AI AGENTS ERP + MICROSOFT 365 Regional distribution and manufacturing group, ~300 employees

An AI layer on top of ERP and Microsoft 365 for a regional distribution and manufacturing group

A regional distribution and manufacturing group with around 300 employees runs its business across an ERP, Microsoft 365, production and quality-control systems, a CRM and a VoIP call centre. We built an AI layer on top of all of them: MCP servers that expose these systems to agents, an orchestrator that routes work between them, and a local LLM layer for data that should not leave the company. Agents work by role, each seeing only the data that role is allowed to see.

Minutes
for a custom report, down from 4 hours to 4 days
15 → 0
analytics requests to developers per month
1–2 weeks
to see whether a hypothesis worked, down from weeks or months

The problem we were asked to solve

The group's day-to-day work runs across many systems: the ERP holds orders, stock, prices, payments and purchasing; production and quality control have their own systems; the CRM and the call centre hold the client side; and Microsoft 365 holds the conversations around all of it — email, Teams, shared files and spreadsheets.

The business reacted to its own metrics too slowly. Product launches and new hypotheses were judged on assumptions rather than on real numbers, because getting the numbers was hard: every custom report needed the development team.

The biggest pain was matching data across systems and domains. Sales, production, purchasing and quality control all needed answers that sat in more than one place.

An AI layer, not another system

Nothing was replaced. The ERP, Microsoft 365, the production, quality-control and CRM systems and the call centre stay the systems of record; the AI layer reads from and acts through them, and joins their data into one picture.

  1. Systems of record
    ERP
    orders, stock, prices, payments, purchasing
    Production & quality control
    MES and QC data
    CRM & call centre
    clients, VoIP calls and their transcripts
    Microsoft 365
    Outlook, Teams, SharePoint, Excel
  2. MCP servers
    ERP MCP server
    ERP data and actions as agent tools
    Production, QC and CRM MCP servers
    operations and client data, call transcripts
    M365 MCP server
    mail, chats and files via Microsoft Graph; contacts and threads stay when a manager leaves
  3. Orchestration
    Agent orchestrator
    routes each request to the right agent, within the user's role
    Model layer
    local Qwen on the client's servers; cloud models get anonymised data only
  4. Agents
    Sales manager
    own client pool and the data around it
    Purchasing, brand, production
    each role with its own data scope
    CEO
    full view: sales, payments, production, stock, cash flow
Architecture of the AI layer. Data and actions flow from the left; agents only reach the systems through the MCP servers.
01

MCP servers over every system

Each system is exposed to agents through its own MCP server: a fixed, reviewable list of tools rather than open access to the database or the mailbox. Agents read data, draft replies, build reports and send the data people need wherever it has to go, once or on a schedule.

02

Agent orchestration

The orchestrator decides which agent handles a request and which tools it may call, within the data the user's role allows. Recurring reports and data deliveries run on a scheduler.

03

Local and cloud models

A local Qwen model runs on the group's own servers. Where cloud models do the work better, the data is anonymised before it leaves the company.

04

Agents by role

Each role gets an agent with its own data scope. A sales manager's agent works only with that manager's pool of clients and the data around them; the CEO's agent has full access to sales, payments, production, stock and cash flow. People use the agents where they already work: in Microsoft Teams, and in Outlook through an add-in.

What changed for the client

Sales got continuity: when a sales manager, buyer or brand manager changes or leaves the company, the client's contacts and the full correspondence stay with the business and are visible to whoever takes over.

Reports and analytics no longer go through the development team, and decisions are made on real numbers.

  • Custom reports — from 4 hours to 4 days, depending on complexity, down to minutes.
  • Analytics requests to the development team — from about 15 a month to none.
  • Hypotheses — from weeks or months to a result, or a clear non-result, within 1–2 weeks of a marketing activity. A/B tests became possible.
Model Context Protocol (MCP) Microsoft 365 / Microsoft Graph API ERP, MES, QC and CRM integration VoIP call transcription Role-based data access Agent orchestration Local LLM: Qwen Cloud models (anonymised data) Microsoft Teams, Outlook add-in Go Client's dedicated servers

Interface screenshots are withheld under client confidentiality.

Running a distribution or trading business on ERP and Microsoft 365?

In 30 minutes we will look at one process you want to take off your team and tell you whether an AI agent can do it — and what a 4-week pilot would look like.

See the 4-week pilot

Book a 30-min call