Meeting preparation with Cowork: connecting evidence, analysis, and action

People connect the applications involved in meeting preparation

Preparing for a customer meeting can mean finding commitments in email, checking figures in a spreadsheet, and explaining them in slides. Even after the materials are ready, someone must reserve preparation time and select the recipients and version to share. Each operation may be small, but between them a person decides what a figure supports and how much to tell each recipient.

The starting question for Cowork is how much of this connecting work can be delegated. Announcements from Anthropic and Microsoft describe products that handle several documents or applications within one assignment. This article uses a hypothetical meeting preparation task to examine the move from reading information to analysis, artifacts, and changes in business applications.

In my interpretation, value depends on more than how long an agent runs. It matters whether preparing the materials takes less work and whether a person can check their basis and move to the next decision. If more plausible artifacts leave the reviewer reconstructing their relationship to the original information, the amount of work successfully delegated remains limited.

The same name covers different sources of context and places of execution

Anthropic's January 12 release notes introduced Cowork as a research preview in the Claude desktop application. Its isolated virtual environment handles local files and external connections through MCP. Files the user permits it to access become the objects of work, extending beyond a workflow where the user creates files from an answer.

Microsoft's March 9 Copilot Cowork announcement describes using WorkIQ to draw on email, meetings, messages, files, and other work context. A request becomes a plan with checkpoints for confirmation, changes, or pausing. Examples include saving a meeting brief, analysis, and slides, as well as proposing Outlook calendar changes and applying them after approval.

Microsoft describes execution in its Microsoft 365 cloud, using existing identity, permission, and audit mechanisms. Access is a research preview for limited customers, with expansion through the Frontier program announced for late March. These are descriptions of architecture and availability plans; they do not measure the success rate of an entire meeting preparation task or establish superiority over another product.

The distinction affects which information can be used without gathering it again. A meeting whose materials sit in a local folder needs different connections from one whose materials are spread across business applications. Permission to read a document also leaves open whether it concerns the right customer and period. The name alone cannot establish the execution conditions; the relevant question is whether the configuration reaches the information the job requires.

Moving analysis from Excel to PowerPoint still requires checking its assumptions

Anthropic's February 24 announcement introduces organization plugins, additional connections, and command forms that organize input fields. It also describes an early research preview for sharing conversation context between the Excel and PowerPoint add-ins. This is a concrete example of carrying an assignment's background from analysis into presentation work. The add-in capability cannot be generalized to every Cowork connection.

In this article's hypothetical example, an account manager prepares a monthly revenue workbook and slides for a customer renewal meeting. Suppose the workbook contains both recognized revenue and the expected value of deals that have not closed. Adding them under “revenue” can produce an arithmetically correct but misleading explanation. Carrying spreadsheet context across applications and choosing the correct business definition are separate problems.

The handoff needs more than a total. The reporting period, customer scope, distinction between confirmed and expected amounts, and missing materials all affect what the presentation can claim. Reducing manual copying has value, but a shared mistaken assumption can spread to both the workbook and slides. Once the artifacts exist, a reviewer needs a way to trace figures through to conclusions.

Computer use adds a route to applications without a direct connection

Anthropic's March 23 computer-use announcement describes a research preview in which Cowork and other supported tools can view a screen, click, and type. It says Claude first uses an available connector and turns to screen interaction when one is unavailable. It asks permission before using a new application, and the user can stop it during the task.

Direct connections and screen interaction provide different routes to the material. If part of our hypothetical meeting preparation depends on a dedicated application screen, retrieving that information could become part of the assignment. Anthropic also describes screen interaction as slower than direct integration and acknowledges errors in this early feature. More reachable applications therefore do not establish that an entire job will finish sooner.

The useful comparison is the operations required to gather the same materials and the amount a person must check. Additional navigation and searches can still have value if they reduce the user's effort moving between applications. If each run requires locating the source again and a person repeatedly replaces the wrong documents, direct connections or manual work may be easier to manage.

Gather information, explain the analysis, and apply changes

Applying these product examples to meeting preparation suggests three roles within the work. This is the article's analytical framework. It does not prescribe a mandatory three-stage sequence or require three separate agents.

The purpose of these roles is to locate where supporting context was lost. Correct figures for the wrong customer require revisiting document selection. A correct workbook with outdated slide text points to the handoff from analysis to explanation. A person can return to the part that needs correction instead of repeating the entire process.

The convenience of one assignment is compatible with inspecting intermediate results. As more work is delegated, it becomes more valuable to preserve a clear account of what has been checked and what remains unresolved.

When new figures arrive, can the full set of materials be updated?

Meeting preparation does not always end after executing the original request. Continuing the hypothetical example, confirmed revenue adjusted for returns might arrive after the slides are drafted. Updating only the spreadsheet total leaves part of the job unfinished if the chart, explanation, and draft sharing email still contain the old figures.

Automatically favoring the newest material would create another problem. If the latest workbook covers a different period, replacing the earlier figures could invalidate the comparison. The user needs to confirm the period and revenue definition and decide which assumption to change before the related artifacts are revised. The operating approach proposed here makes those relationships traceable.

For example, showing the changed table together with its dependent chart, affected slides, and unsent email would help the reviewer understand the extent of the revision. Neither company has thereby demonstrated automatic updates to every dependency. A trial should also evaluate whether the system identifies the places a change needs to reach.

Confirmation before applying a change should refer to a version. Approval for the earlier materials should not be reused to send revised content without the necessary review. If a recipient already received a document, deciding whether to send a correction is a separate task. Stopping the workflow on screen does not retract an earlier delivery.

Include review and correction when measuring delegated work

A trial of this kind separates the account of the work from its actual results. Anthropic's guide to evaluating agents likewise distinguishes the transcript of conversation and operations from the outcome left in the environment. For our meeting preparation task, checks would include opening the files, comparing the figures in tables and explanations, and verifying the intended time in the calendar.

Compare the task with a person preparing the same materials, recording time spent finding information and checking its basis, as well as the number of corrections. This is the article's evaluation proposal, not a measurement of either product. A fast first draft may leave total work unchanged if the reviewer must locate every source again. Conversely, a draft that identifies uncertainty can help establish which issues the meeting needs to resolve.

Anthropic's February 24 announcement also includes OpenTelemetry support for observing usage, cost, and tool activity. Such information can help locate where activity or time is concentrated. Recording operations alone does not establish the correctness of figures or outgoing content. The records need to be interpreted alongside the results of human checks and corrections.

A first trial need not include document production, sharing, and calendar changes together. Limiting the meeting and materials, then comparing drafts whose basis can be checked, makes it easier to identify where work became easier. The effect on review effort can guide the next part of the task to delegate.

One assignment should advance the user's next decision

A fixed process may be sufficient when the job is always to produce the same single table. The broader ambition visible in the Cowork examples is to handle work across applications as document locations and meeting context change. That leaves room to reduce the connecting work a person currently performs, while the jobs for which this helps still require individual comparison.

Successful delegation of meeting preparation should be judged through more than file counts or autonomous runtime. A person needs to understand which information supports the explanation, which artifacts a revision affects, and what was applied for other people. If the task becomes easier with review and correction included, Cowork can help the user move to the next decision.