Copilot Cowork reached general availability on 16 June 2026. I first wrote about it in March, before launch, when most of what anyone could say was a forecast. Three months of a shipping product is a better basis for advice, so I have rewritten this post from scratch rather than add another update to the top of it.
The short version: Cowork is a real product that does useful multi-step work inside Microsoft 365. It is also metered, it acts with the permissions of whoever is using it, and it has changed a good deal in its first three months. Whether your business needs it depends less on what it can do than on whether you have somewhere sensible to point it and someone watching the meter.
The distinction that matters has not changed. Copilot Chat answers a question and leaves the next step to you. Cowork takes a goal, breaks it into steps, carries them out across Outlook, Teams, Excel, SharePoint and OneDrive, pauses at checkpoints for approval, and logs what it did. One answers; the other acts.
What has changed is where you find it. Chat and Cowork are now two tabs of one experience rather than separate destinations. That sounds cosmetic and is not. When delegation sits one click away from the chat box everyone already uses, people try it, including the people you had not planned to pilot it with. The governance work below needs doing before that happens, not after.
For the timeline, because older articles (including the earlier version of this one) blur it: Microsoft 365 E7 and Agent 365 reached general availability on 5 May 2026. Cowork followed on 16 June.
The App skill is the one that changes the wider conversation, because “Copilot builds a small internal app” overlaps with work that used to need a developer. I have covered where that leaves the build-or-buy decision in Copilot vs custom AI: where the line sits now.
At launch the story was that Cowork ran on Anthropic’s Claude. It is now a choice. The model picker is split into a GPT section and a Claude section, and September added a frontier model to each: Claude Fable 5.1 from 1 September and GPT-6 Astra from 4 September.
The detail that is easy to miss is that Microsoft limits both of those models to Cowork and Copilot Studio. They are not available across Copilot generally. So the question for your acceptable use policy is no longer “which AI vendors do we approve?” but “which model, on which surface?” Claude has been on by default in Word, Excel and PowerPoint since May and selectable in Copilot Chat since July. Fable 5.1 and Astra are a different tier, in different places. A policy that names approved models per vendor is already out of date.
There are two layers, and the earlier version of this post only described one of them.
The licence. Access rides on a Microsoft 365 Copilot licence on top of your existing subscription. The standalone licence is around £24.70 per user per month in the UK (US$30 list), a figure the 1 July price round did not change. Smaller organisations have cheaper routes to a Copilot licence: the Copilot Business add-on at US$21 (roughly £16.10) now that the introductory rate ended on 30 September, and the Business Standard and Business Premium “with Copilot” bundles at US$23.50 and US$32. Confirm that the SKU you are buying includes Cowork before you choose on price. At the top end, Microsoft 365 E7 bundles E5, Copilot and Agent 365 at US$99 per user per month. Copilot Pro, for the avoidance of doubt, is the consumer product and is not a route to Cowork.
The meter. Agentic work is counted in Copilot Credits: US$0.01 per credit pay-as-you-go, or US$200 a month for a prepaid pack of 25,000. A classic answer is about 1 credit, a generative answer 2, an agent action 5, grounding on Microsoft Graph data 10, and an autonomous action, one triggered with nobody in the loop, 25. Users who hold a Microsoft 365 Copilot licence do not draw credits for classic answers, generative answers or Graph grounding inside Copilot Chat or Teams. That exemption is specific to those three items, so do not assume everything a licensed user delegates is free.
An illustration, using Microsoft’s published rates. An Automation that fires once each working day, grounds itself on tenant data, writes a summary and takes one action is roughly 25 + 10 + 2 + 5 = 42 credits a run. Over 21 working days that is under 900 credits, or about US$9 a month at the pay-as-you-go rate.
Run the same shape 3,000 times a month, once per inbound email for example, and it is around 126,000 credits: five or six prepaid packs. The task did not change. The volume did.
Cowork is cheap for occasional, person-scale delegation and expensive as a batch engine. The fuller comparison is in Power Automate vs Copilot Studio vs Claude API.
Two September changes make this manageable. Per-task credit visibility means the cost is visible before the run rather than on the invoice. And credit governance now has real controls: limits can be set per agent, and there is a tenant-wide view of credit spend. One thing to know before you let people loose in Copilot Studio: credits are now consumed while makers build agents, not only when the agents run.
The Consumption Dashboard now reports Cowork assisted hours and value across eight task types, and Microsoft has published the methodology behind it. As with the App skill, I cannot confirm it has reached every tenant; it may be in targeted release.
If you have it, use it, but be clear about what it is. It is a modelled estimate of time assisted, produced by the vendor selling the product. I treat it as a conservative proxy to baseline against: note the figure each month, read the methodology, and watch the trend. I would not put it in front of a finance director as realised savings. For that you still need your own before-and-after measure of a specific process, which is the approach in how to measure AI ROI. Data export from the Copilot and Agent 365 dashboards, also new, makes it easier to keep both sets of numbers in one place.
Worth piloting if: your people already hold Copilot licences; the work lives in Microsoft 365; the tasks are multi-step and scoped to one person or one team; a human reviews the result; and volume is low to moderate. Preparing for a client meeting from email, files and a tracker, or pulling a weekly status pack together from three sources, is the kind of task Cowork handles well.
Not the right tool if: the process is fully rule-based, in which case Power Automate does it without a meter; the work is high-volume and unattended, where per-token pricing on a custom build is far cheaper; the data sits outside Microsoft 365 in a CRM, finance or line-of-business system; the output goes straight to customers; or your permissions are not yet in a state where you would want an agent acting on them.
If you do pilot, keep it small: five to ten people with the messiest cross-app work, a credit budget, a baseline of how long those tasks take today, and a fixed date to decide. There is no prize for being first to switch on a metered agent across the whole company.
Three months in, Cowork has earned a place in the toolkit, and that place is specific: person-scale and team-scale delegation, inside Microsoft 365, with someone reviewing the result. None of it helps if nobody owns the budget or the permissions underneath it.
If you are deciding whether to pilot it, start with the list of tasks, not the licence. That is the first thing I ask for in a Microsoft 365 AI readiness review.
I help UK SMEs work out which tasks belong in Copilot Cowork, which need a custom tool, and what each will cost to run, with the credit maths done before anything is bought. Book a free 30-minute discovery call.
Book a Free Discovery Call