Comparing Word and PDF documents using Copilot Chat: Categorizing changes down to 'unverifiable' status and tracing back to the source text

Microsoft 365・Azureカテゴリを表すパンダのイラスト Microsoft 365 / Azure

About this article
This article is produced using an automated generation workflow powered by generative AI. Based on official Microsoft documentation, it outlines the capabilities, human oversight boundaries, and contractual terms when comparing differences between Word and PDF files uploaded to Copilot Chat.

Verification Status: 📘 Confirmed via official Microsoft documentation, unverified in a live tenant
Information Verification Baseline Date: October 5, 2026

When comparing old and new versions of regulations, proposals, or procedure manuals, you can generate candidate differences simply by providing two files to Copilot Chat and asking it to show the differences.

However, for business use, it is safer to:Categorize items into additions, deletions, modifications, no difference, and unverifiable, and structure them so you can trace back to the source evidence.This approach is recommended.

Is this possible under the current agreement?

According to official Microsoft information as of October 5, 2026, Copilot Chat is included in organizations with applicable Microsoft 365 licenses. File uploads allow users to provide Word, Excel, PDF, PowerPoint, and other files to the chat for content analysis.

ObjectivePrimary conditions
Use Copilot ChatApplicable Microsoft 365 / Office 365 subscription + Entra ID
Upload Word/PDF files to the chatCopilot Chat file upload feature
Web-based Copilot ChatScope included at no additional cost within the eligible subscription
Work-based chat leveraging organizational data broadlyCheck the Microsoft Copilot license prerequisites
Advanced Copilot features within the Word applicationCheck for Microsoft Copilot add-on licenses and related requirements
Agents using shared tenant data such as SharePointCheck the Copilot license or pay-as-you-go conditions

In other words,Uploading and comparing two Word/PDF documents is not subject to the same licensing terms as cross-searching internal information across all of Microsoft 365.

Start with dummy documents

Instead of using actual contracts or HR documents from the beginning, determine the output format using dummy old and new materials.

For example, make the request as follows.

2つの資料を比較してください。

比較項目:

- 対象者

- 申請方法

- 期限

- 承認者

- 金額

- 禁止事項

- 例外

各項目を
「追加」
「削除」
「変更」
「差なし」
「確認不能」
に分類してください。

資料にない内容は推測しないでください。

This makes it clearer what to compare and which labels to return compared to just saying "Tell me the differences."

Why keep "Unable to Verify"

Designing the system to force the AI to fill in blanks may cause it to invent natural-sounding sentences for information not present in the source material.

Therefore,

確認できないものは「確認不能」としてください。
推測で補完しないでください。

set it as follows.

Instead of treating Copilot's response as the final version,use it as a candidate list for human verification of the original text.Require the generation of sources

flowchart LR
    A["旧版 Word/PDF"] --> C["Copilot Chat"]
    B["新版 Word/PDF"] --> C
    C --> D["追加・削除・変更・差なし・確認不能"]
    D --> E["人が原文と照合"]
    E --> F["確定した差分"]

Require the source to be output

Checking diffs alone takes too much time.

Next,

「変更」と判定した項目について、
旧版と新版それぞれの根拠になった記述を示してください。

根拠を特定できない項目は
「根拠確認不可」
としてください。

set as.

The key is not to trust Copilot's explanations blindly, butto make it easy to refer back to the original text.

Move numbers into a separate table

Extracting dates, amounts, quantities, percentages, and version numbers separately makes verification easier.

2資料に含まれる
日付・期限・金額・数量・割合・バージョンを抽出し、
旧版と新版を横並びにしてください。

片方にしかない値はその旨を明記し、
推測で対応付けしないでください。

Since numerical discrepancies have a high impact if incorrect, separate them from the main text summary.

Divide into two stages

Instead of cramming the comparison task into one giant prompt,

  1. Check the document structure

  2. Extract comparison items

  3. Add difference labels

  4. Provide evidence

  5. Re-verify numbers only

  6. Human cross-checks with the original text

Splitting it this way makes verification easier.

Where are the files stored?

According to Microsoft's FAQ, files uploaded to Copilot Chat are stored in the user's OneDrive for Business and can be deleted by the user.

It is also explained that under Enterprise Data Protection, uploaded files are not used for model training.

This does not mean that you can upload anything.

Internal

  • confidentiality classification

  • personal information

  • customer information

  • contract information

  • DLP

  • retention rules

  • AI usage guidelines

must still be followed.

Do not confuse Copilot Chat with Copilot in Word

Differentiate between the method of uploading Word/PDF files to Copilot Chat accessed via browser or Microsoft 365 apps and the Copilot feature within the Word app.

According to official Microsoft licensing documentation, the web-based chat of Copilot Chat is included in eligible Microsoft 365 subscriptions. On the other hand, work-based chat and integrated Copilot features within Word, Excel, Outlook, and Teams are subject to Microsoft Copilot license conditions.

Do not assume that E3 covers everything or that Copilot Chat implies Copilot in Word is the same.

What is measured in the pilot?

When testing in an enterprise, you cannot compare based solely on whether it was convenient.

For example, in comparison work between the old and new versions,

metricsmeasurement method
Initial diff candidate generation timeManual vs. Copilot usage
Number of oversightsDifferences discovered during final human review
Number of false positivesNumber of instances where Copilot flagged a change that was actually incorrect
Grounding verification timeTime to return to the original text
Correction timeTime spent by humans correcting Copilot output

shows that effectiveness should be measured not only by AI response timebut also by the total time including human corrections.measure the impact.

Do not delegate legal judgments

Even when used for contract comparison,

この契約は法的に問題ないか

do not leave final decisions like this solely to Copilot.

Copilot is

  • diff candidates

  • rationale candidates

  • numeric extraction

  • Identify overlooked verification items

used up to that point, with final decisions made using the original text and by the responsible person or expert.

Items to be verified by the administrator

For an enterprise pilot, verify at least the following before starting use.

  • Microsoft 365 license for target users

  • Sign-in with Entra ID

  • Availability of Copilot Chat

  • Web search settings

  • Upload and storage in OneDrive for Business

  • DLP, sensitivity labels, and retention policies

  • File upload restrictions

  • Functional differences compared to users with additional Microsoft Copilot licenses

  • Presence or absence of pay-as-you-go billing when using agents or SharePoint

Because features are rolled out gradually, the user interface or behavior may differ by tenant or user even under the same subscription.

Official documentation

Document information

Article title
Comparing Word and PDF documents using Copilot Chat: Categorizing changes down to 'unverifiable' status and tracing back to the source text
Published
Updated
Source
https://papanda925.com/?p=18042&lang=en

License: Text and original figures for which this site holds the relevant rights are available under CC BY 4.0 , unless otherwise noted. This article may include content created or edited with generative AI. If code has a separate license notice or a linked GitHub repository license, that license takes precedence for the code. Quotations, third-party materials, images, and trademarks are excluded from this license. Usage policy

Copied title and URL