We build Office.js Word add-ins that put Anthropic Claude and OpenAI inside the document — tracked changes, citations, and Microsoft Graph integration, shipped to AppSource.
We have shipped Office.js add-ins to production. Every failed pilot we have seen comes back to one of these.
Lawyers and reviewers alt-tab from Word into a separate AI tool, paste a clause, then paste the answer back as a comment
Naive Office.js add-ins call OpenAI directly from the client and wait 6-12 seconds per redline
Manifest schema, permissions, and AppSource validation reject every other submission
Add-in cannot reach SharePoint, NetDocuments, or iManage because Azure AD scopes were never wired
Tracked changes inserted at stale ranges, comments attached to the wrong paragraph after a re-flow
Lawyers and reviewers alt-tab from Word into a separate AI tool, paste a clause, then paste the answer back as a comment
Context lost between hops, tracked changes broken, no audit trail of what the AI actually saw
15-25 minutes wasted per document on tool-switching alone
How we avoid it: Office.js batched proxy + edge-worker LLM routing + Azure AD SSO wired from day one.
The full Office.js + LLM + Microsoft 365 stack, delivered as one production add-in.
Anthropic Claude or OpenAI returns a structured diff. Our Office.js layer applies it as native Word tracked changes — accept/reject preserved, author attributed to the AI.
Your precedent clauses, playbook positions, and prior matters are indexed in Pinecone or Qdrant. Every suggestion is grounded against your library with citations rendered in the sidebar.
Cloudflare Workers and Vercel Edge Functions sit between the task pane and the LLM provider. Streaming, retries, rate-limiting, prompt-injection filtering, and per-tenant logging all happen at the edge.
Office.js single sign-on hands us an Azure AD token. We exchange it on the server for Microsoft Graph access to SharePoint, OneDrive, and the user's matter folder — zero re-auth prompts.
First-class connectors to NetDocuments and iManage for legal DMS, Salesforce Document Generation for sales contracts, and DocuSign for downstream e-signature — all reachable from the same task pane.
We ship the Office Add-In manifest XML, security questionnaire, and AppSource validation package. Enterprise distribution through the Microsoft 365 admin center or the public AppSource listing.
Six stages, sub-2-second round-trip. Each stage is engineered against a specific latency or correctness failure mode.
The user highlights a clause or clicks 'Review document'. The task pane (React + TypeScript) reads the selection and document context through Office.js.
We pull the active selection, surrounding paragraphs, document-level metadata, and any existing tracked changes. The Office.js batched proxy minimizes Word-process round-trips.
A Cloudflare Worker or Vercel Edge Function authenticates the Azure AD token, enriches the prompt with the user's clause library from Pinecone or Qdrant, and proxies the LLM call.
Anthropic Claude or OpenAI returns a JSON diff: spans to delete, spans to insert, comments to attach, and citations to the source clause in the library.
The task pane applies the diff as native Word tracked changes, attributed to the AI author. Range bindings keep insertions stable even if the user is still typing.
Each suggestion is mirrored in the comment pane with a citation back to the playbook clause, the LLM reasoning summary, and an accept/reject control wired to Office.js.
Office.js (TypeScript) inside Word, Office Add-In manifest XML for distribution, Cloudflare Workers or Vercel Edge Functions for the low-latency LLM proxy, Microsoft Graph API for SharePoint and OneDrive, Azure AD for enterprise SSO, Anthropic Claude and OpenAI for the language model, and Pinecone or Qdrant for the clause-library RAG index. Every box on the architecture diagram has a name.
Anywhere Microsoft Word is the contract surface, the proposal surface, or the memorandum surface.
In-document NDA and MSA review for a Fortune 500 legal firm. Lawyers see redlines appear in Word with citations back to the firm's playbook, no copy-paste to ChatGPT.
Procurement teams at a global manufacturer triage inbound NDAs and master service agreements directly in Word. Risky clauses are flagged before they hit Legal.
Sales engineers draft customer proposals in Word with the add-in pulling approved language from SharePoint and Salesforce Document Generation. Brand and legal language stay consistent.
Investment teams review information memoranda and term sheets in Word. The add-in extracts deal terms, compares to comparable deals indexed in Qdrant, and surfaces inconsistencies.
Accounting and consulting firms automate engagement letter drafting. The add-in personalizes templates from SharePoint, applies firm-standard scope language, and hands off to DocuSign.
These are the engineering targets the pipeline is designed to hit.
Prototype installed via sideloading in week 3. Production add-in published through Microsoft AppSource and Microsoft 365 admin center by week 8.
Office.js task pane shipped behind a sideloaded manifest XML. Wired to Anthropic Claude or OpenAI through a Cloudflare Worker. Tracked changes flowing inside Word desktop and Word on the web.
Pinecone or Qdrant index built from your playbook. Azure AD app registration, OAuth 2.0, Microsoft Graph scopes for SharePoint and OneDrive. NetDocuments or iManage connector if needed.
Manifest XML hardened against AppSource validators. Security questionnaire, privacy policy, and Microsoft Partner Center submission completed. Centralized deployment configured in the Microsoft 365 admin center.
Same manifest extended to Excel and Outlook task panes. AI layer follows the user across Microsoft 365. Optional: Teams message extension reusing the same edge worker.
The questions we get from CTOs, GCs, and platform leads before they sign.
Sideloaded prototype in week 3. AppSource-published production in week 8. Office.js, Microsoft Graph, Azure AD, and your LLM of choice — wired by engineers who have shipped this stack before.
Founded 2019 — 50+ projects delivered, 96% client satisfaction, 4 production AI products.