tanguy Catlin named removing those as one of the ways to keep AI spending under control. And the computing power was paid for under a fixed price as well; you just couldn't see what for. Now you see what a single task costs. That makes possible a comparison that did not exist before: with what the same task costs you today in an employee's hours.

Stay sharp with the newest form too: paying per resolved request. Atlassian counts a request that an agent completes without a person needing to step in. That is a clear, measurable agreement. But a request that never reached a person is not automatically a request that was answered well. What counts as good is something you had better define yourself.

The choice to make now

If you buy Microsoft 365 through a partner, or your teams work in Jira and Confluence, this choice will land on your desk in the coming weeks. There are two things you can do now.

First, set the limits yourself, before the defaults do it for you. In Microsoft's admin center you can set spending limits for the whole organization and per user, and turn on alerts. Also look at the setting that automatically places new AI services under an existing budget. According to Microsoft, it is on by default. If you turn it off, you review every new service yourself first.

Then pick three recurring tasks you would like to hand to an agent. For each task, write down three things: how much time it takes an employee now, how much time it takes to check one result, and how you can tell it is right. If checking is almost as much work as doing, leave that task with your people, however low the price per task becomes. If checking comes down to opening a source or comparing an amount, you have found the task to start with.

The choice in front of you now is whether you keep buying AI as access per person, or as work per task. If you choose work, make sure you have those numbers before the first bill arrives, not after.

Sources

Is there work in your business that comes back every week? That is what I build software for.

Book a call →