M365 Copilot Cowork pricing: a barrier to business adoption?

Microsoft has introduced some genuinely welcome improvements to Microsoft 365 Copilot.

The arrival of newer models, including GPT 6 Astra in Cowork, gives users access to stronger reasoning and more capable AI. Copilot is also moving beyond answering questions and drafting content. Cowork can plan work, use tools, connect information and take action.

That should be an important step forward. Unfortunately, the way Microsoft charges for it risks putting many businesses off before they have discovered whether it is useful.

The problem everything is Cowork is metered and charged for

I understand why Microsoft is introducing usage-based pricing.

Advanced AI tasks require substantial computing power. Providers need to create sustainable products, and businesses should expect to pay when AI completes valuable work on their behalf.

The problem is not that Cowork costs money. The problem is that users cannot reliably know what a task will cost before they start it.

Microsoft measures Cowork usage in Copilot Credits. According to its guidance, credit consumption can depend on factors including the model selected, the volume of context retrieved, the amount of orchestration involved and the tools used. Higher reasoning settings can also use more credits.

In other words, the price depends on decisions made by the system while it is doing the work.

Microsoft does provide cost controls. Administrators can set limits for users and groups, monitor consumption and restrict the models available. Users can also enter /cost during a Cowork task to see an approximate total for the work completed so far.

But Microsoft confirms that you cannot check the cost before running a task. The /cost figure is also an estimate, rather than the official billing record.

That is visibility after spending has started. It is not price certainty.

The same prompt may not mean the same bill

The uncertainty becomes more difficult when we consider how generative AI behaves.

These systems are not entirely consistent. Give an AI tool the same broad task twice and it may not approach it in exactly the same way. On one occasion, it might produce a fairly direct answer. On another, it might retrieve more information, call additional tools, reason for longer or decide that a more detailed piece of work is required.

Microsoft gives users some control through model choice and reasoning levels. Its own guidance explains that usage varies according to the model, context, orchestration and tools involved. But most employees will not understand how those factors translate into credits, let alone pounds and pence.

Depending on the task, I have seen AI work cost anything from a few pence to £10, £20 or £30. That might be reasonable when the result creates enough value. It is much harder to accept when the user believed they were starting a simple task and the system decided to go considerably further.

Imagine asking a member of staff to use Cowork without being able to tell them whether the task will cost 50p or £20. Now multiply that uncertainty across a team.

Yes, you can impose spending limits. But a limit only controls how much you can lose. It does not tell you whether you are receiving good value from each task.

Experimentation is part of adoption

This matters because most organisations are still experimenting with AI.

They are working out which tasks are suitable, where human judgement is still essential and whether the time saved justifies the cost. Employees need room to try things, make mistakes and develop their understanding.

Unpredictable usage charges change that behaviour.

Instead of thinking, “Let’s see whether Cowork can help with this”, people begin thinking, “How much might this cost me if it goes too far?”

That creates hesitation at exactly the point when Microsoft needs users to build confidence.

The organisations I speak to are not rejecting AI. They are becoming cautious about Cowork. Some are looking at how much they can complete in the main Copilot chat interface, with people stitching the thinking together and taking the final actions themselves.

That is less automated, but it feels more controllable.

Are businesses paying too much for the final action?

Cowork’s attraction is its ability to do more than produce an answer. It can work across information, use tools, send notifications, support automated tasks and respond to scheduled or event-based triggers.

Those actions can be useful. But in many cases, the most valuable part of the task has already happened in the chat.

The AI has helped someone analyse the information, structure the work, draft the email or decide what should happen next. A person can then review the result and take the final action.

That human step may take a few minutes. It also provides a useful accuracy and judgement check.

For some high-volume or repeatable processes, paying Cowork to take the action will make complete sense. For occasional tasks, the economics are less convincing. Businesses may conclude that the final layer of automation is not valuable enough to justify an uncertain bill.

Microsoft has built controls, but not confidence

Microsoft deserves some credit here. Administrators can create spending policies, set monthly limits and review consumption through the Microsoft 365 admin centre. Cowork also includes model and reasoning controls that can help users balance quality, speed and cost.

You can download Microsoft’s Copilot Credits guide for more clarity, but it's not really compltely clear! The documentation still leaves the most basic business question unanswered:

“What will this task cost before I approve it?”

Until users can answer that question with reasonable confidence, I think many organisations will restrict Cowork, reserve it for selected people or avoid it altogether.

My concern is what happens next

I use Cowork, but most of my day-to-day work still happens in ChatGPT. My subscription includes enough working capability for me to experiment, understand where the tool performs well and build confidence before committing extra credits to a more advanced task.

That experimentation matters.

Businesses need a similar route into agentic AI. Give people a meaningful included allowance, clearer upfront estimates or a predictable way to test real tasks. Let them experience the value before asking them to approve an open-ended cost.

Otherwise, Microsoft risks building an impressive new way of working that business leaders are too nervous to switch on.

Am I being too cautious? Are you using Cowork in your organisation, and do you feel confident about controlling what it costs?