Copilot Cowork. A powerful assistant, but not a tool for everything

Microsoft made Copilot Cowork generally available on June 16, 2026, and the reactions are exactly what could be expected. On one side, there is enthusiasm. More than half of Fortune 500 companies already use Cowork. On the other side, there is a reality check. The three month period when Cowork was available for free in the Frontier program is over. Normal operation is here, together with the question that had to be asked from the beginning.

Who is going to pay for it.

Copilot Cowork is paid through Copilot Credits. It is not another feature included in the license. It is agentic work that consumes compute capacity based on what you ask it to do. Billing started on June 16, 2026. Organizations whose users actively used Cowork in Frontier between March 30 and June 16 received a transition period until July 1, 2026, during which usage is not charged.

That is not wrong. It just needs to be understood. Above all, organizations need to understand what Cowork is, when it makes sense, and what they already have available in the existing license.

What is Copilot Cowork

Copilot Cowork is an agentic work environment inside Microsoft 365 Copilot. It differs fundamentally from standard Copilot Chat. Chat helps you explain, summarize, or draft something. Cowork is designed to do the work. It plans the steps, executes them across applications, and returns a finished result, not just a draft or recommendation.

Five properties form the foundation of Cowork.

• It runs in the cloud. Files are not stored locally, security is strong, and tasks continue even after the session is closed.
• Native Work IQ. Cowork is grounded in the company context and follows Microsoft 365 security and compliance rules. This includes emails, calendar, documents, Teams, meetings, search, and custom skills.
• Enterprise security and compliance. Cowork operates within the Microsoft 365 trust boundary, including audit log, DSPM, eDiscovery, Insider Risk Management, Data Lifecycle Management, which reached GA on June 22, and Communication Compliance. DLP is in preparation.
• Multi model design. At GA, it runs on Anthropic models including Opus 4.8 and Sonnet 4.6. In Frontier, OpenAI GPT 5.5 is available, and the model picker also supports the Sonnet plus Opus Advisor combination.
• Lower costs compared with competition. Microsoft internal testing across 125 test runs showed that Copilot Cowork is on average 30 to 40 percent cheaper than Claude Cowork with the Microsoft 365 connector. Both variants were tested on Opus 4.8.

Cowork 1 is coming soon. It is Microsoft’s own fine tuned model, post trained specifically for Cowork tasks. The goal is a significantly lower price while maintaining quality and removing model bias.

The key difference is simple. Copilot Chat advises and responds. Copilot Cowork plans and acts. You delegate a task, and it completes it end to end across multiple systems. That is why Cowork is not just about the model. It is mainly about orchestration of tasks, context, tools, permissions, and the ability to get work to a concrete outcome

Where and how Cowork helps

Cowork excels in situations where work consists of multiple steps, requires broad context, and needs a level of judgment that a simple chat or basic automation cannot cover. Practical examples include the following scenarios.

• Preparation for an important meeting. Cowork can review relevant emails, documents, records from previous meetings, calendar entries, and chat conversations. The output does not have to be only a summary. It can produce an agenda, a list of open points, a briefing document, and a draft follow up email and meetings.
• Project status and risk management. Cowork can process project documentation, emails, and meeting records, identify risks, prepare a status report, propose next steps, and generate communication tailored for different audiences, such as management, the team, or the customer.
• Sales pipeline and customer scenarios. During the Frontier program, one sales lead asked Cowork to investigate a stagnant pipeline. The result was a prioritized list of at risk opportunities with a precise description of where communication had gone cold, compressing a full week of manual review into one morning.
• Large scale comparison and documentation. During Frontier, an engineering team compared nearly 4,000 files between two product versions. That work would otherwise have taken weeks.
• Management agenda. Daily briefing, meeting preparation, priority synthesis, and regular risk overviews can save time. But this is also where the financial side starts to matter, because recurring automations generate ongoing credit consumption. See section 5.

Cowork is also the fastest growing feature in the history of the Frontier program. According to Microsoft, it has the highest user satisfaction of all Copilot and agentic experiences the company has ever released.

Cowork is not a silver bullet

When you need to drill a hole for a wall plug, you do not use a jackhammer. You use a drill. A jackhammer is powerful, but for that task it is unnecessary, expensive, and can cause more damage than value.

The same applies to Cowork. The fact that a tool can complete a complex task does not mean you should use it for every task. In practice, there is a clear line where the consumption meter starts. It is the shift from you working with AI assistance to AI working on your behalf.

• Included without additional cost. You open Copilot and ask it to summarize a document, draft an email, analyze a dataset, or do research. You are in the lead. You review the output.
• Metered and charged in credits. You ask Cowork to execute a multi step workflow from start to finish and return a finished result. AI is in the lead. It plans, executes, iterates, and delivers while you do something else.

This shift from assistant to agent is exactly the point where charging begins. It is also surprisingly easy to cross that line without the user noticing.
The principle is the right tool for the right task. Use Copilot Cowork only when a simpler tool is not enough. Otherwise, you risk paying too much for a trivial task.

What you already have in Microsoft 365 Copilot and when to use it instead of Cowork

Before an organization starts enabling Cowork broadly, it should ask one practical question. Are we using well what we already have in Microsoft 365 Copilot. The Microsoft 365 Copilot license already includes a significant part of AI productivity that many organizations still do not fully use.

Copilot Chat with frontier model selection. The model picker gives every licensed user access to Opus 4.8 and GPT 5.5 without metered consumption for standard chat interactions. Using one of the most capable AI models for a business conversation is included in the license price.

Agent Mode in Word, Excel, and PowerPoint. Since April 22, 2026, Agent Mode has been the default mode in all three apps. No new license. No additional cost. Agent Mode does not just help. It does the work. You describe what you want, and Copilot plans the steps, executes them, and returns the result. In Excel, it can run Python for complex data analysis. In the 30 days before GA, Excel engagement grew by 67 percent and retention by 50 percent.

Researcher. Built on the OpenAI deep research model in combination with Microsoft 365 context and Bing search. It synthesizes data from internal emails, meetings, files, and chats, as well as from the web and third party connectors such as Salesforce, ServiceNow, and Confluence. It includes Critique mode, where GPT generates the research and Claude reviews it, and Model Council, where both answers are produced in parallel together with a synthesis of agreements and differences.

Analyst. Built on OpenAI with Python execution. You attach an Excel file or CSV, ask a question in natural language, and Analyst writes and runs Python code, creates charts and visualizations, and documents every step.

Researcher and Analyst are preinstalled in the Microsoft 365 Copilot app and available to every licensed user, with up to 25 combined queries per user per month. The limit resets on the first day of each month.

Planner Agent. Helps with tasks, plans, priorities, and work status.

Custom agents through Agent Builder and Copilot Studio. Building is included in the license, and usage for licensed users in internal scenarios is also covered within fair use limits.

Overview table. Which tool for which task

Users should receive a clear map of these tools. They need to know what to use, when to use it, and what is cheaper or more expensive. Without that, people can easily start using the strongest tool even where a simpler one is enough, and burn credits unnecessarily.

Financial implications of Cowork

Billing structure

Copilot Cowork requires an active Microsoft 365 Copilot user subscription license, and then usage is billed based on Copilot Credits. This can be done through pay as you go at $0.01 per credit without commitment, or through a P3 annual commitment with a lower rate per credit. A prepaid Copilot Credit Capacity Pack is also available, with 25,000 credits per month for $200.

The cost of each task depends on four inputs. The model used, the volume of retrieved context, tool calls, and runtime.

Task categories. Light, Medium, and Heavy

Microsoft derived three task patterns from Frontier program data. They are not fixed boxes, but indicative calculations for budget planning.

These categories are complemented by four user personas with estimated monthly consumption. The estimates are based on Frontier program data as of May 27, 2026, and assume the Anthropic Opus 4.8 model. 

Source. Microsoft CustomerCoworkEstimator, aka.ms/CustomerCoworkEstimator

Why the Technical Worker persona is the most expensive

The Technical Worker persona does not have significantly more prompts. It has 35 per month, fewer than the Corporate Knowledge Worker with 38. But 14 of them are heavy. At the same time, $350 of the total $410 monthly cost comes from heavy tasks. The calculation is 14 heavy prompts times roughly $25, which equals $350. That is about 85 percent of total cost from one category. Fourteen heavy tasks across 21 working days is less than one per day. Even so, it is a conservative baseline that an active technical worker will probably exceed.

The asymmetry. Opus 4.8 included in chat vs. $25 for a heavy Cowork task

This asymmetry matters. The same model, Opus 4.8, is included in Copilot Chat. Delegating a heavy task to Cowork on the same model costs roughly $25, which equals 2,500 credits at $0.01 per credit. What changed is not the model. What changed is who is doing the work, how the work is orchestrated, which tools are needed, and how long it runs.

Scheduled tasks. The number the estimate does not show

The estimator models only interactive prompts. These are tasks that a user starts manually. It does not model scheduled tasks. Yet automation is one of Cowork’s key capabilities. A daily morning briefing, preparation 30 minutes before every meeting, a weekly pipeline summary, or a monthly competitive intelligence report. These tasks run regardless of whether the user is at the computer, and they consume credits at the same rate as interactive tasks.

Example. Daily email summary with suggested replies. For a leader who receives hundreds of emails per day, Cowork aggregates a high volume of data, evaluates priorities, and proposes context aware replies across threads. That is a heavy task at roughly $25 per run. With 22 working days per month, the calculation is 22 times $25, which equals $550 per month from a single automation. That is 4.5 times the entire modeled monthly cost for the Managers and Senior Leaders persona, which is $121.25.

In model scenarios that also include scheduled tasks, real consumption can be significantly higher than the baseline estimate. This is not an official Microsoft forecast. It is a practical warning that scheduled tasks can change the economics of the entire deployment.

Copilot Studio as another usage layer

Copilot Cowork is not the only usage layer. Organizations running Copilot Studio agents must also consider additional items that consume credits.

Note. Native Work IQ grounding inside a Cowork task is included in the task price. Work IQ API is charged only for custom builds outside Cowork.

The Cowork estimator and the Copilot Studio estimator at microsoft.github.io/copilot-studio-estimator model different things. Organizations running both, which means any serious AI deployment on Microsoft, need both to understand the full bill.

Estimate vs. reality. What to tell the CFO

The right approach is not to take an average persona, multiply it by all users, and send the result to leadership. Example. 100,000 users in one persona produces a number above $20 million per month. That will trigger a predictable reaction, such as freezing the rollout or scheduling a crisis meeting, but it is real math applied to an unrealistic input.

A proper estimate must reflect the real distribution of roles. Use all four personas in the estimator, not one. Add a line for scheduled tasks. Then pressure test the estimate against the value actually delivered.

How to manage limits and costs

Copilot Cowork is disabled by default. The administrator decides when it is enabled in the tenant and who gets access. At GA, Microsoft delivers cost management as one of the key areas, structured around three pillars.

Control

• Consumption limits at tenant, group, and user level. Admins define budgets, including user level caps inside group policies.
• Configurable alerts. Notifications when defined consumption thresholds are reached.
• User initiated credit requests. A user who needs more credits to complete a task can request them directly from Cowork.

Visibility

• Usage reporting at tenant, group, and user level in the Microsoft 365 admin center, through the new Cost Management Dashboard.
• Real time task price in credits. Planned soon after GA.

Efficiency

• Two payment options. Pay as you go at $0.01 per credit for flexibility, and P3 as an annual commitment with the lowest rate per credit.
• Model picker in Frontier. The option to manually choose a cheaper model for a specific task.

Recommended rollout approach

• First map the scenarios.
• Then determine the right tool for each type of work.
• Then enable Cowork only for pilot groups.
• Set limits and alerts.
• Only then expand based on value, not enthusiasm.

Architecture consideration. When Cowork is not the right architecture

Organizations that distinguish when to use Cowork and when to build deterministic software will have a structural cost advantage. Cowork is appropriate for ambiguous, judgment heavy, variable work. Deterministic software is more appropriate when the inputs and logic are known. A daily dashboard built in Cowork is generated from scratch every time. Cowork loads context from Work IQ, reasons across emails, calendar, files, and transcripts, and does so at model rates. Every day. Whether anything changed overnight or not.

The right architecture for such a dashboard is a database query, AI analysis only where judgment is needed, and data binding into a tested interface. Deterministic where possible. AI where it adds something a WHERE clause cannot. The cost difference between these approaches is not incremental. It is an order of magnitude.

Cowork makes it easier to build the first version. That does not make it the right choice for every recurring workflow.

Governance is not optional

Introducing Cowork moves Copilot from assistance into agentic work. That changes the requirements for governance. Topics that have often been underestimated now move to the foreground.

• AI Governance. Who can use which AI capabilities and for what purpose. Rules and policies for different generative AI scenarios.
• Agent Governance. Who can create agents, including skills, plugins, and Copilot Studio agents. Who approves deployment, who is accountable for outputs, and who owns the lifecycle.
• Agent FinOps. Who monitors and owns costs, who approves limit increases, and how ROI of agentic tools is measured.
• Token and Context Economics. How to work with context, file size, task frequency, model choice, and level of autonomy so that credit consumption makes economic sense.
• Adoption. Users must understand when to use Chat, Researcher, Analyst, Planner Agent, and Cowork.

A common bad habit is attaching a 30 MB PowerPoint template just to change the visual style of a presentation. A large file means more context, more processing, and higher consumption. For many scenarios, a clearly described skill with style rules and brand instructions would be enough. Context is not free. It matters whether you insert one page or a large set of documents into the task.

The future of Cowork. Model neutrality and declining costs

Copilot Cowork is a model neutral orchestrator. It is not tied to one LLM. At launch, Microsoft states support for Anthropic models including Opus 4.8 and Sonnet 4.6. Documentation for the model picker also mentions additional options, including GPT 5.5 and the Sonnet plus Opus Advisor combination. Availability of specific models may therefore vary by environment, release ring, and administrator settings.

The important point is that Cowork is not built on a single model. It is a work orchestrator that combines model, context, tools, permissions, and runtime based on the task type. This approach allows Cowork to gradually match the right model to each part of a task and reduce costs without sacrificing quality.

Three factors will continue to reduce cost.

1. Models will become cheaper, including the arrival of Cowork 1, which is designed for everyday tasks at substantially lower cost.
2. Cowork will improve at matching models and tasks, automatically selecting a more efficient model where it is enough.
3. Context retrieval and tool calls will become more efficient.

That does not mean governance can wait. It means the opposite. Organizations that learn to manage consumption now will have an advantage. They will know which scenarios create value, where Cowork saves hours of work, and where it only replaces a simple process in an expensive way.

What is worth watching

• The asymmetry between models included in chat and Cowork. Selecting a frontier model such as Opus 4.8 in Copilot Chat is currently included in the $30 license. Whether this remains fully included as usage grows is an open question.
• The 25 query limit for Researcher and Analyst. Today, it is a soft cap for included deep reasoning. What happens at query number 26 is not clearly documented.