AI tools like ChatGPT and Claude cannot natively produce Word documents with tracked changes. They can read your document, suggest edits in plain text, and even rewrite sections — but they output text, not a .docx file with accept/reject markup embedded in the Word XML. Closing that gap requires a tool that sits between the AI and the Word format, applies the AI's proposed changes as structured tracked-change data, and returns a file Word can actually open in Review mode. That is what MCP connectors like the Scaffold MCP connector are built to do.
This guide explains the full picture: how Word tracked changes work technically, why AI chat interfaces fall short, how MCP bridges the gap, and how professionals across legal, HR, and consulting can build reliable AI-assisted review workflows.
What Are Word Tracked Changes, and Why Do They Matter?
Word tracked changes are not visual annotations painted on top of text. They are structured data embedded inside the document's XML — the Open XML format that every .docx file uses. Each tracked change is a <w:ins> (insertion) or <w:del> (deletion) element that wraps the affected text, carries metadata about the author and timestamp, and is hidden or shown depending on the document's display settings.
When a reviewer opens a file with tracked changes in Word, the Review tab surfaces all of this: who changed what, when, and in what order. A second reviewer can accept or reject each change individually, leave comments, or layer their own redline on top. This sequential, attributable markup is the standard workflow for professional document collaboration — in contract negotiations, employment agreements, policy approvals, and consulting deliverables. The reason it matters is accountability: every change is on the record, and every decision to accept or reject it is deliberate.
Why Can't ChatGPT or Claude Produce Tracked Changes on Their Own?
AI language models generate text tokens in sequence. When you ask ChatGPT to "redline this contract," it can read the text you paste, reason about it, and return revised text — but that output is a stream of characters, not a structured document file. To produce a .docx with tracked changes, a process needs to:
- Parse the original document's XML structure
- Identify exactly which text elements change
- Wrap deletions in
<w:del>elements and insertions in<w:ins>elements, with correct author, date, and revision ID attributes - Reconstruct the full document XML with all unchanged content preserved, including styles, tables, headers, footers, and embedded objects
- Repack the result into a valid .zip-based .docx archive
This is a software engineering task, not a text-generation task. The AI can determine what should change. Separate software has to determine how those changes get encoded in the Word format. Without that second layer, the AI's suggestions exist only as chat output — and the human has to manually apply each one in Word.
How MCP Bridges AI and Word's Tracked-Change Format
The Model Context Protocol (MCP) is an open standard that lets AI models call external tools during a conversation. An MCP server exposes capabilities — uploading files, running computations, reading databases — as callable functions. The AI decides when and how to call them based on the user's request.
The Scaffold MCP connector is an MCP server built specifically for Word document operations. When you ask an AI with Scaffold connected to "redline clause 4 to add a mutual NDA obligation," the sequence is:
- The AI reads the document via Scaffold's
read_documenttool - The AI identifies the specific changes — new text to insert, old text to remove, language to restructure
- The AI calls Scaffold's
redline_documenttool, passing the change instructions - Scaffold's server-side code executes the XML operations, writes the tracked-change markup, and produces a revised .docx
- The AI returns a download link; the user opens the file in Word with full tracked-change markup intact
The AI contributes legal or domain reasoning. The Scaffold MCP connector contributes the document engineering. Neither could achieve the full result alone.
How This Works Across Different Professions
Legal: Contract Negotiation and Clause Drafting
Attorneys typically receive opposing counsel's draft and need to mark it up with their client's positions. Historically this means reading the contract, making manual changes in Word, and sending the redlined version back. With the Scaffold MCP connector in Claude or ChatGPT, an attorney can upload the opposing draft, describe the client's requirements in plain English, and get back a tracked-changes file ready for review — in minutes rather than hours for routine provisions.
The attorney still reviews every change. The AI accelerates the drafting of positions the attorney would have written anyway; it does not replace professional judgment. For non-standard provisions or high-stakes clauses, the attorney drafts those manually and uses AI assistance only for the boilerplate.
HR: Policy Updates and Employment Agreement Revisions
HR teams regularly update policy handbooks, offer letter templates, and employment agreements to reflect regulatory changes, new benefits, or revised company positions. These updates touch many documents, and tracking what changed from version to version is important both for compliance and for employee communications.
With an AI-connected workflow using the Scaffold MCP connector, an HR specialist can describe the policy change in plain language — "update the remote work policy to reflect the new three-day minimum in-office requirement" — and receive a tracked-changes version of the affected document. Legal can review the markup before final approval. The change history is preserved in the file rather than in someone's email thread.
Consulting: Deliverable Review and Client Markup
Consultants frequently deliver documents — reports, proposals, engagement letters, SOW agreements — and receive client feedback that needs to be incorporated with a clear audit trail. When a client's legal or procurement team redlines a consulting SOW, the consultant needs to review and respond with their own markup.
The Scaffold MCP connector supports this workflow in both directions: consultants can use it to apply client-requested changes and track them, or to generate initial draft markup based on their standard positions before a negotiation round.
What the AI Gets Right, and Where You Still Need Professional Judgment
AI models are good at applying rules consistently across a long document. They can find every instance of a specific clause pattern, apply the same change uniformly, and flag language that conflicts with stated requirements. They are less reliable for:
- Novel legal arguments — AI can suggest standard alternatives, but creative or jurisdiction-specific positions require legal expertise
- Business context — the AI does not know your client relationship, deal dynamics, or what you already agreed to verbally
- Ambiguous instructions — if you ask for "more balanced" indemnification language, the AI will pick an interpretation; you need to verify it chose the right one
- Formatting-sensitive documents — heavily formatted files with tables, signatures blocks, or embedded objects can behave unexpectedly after programmatic edits
The practical rule: use the AI to accelerate the mechanical work of drafting and marking up standard provisions. Reserve your own review time for the judgment calls.
Best Practices for Reviewing AI-Proposed Tracked Changes
Review every change before accepting. Never click "Accept All" on an AI-generated redline without reading through the markup. The AI may have made a change you did not intend, or interpreted an ambiguous instruction in a way that changes meaning.
Use Word's Reviewing Pane. The Reviewing Pane (View > Reviewing Pane) shows a sequential list of all changes. This is faster for long documents than scrolling through inline markup.
Compare against your original. Word's Compare feature (Review > Compare) lets you diff any two versions. Run it between your original and the AI-redlined version to confirm no unintended changes appeared outside the sections you targeted.
Document your instructions. Save the prompt you gave the AI alongside the document version. If a change is questioned later, you want a record of what you asked for.
Set a reviewer name in Scaffold. By default, tracked changes are attributed to "Scaffold." In your Scaffold account settings, you can set the reviewer name to your own name or your firm's name, so the attribution in Word is accurate.
Frequently Asked Questions
Does the Scaffold MCP connector work with both Claude and ChatGPT?
Yes. The Scaffold MCP connector is compatible with any AI that supports the Model Context Protocol. That currently includes Claude (claude.ai web and Claude Desktop), ChatGPT Plus and Team via the ChatGPT connector system, and any other MCP-compatible AI client. The workflow is the same regardless of which AI you use.
Can I use this for documents with complex formatting — tables, signatures, headers?
The Scaffold MCP connector preserves document structure during redline operations. Tables, headers, footers, and text boxes are maintained. Very complex layouts with embedded images or custom macros may require manual review after processing, but standard legal and business documents work reliably.
If you want to see how this works in practice, start a free 7-day Scaffold trial. You can have a full AI-assisted redline workflow running in Claude or ChatGPT within a few minutes of setup.