When does experimenting with artificial intelligence end – and when does productive use begin?
For many organisations, that point has now arrived. On 16 June 2026, Microsoft announced the worldwide general availability (GA) of Copilot Cowork, and the new usage-based pricing model takes effect from 1 July 2026. This shifts the question from “What can AI do?” to “Which tasks do I delegate deliberately – and what will it cost me?”. This Copilot Cowork pricing model is exactly what determines whether an experiment becomes a predictable productivity lever.
In this article, I will put both sides into perspective: what Cowork does fundamentally differently from a technical perspective, and how you can calculate costs reliably.
What Copilot Cowork really is
What may initially sound like another AI feature is a new concept within Microsoft 365: Cowork carries out complex, long-running tasks using several tools. You define the work – and Cowork completes it from start to finish and returns a finished result. Not a draft, not a recommendation, but a completed work product.
According to Microsoft, Cowork is the fastest-growing feature in the history of the Frontier programme and one of the best-rated Copilot experiences the company has ever delivered. Following three months of preview, Microsoft states that more than half of Fortune 500 companies are using Copilot Cowork, including Accenture, Avanade, Capital Group, Koch, Ooredoo Qatar and Zurich Insurance.
The examples demonstrate the range of use cases: one engineering team taught Cowork how to safely edit batch-job tables and create dependency flow diagrams after every change. Another team compared almost 4,000 files across two product versions – work that had previously taken weeks. A sales leader directed Cowork towards a stalled pipeline and received a prioritised list of at-risk opportunities, including the follow-up email due for each one – what had been a week of manual analysis became a single morning’s work.
Five factors – and one important caveat
Cowork is designed to work more accurately, securely and cost-effectively than comparable approaches. Five points make this possible:
- Cloud hosting: Files are not stored locally, security policies are consistently enforced, and tasks continue running even when the laptop is switched off.
- Work IQ: grounds every task in the systems your organisation already uses.
- Enterprise security: Cowork operates within the Microsoft 365 Trust Boundary.
- Multi-model design: Cowork selects the model required for each individual task.
- Lower costs: through a runtime that efficiently identifies the right information and tools.
The multi-model design is one of the strongest differences compared with other solutions: you are not tied to a single model. The following models are available at GA launch:
| Model | Availability | Characteristics |
|---|---|---|
| Anthropic Claude Opus 4.8 | GA (standard) | Most powerful model, basis for the pricing calculator |
| Anthropic Claude Sonnet 4.6 | GA | Faster and more cost-effective for medium-sized tasks |
| GPT 5.5 | Frontier programme | OpenAI’s latest model for early adopters |
| Cowork 1 | Coming soon | Microsoft’s own, significantly less expensive model |
Cowork 1 is expected to be released in the coming weeks as a secure, fine-tuned model. It is optimised for a substantially lower price and explicitly addresses the reduction of model bias. For cost-sensitive, everyday workloads, it is likely to become the most economical option.
At GA launch, nine new partner plugins are also available (Enosix, Harvey, LSEG, Miro, monday.com, Moody’s, Morningstar, S&P Global Energy and TeamsMaestro); others including Adobe, Atlassian, Box, Canva, CB Insights, Databricks, MoneyForward and Templafy will follow. Within the Frontier programme, Cowork can also search the web using a local Edge browser while complying with the corporate policies already applicable to users.
One caveat I want to state clearly: even at GA, there is no EU Data Boundary – for EU customers, data processing does not necessarily take place within the EU. This is an exception we had not expected, as regional processing is normally ensured by the time of GA. For regulated industries, this point should be carefully assessed before activation.
The pricing model: two layers, clearly calculable
The pricing model follows a two-layer principle. Once understood, budgets can be planned reliably.
Layer 1: the fixed licence
Copilot Cowork requires the Microsoft 365 Copilot User Subscription Licence (USL). This includes the complete AI productivity experience – Copilot Chat, Copilot in Word, Excel, PowerPoint, Outlook and Teams, the Work IQ context engine, the multi-model system, prebuilt agents such as Researcher and Analyst, as well as custom agents from Agent Builder – for a predictable monthly fee per user. The following prices currently apply in Germany (plus VAT):
| Plan | Promotional price (until 30 June 2026) | Standard price |
|---|---|---|
| Microsoft 365 Copilot Business | €15.60 / user / month | €18.20 |
| Microsoft 365 Copilot Enterprise | €27.60 / user / month | approx. €30.00 |
Layer 2: Copilot Credits (usage-based)
In addition to the fixed licence, Cowork usage is billed based on consumption, denominated in Copilot Credits. The price per task is determined by four elements: model usage, context retrieval, tool calls and runtime.
- PayGo: $0.01 per credit – at the current exchange rate, approximately €0.009 per credit.
- Capacity Packs: $200 per tenant per month for 25,000 credits; paid in advance and expiring monthly.
- P3 plan (pre-purchase): annual volume with a discount for larger quantities.
Microsoft describes three typical task patterns. The following credit ranges come from the GA blog and are illustrative estimates based on Opus 4.8 – not fixed billing tiers:
| Task type | Example | Credits | Approx. cost (PayGo) |
|---|---|---|---|
| Light | Monday status report | 100–300 | $1–3 (€0.93–2.79) |
| Medium | Meeting brief + Excel + slide | 400–700 | $4–7 (€3.72–6.51) |
| Heavy | 6 months of usage data → executive report | 700+ | $7+ (€6.51+) |
For practical use, content processing is particularly tangible: it costs 8 credits per page. This makes document analysis very easy to calculate:
| Task | Credits | Approx. cost (PayGo) |
|---|---|---|
| Analyse one PDF page | 8 | ~€0.07 |
| Summarise a 10-page document | 80 | ~€0.74 |
| Fully analyse a 50-page document | 400 | ~€3.72 |
| Weekly report from emails and appointments | 100-300 | ~€0.93–2.79 |
| Prepare a customer proposal (multiple sources) | 350 | ~€3.26 |
| Analyse and prioritise a sales pipeline | 550 | ~€5.11 |
| Monthly report for management (Opus 4.8) | 900 | ~€8.37 |
To model expected costs for your own organisation, you can use the Copilot Cowork Cost Calculator developed by Anastasios Ntaflos.
According to Microsoft, Cowork is noticeably more cost-effective when using the same models than operating Claude Cowork through the Microsoft 365 connector.
Cost control: you retain control
From my perspective, the most important signal for IT is that Cowork is disabled by default. You decide when it is activated in the tenant, who receives access and how much may be spent.
- Spending limits at tenant, group and user level through defined billing policies, including user-specific limits within group policies.
- Customisable usage alerts: You define the thresholds and determine who is notified as soon as spending reaches that level.
- Transition period: Tenants in which at least one user used Cowork in the Frontier programme between 30 March and 16 June 2026 will not be charged before 1 July 2026.
Management is handled via the new Cowork navigation item in the Admin Centre. There, you define cost profiles and limits and link the relevant Azure subscription.
Conclusion
Copilot Cowork is not a chatbot upgrade. It is a step towards genuine, AI-supported task delegation – secure, traceable and directly integrated into the familiar Microsoft 365 environment. The usage-based model makes costs transparent through credits: those who know which tasks they delegate can forecast spending precisely. For a medium-sized company in Germany, this often means only a few euros per complex task – with a return on investment that pays off as early as the first working day saved.
The decisive step therefore no longer lies in the technology, but in discipline: clearly define which tasks are suitable for delegation, set limits properly and – particularly for regulated industries – consciously address the question of data processing outside the EU. This is exactly where we support organisations at evoila, from assessment through to controlled implementation.
