What Are Microsoft Copilot Credits, and Why Are They Losing UK Businesses?

Somewhere in a boardroom, right now, a Finance Director is staring at a line item that says “Copilot Credits” with no real idea what it means.

It isn’t a sharpness problem. These are sharp people. It’s that Microsoft has introduced a pricing currency for Copilot Cowork, their agentic AI offering, that’s genuinely, structurally impossible to benchmark.

Here’s what Copilot Credits actually are, why the pricing is a problem, and why seven businesses I know of have either reduced or cancelled their Copilot Cowork deployments as a direct result.

What are Copilot Credits?

When you use Copilot Cowork, Microsoft’s AI agent product that can take action across your M365 environment, you’re not paying per seat any more. You’re paying per task, in a currency called Copilot Credits.

Microsoft deducts credits from your account each time Copilot Cowork completes an agentic task. The more complex the task, the more credits it burns through.

In principle, that’s a reasonable idea. Consumption-based pricing means you pay for what you use, and for businesses with highly variable AI usage, it could genuinely be more cost-effective than a fixed subscription.

The problem: nobody knows what a credit is worth

Here’s the core issue with Copilot Credits: Microsoft has not published a conversion rate from tokens to credits.

When you use Claude, ChatGPT, or almost any other AI tool on a consumption model, you pay in tokens. A token is a defined unit, roughly three-quarters of a word in most models. Token prices are published. You can calculate exactly what a task will cost before you run it, and model your monthly bill at different usage levels with real confidence.

With Copilot Credits, none of that is possible. There’s no official figure for how many tokens equal one credit, and no published conversion from credits to pounds, dollars, or any other currency that means something to a Finance Director.

What you can do is buy credits in advance, in packages priced by Microsoft, without knowing how many credits a typical task will consume until you’ve already spent them.

That’s not a pricing model. That’s a black box.

Why this is a boardroom problem

Most AI purchasing decisions in UK SMEs go through some version of board review. The board asks reasonable questions: what are we spending, what are we getting, how does this compare to the market?

With a transparent pricing model, those questions are answerable. With Copilot Credits, they’re not.

We sat in a boardroom with a client who’d deployed Copilot Cowork across part of their team. The board wanted to understand the value they were getting and asked the IT Director to benchmark the cost. He couldn’t, not through any lack of skill, but because the information required to benchmark it simply doesn’t exist in publicly accessible form.

The result was a loss of confidence in Copilot’s billing specifically, rather than in AI generally, and that loss of confidence cost Microsoft seven licences from that one client alone.

The cost comparison we actually ran

To give you a real-world data point: we ran a controlled test.

Same task. Same underlying model (Opus 4.8, Anthropic’s flagship at the time). Same brief. We ran it through Claude Cowork and Copilot Cowork and compared both the outputs and the costs.

Claude: $0.92. Copilot Cowork: $2.22.

More than double. For the same output quality.

I’d expected a surcharge. Microsoft adds integration value, enterprise support, and procurement simplicity, and a reasonable premium for that would have made sense. More than double wasn’t what I’d modelled, and I’ve been watching AI pricing closely for two years now.

What you should do

If you’re already on Copilot Cowork:

Track your credit consumption carefully and correlate it against the tasks that consume the most. A small number of task types will usually account for a large share of credit usage. Knowing which ones are expensive helps you decide what stays on Copilot Cowork and what gets routed elsewhere.

Ask your Microsoft account manager for a cost breakdown. They can’t give you the global token-to-credit conversion, but they can give you consumption data that helps you understand your own patterns.

If you’re evaluating Copilot Cowork:

Don’t make the decision without running a cost comparison against Claude or another alternative for your specific use cases. The headline licence cost isn’t the number that matters. The cost per task at realistic usage volumes is.

If you’re presenting AI costs to a board:

Use tools with transparent pricing wherever you can. If you need Copilot Cowork for specific reasons, be upfront about the pricing limitations and commit to reporting on credit consumption alongside business outcomes. Boards can work with uncertainty as long as it’s disclosed. What erodes confidence is finding out about it later.

The bigger picture

The Copilot Credits situation is a useful lens on something broader about AI procurement.

The AI market is moving fast enough that vendors are making pricing decisions quickly, under competitive pressure, without always prioritising customer clarity. Consumption billing is genuinely the right model for AI at scale, but only if customers can actually understand what they’re consuming.

The businesses that get the best results from AI investment over the next three years will be the ones treating AI procurement like any other significant business spend: clear metrics, transparent costs, and regular review against outcomes.

That discipline is available to you regardless of which tool you’re using. The Copilot Credits situation is simply a reminder that it matters more now than it used to.

Artificia1 is a vendor-neutral AI consultancy helping UK SMEs make smart AI investment decisions. Get in touch if you want an independent view on your AI tool stack.