About this article
This article was created using an automated generation workflow powered by generative AI. Official Microsoft documentation was verified on October 20, 2026, and prompts and verification procedures were created to be tested with fictional data. Execution on the Copilot screen has not been performed.Verification Status: 📘 Official Information Verified – Prompt Execution Unverified
- Supported Environments, Capabilities, and Limitations (Official Verification: October 20, 2026)
- Extracting decisions and action items from brief notes
- Compare the expected classification with the original text
- Retaining reference lines reduces the unit size for verification.
- Change a single condition and observe assignee inference.
- Expanding to longer memos containing ambiguous statements like "I'll take care of it".
- Do not let completion criteria fall outside the task
- Reduce input and retain information necessary for confirmation
- Official Information and Primary Sources
Supported Environments, Capabilities, and Limitations (Official Verification: October 20, 2026)
| Item | Scope and Target of this Example |
|---|---|
| Product / Entry Point | Microsoft Copilot Chat for work (formerly Microsoft 365 Copilot Chat).Assumes standard chat on the web versionAssumes standard chat on the web version |
| Account / License | Work account authorized for use within the organization. Confirm availability with your contract and administrator settings. Personal Copilot is out of scope for this example. |
| Additional Contracts / Billing | Additional features within Office apps and usage-based agents are not used in this example. Do not interpret this as enabling features that require additional contracts. |
| Input / Storage Location | Paste fictional text directly. Do not connect local paths, OneDrive, SharePoint, or the inbox. |
| Actions to Test | Referencing, organizing, comparing, and drafting inputted text. Content must be cross-checked against the original text by a human. |
| Out of scope for this example | File search, direct document editing, moving, or deletion, email transmission, and registration in business systems |
| Limits and availability status | Preview features are not assumed. Numerical values for input length, frequency, region, and screen differences, as well as actual device builds, are unconfirmed. Split long inputs and verify availability on the screen. |
| Information management | Use only dummy data initially. Before expanding to business information, check organizational approval and information classification. Do not rely solely on prompt instructions as a guarantee for access control or the suspension of web searches. |
When meeting notes contain a mix of "to be considered," "would like to request," and "decided," neatly formatted summaries can be dangerous. This is because suggestions can inadvertently become finalized decisions, and tasks may be assigned to individuals who were merely mentioned. Restricting the work assigned to Copilot Chat to extracting task candidates that allow you to trace back to the original text, rather than completing the meeting minutes, helps reduce the effort required for verification.
Extracting decisions and action items from brief notes
The following is a fictional inventory meeting for equipment. It does not contain information on real departments or individuals. Paste the instructions and notes together into a new Copilot Chat conversation. Unlike features that use Teams meeting recordings or transcriptions, this process handles only the text you input.
目的：次回会議で確認できるタスク候補表を作る。 根拠は下のメモだけ。外部検索や別の文書で補わない。 「決定」「提案」「未確定」を区別する。 担当者・期限の記載がなければ「未定」と書く。 列はID／分類／作業・論点／担当／期限／根拠行。 一つの行に複数の担当や作業を詰めない。 表の後に、確認が必要な質問を3件以内で示す。 L1：棚卸し用の一覧表はA係が作成し、10月28日までに共有することで決定。 L2：現物確認は11月上旬を候補とする。実施日はまだ決めない。 L3：B係から、破損品の写真も残してはどうかという提案があった。 L4：写真の保管場所と担当は次回決める。 L5：C係は、予算が確認できるまで追加購入を確約できないと説明した。
The only task finalized in this example is the creation of the spreadsheet by Section A. The implementation date for physical inspection, the storage location for photos, and the person in charge of photography have not been finalized. The statement by Section C does not constitute taking responsibility for purchasing. During the initial review, check whether this boundary is maintained rather than focusing on the appearance of the table.
Compare the expected classification with the original text
The following table is an expected example for verifying correct answers and does not represent Copilot's execution output. Phrasing and line breaks may vary even with the same instructions. Evaluate based on whether unauthorized commitments have been added, rather than on exact text matching.
| Target | Expected handling | Example to return as an error |
|---|---|---|
| Spreadsheet creation | Decided, Section A, October 28, L1 | Rounding the deadline to the end of October |
| Physical inspection | Candidate schedule, implementation date undecided, L2 | Confirm November 1st |
| Photo storage | Proposed or unconfirmed, L3/L4 | Assign to Section B |
| Additional purchase | Conditional and uncommitted, L5 | Assume purchase request has been submitted to Section C |
Preventing the AI from inferring assignees based solely on proximity in names is central to this example. If "Proposal from Section B" changes to "Execution by Section B", the operational meaning differs entirely even if the table looks well-structured. Including the reference line allows you to instantly return to the discussion and make corrections.
"Early November" is a candidate timeframe rather than a specific deadline. Prioritizing the completion of date fields may cause the AI to insert arbitrary dates. The design treats blanks not as failures, but treats "TBD" as useful information.
Retaining reference lines reduces the unit size for verification.
Comparing full memos and summaries becomes increasingly burdensome as their length increases. Including the original line number in each row narrows down where to check—such as verifying the validity of the assignee in L1 and the handling of photos in L3 and L4. Line numbers serve as markers specific to this input rather than links to the document.
The tables created here are not official operational instructions. They move to the execution list only after meeting participants review them and a responsible person approves them as finalized decisions. Distributing AI extraction results directly to assignees risks having unverified assumptions treated as instructions.
Change a single condition and observe assignee inference.
Next, in a new session, replace only L3 with: "Decided that Section B will take photos of the damaged items and share them by October 30." Leave all other rows and instructions unchanged. The expected change is that photography becomes a decided task, complete with an assignee and deadline.
Meanwhile, the photo storage location remains as L4. Just because the person taking the photos has been decided does not mean the storage location has been determined as well. Observing the extent of the impact from a single-line change makes it easier to catch errors where related matters are prematurely grouped and finalized.
Separating conversations helps with comparison. Passing revised memos consecutively in the same session can cause previous and subsequent notes to merge. Testing revisions and previous versions in separate sessions and viewing them on the same evaluation table makes it easier to distinguish between the effects of instruction improvements and chat history bias.
Expanding to longer memos containing ambiguous statements like "I'll take care of it".
In practical discussions, statements often lack explicit subjects. If it is unclear who spoke immediately before "I'll take care of it," the assignee cannot be confirmed. Input data should include speaker identifiers, and any sections where the speaker cannot be restored should be flagged for questions. It is safer to prompt for confirmation than to let common sense fill in blanks.
Tasks and decisions should also be separated. "Reviewing the estimate" is a task, whereas "determining whether to make a purchase" is a decision. Combining the two can result in the person tasked with reviewing the estimate being treated as the purchase approver. Classifying items into table columns is meant to identify the type of work before assigning responsibility.
When dividing memos, pay attention to whether decisions have been modified later. If the first half contains "October 28" and the second half states "Reschedule dates," the first half's table alone does not reflect the latest state. In the final consolidation step, create a correspondence table for retractions, changes, and added conditions.
Do not let completion criteria fall outside the task
Even for a task like creating a list, the state changes depending on whether simply creating it is sufficient, or if it is only complete once someone reviews and shares it. If the original notes lack completion criteria, do not arbitrarily add them to the task description; instead, leave them as a confirmation question. Even if the assignee and deadline are filled in, work can remain ambiguous regarding how it is finished.
Because line numbers in meeting notes shift when the source text is edited, fix the version used for verification. If you append notes after verification is complete, save them as a new version and either recreate the table using that same version or check the diff. Combining an old table with new notes can sometimes point to statements referencing different source lines.
Even when breaking tasks down into smaller steps, make sure not to add operations that are not in the input. While one might propose splitting "create a list" into four phases—design, input, review, and approval—these are not the four tasks agreed upon in the meeting. Process proposals should be separated from extracted decisions and submitted as proposals for confirmation.
Trial records can be reused if the input version, prompts used, output, and reason for correction are saved as a set. The next time the same error occurs, you can compare whether it was the result of changing the prompt or if the input was ambiguous. Rather than adding long prompts every time, it is easier to manage if you gradually refine the conditions handling frequent errors.
Reduce input and retain information necessary for confirmation
Initial trials will be conducted using this dummy data. When scaling to actual meetings, check the account permissions and information handling scope approved by your organization, and remove unnecessary personal names and contract information. Even if names are replaced with symbols, individuals can sometimes be identified from combinations of dates and cases, so substitution alone does not constitute complete anonymization.
Even if Copilot returns a table, it does not mean it has been registered in the task management service. When copying and moving to another table, check the sharing permissions accessible to assignees and the storage location of the original notes. If the sharing destination for the summary is broader than the original notes, it can lead to unintended dissemination of meeting content.
Evaluate trials using the number of corrections rather than the number of extracted rows. Slightly increase the dummy data and record the number of decisions misclassified as proposals, undetermined assignees supplemented, and unsubstantiated deadlines set. If problems are visible with short inputs, you can revise the prompts before increasing real data.
Final verification depends on whether readers of the table can return to the source text. A warning note simply saying "please check" does not indicate where to check. By retaining the three elements of source lines, pending status, and confirmation questions, meeting participants can specify the points that need to be answered.
Organizing meeting notes is not just about shortening text. It is the work of locking down what has been decided while keeping what has not been decided visible. If this distinction is maintained, the next meeting will not require reading aloud every row of the table, allowing focus on undetermined assignees, deadlines, and conditions.
Official Information and Primary Sources
Refer to official documentation for product scope and input methods. Dummy data, prompts, expected examples, and checklists are verification proposals for this article and do not represent output examples guaranteed by Microsoft.
