# Connecting Microsoft 365 Copilot, Teams and Claude to the Same Governed Company Numbers

> Why raw connectors leave AI tools disagreeing on basic business figures, and how one governed knowledge graph underneath Copilot, Teams and Claude fixes it.

Published: 2026-11-22T21:00:00Z
Updated: 2026-11-22T21:00:00Z
Author: Archie Norman (Founder, Mercury Labs)
Category: Twin
Tags: copilot, microsoft-teams, claude, mcp, governed-knowledge-graph, entra, enterprise-ai
Canonical URL: https://mlx.systems/blog/connecting-copilot-teams-and-claude-to-governed-numbers

## TL;DR

- Microsoft 365 Copilot reads email, Teams and SharePoint well, but cannot see reconciled figures across your finance ledger, ERP, CRM and project tracker.
- Connecting AI tools straight to raw SaaS APIs or SQL tables leaves the language model guessing entity joins and doing arithmetic on the fly, so Copilot and Claude often return different numbers.
- A Company Twin runs a deterministic, model-free Rust graph engine underneath Copilot (custom federated MCP connector with OAuth 2.1 + PKCE), Teams (Entra SSO agent), Slack and Claude so every tool returns the same figure with source links and role permissions.

---

Most UK mid-market organisations that rolled out Microsoft 365 Copilot over the last twelve months hit the same ceiling within six weeks.

Copilot summarises Teams meetings well. It drafts emails and finds SharePoint documents quickly. Then an Account Director or CFO asks an ordinary commercial question:

> *"What does Northgate Retail owe us right now, and which delivery projects are still open?"*

Copilot searches emails and spreadsheets, pulls an outdated attachment from three months ago, or says it cannot see the accounting ledger. Meanwhile, the finance lead opens Claude, exports two CSVs from Xero and HubSpot, and gets a different total because one tool counted an on-hold job and the other blended GBP and USD invoices.

Within a quarter, two things happen: IT starts hearing *"Copilot isn't worth the renewal,"* and leadership quietly stops trusting AI answers on numbers.

## Why raw connectors do not solve the problem

The intuitive first fix is to wire Copilot or Claude straight into raw SaaS APIs, Power BI tables, or database MCP servers. In practice, raw connectors leave four hard problems sitting inside the language model on every single prompt:

1. **Cross-system entity matching is guessed on the fly.** Your accounting ledger says `Northgate Retail Ltd`, your CRM says `NORTHGATE RTL`, and your project tracker says `Northgate (UK)`. A model guessing joins at query time misses rows silently.
2. **Arithmetic happens inside the model.** Pulling raw invoice rows into a context window and asking an LLM to sum overdue balances, filter out credit notes, and respect complete-period accounting rules produces drift from one run to the next.
3. **Legacy systems and spreadsheets stay invisible.** Mid-market IT estates rarely live in three tidy cloud apps. Older ERPs (such as Infor SunSystems or Dynamics GP), on-premise SQL databases, and operational spreadsheets have no clean cloud connector.
4. **Permissions fracture across tools.** Giving Copilot one connector, Claude a pasted API key, and Teams a separate bot means payroll or margin data can leak in one surface even when restricted in another.

In our audited internal test on 77 everyday business questions about our own company (3 runs each), even giving a frontier model direct read-only SQL access to our raw synced tables only matched our agreed business definitions roughly **60% of the time** — compared with **86%** through our governed knowledge graph over MCP, finishing about a quarter faster in roughly 40% fewer steps ([read the full benchmark write-up on Mercury Labs](https://mercurylabs.io/blog/raw-sql-vs-governed-knowledge-graph)).

## One governed knowledge graph underneath every AI tool

A **Company Twin** sits between the business systems you already run and the AI tools your team already opens every morning — without migrating a single source record.

Instead of asking a language model to stitch tables together on the fly, the Twin does the hard engineering work ahead of time:

- **Read-only system connections:** Finance ledgers, legacy ERPs, CRM, project trackers, SQL databases, SharePoint and operational spreadsheets sync read-only into one reconciled entity graph.
- **Model-free Rust graph engine:** Agreed business definitions (overdue receivables, complete-month invoiced revenue, unbilled WIP, active pipeline) are calculated deterministically in memory — zero LLM arithmetic, sub-second response times, and explicit refusal when a question falls outside coverage.
- **Click-through provenance:** Every answer carries the sync timestamp, definition version, and direct links back to the underlying source records so nobody has to take an AI summary on trust.
- **One role-based permission model everywhere:** Sales invoices can be open to account managers while payroll and management accounts remain restricted to finance. The graph engine enforces that boundary identically whether a user asks in Copilot, Teams, Slack, or Claude.

## How it runs across Copilot, Teams and Claude together

In real UK organisations, teams rarely standardise on a single AI client:

- **Microsoft 365 Copilot (`/copilot`):** Connects via Microsoft's **custom federated connector (MCP)** to your Twin's strictly read-only `/mcp/graph` endpoint using OAuth 2.1 with PKCE. Works inside Copilot Chat, Word, Excel, PowerPoint and Outlook.
- **Microsoft Teams (`/teams`):** Runs as a native Teams agent with **Microsoft Entra sign-in** inside channel threads and 1:1 chats — ideal for giving operating staff governed answers without buying a Copilot licence for every seat.
- **Claude & Claude Code (`/claude`):** Leadership, finance analysts, and private equity operating partners connect Claude to the exact same `/mcp/graph` endpoint over OAuth 2.1 + PKCE with zero pasted API keys.

Graph reads from Copilot, Teams, Slack and Claude consume **zero AI credits** on every Company Twin plan. You use the AI licences you already pay for; the Twin supplies the reconciled company truth underneath them.

## What a two-week staged rollout looks like

You do not need a six-month data warehouse programme to prove this works:

1. **Days 1–5 — Connect 2 to 3 core systems read-only:** We wire your accounting/ERP system, CRM and project tracker, reconcile client and project entities, and check every core figure against your existing month-end reports.
2. **Days 6–8 — Register the federated connector & Entra roles:** Your Microsoft 365 admin registers the OAuth 2.1 + PKCE connector in the Teams Developer Portal and M365 admin centre (or enables the Teams package) and maps Entra groups to Company Twin roles.
3. **Days 9–10 — Staged rollout to 5 pilot users:** Your pilot group asks real commercial and finance questions inside Copilot, Teams and Claude, clicks through to verify source records, and expands once the numbers match your books.

## See it on your own question

Bring one cross-system question your team currently chases across spreadsheets every week. We will show the same question answered live in Microsoft 365 Copilot and Claude — with source links and Entra role enforcement — in a 20-minute walkthrough.

- Explore [Company Twin for Microsoft 365 Copilot](/copilot), [Microsoft Teams](/teams) and [Claude](/claude)
- Read the [Teams & Copilot federated connector setup guide](https://docs.mlx.systems/docs/connect/teams-and-copilot#connect-as-a-custom-federated-connector-mcp)
- View [transparent subscription pricing](/pricing) or [book a 20-minute Copilot demo](/copilot#waitlist)
