Connecting AI to your email, your ERP and your documents

What it looks like when it is done right

Group NDL bridges Canadian manufacturers and factories in Asia. Its history is 20 years of email in French and English. Today, AI has read and sorted close to 240,000 of those emails, and it finds 352 part numbers with their exchanges and prices. Every morning, the inbox is sorted and replies to clients are already drafted.

For the team, a question like “what did we promise this client on this part?” takes a few seconds, with the emails that prove it. The full case study tells how.

Your company’s three memories

In almost every business we visit, what the company knows lives in three places: email, the ERP or management software, and the documents that circulate. Email is often the richest source, because that is where the real conversations with clients are.

An AI project that really changes something connects to at least two of those memories. An assistant that only sees what you paste into its window helps you write. It does not yet help you work.

Before AI, put your data in the right place

This is the step nobody sees, and sometimes the longest one. At Group NDL, email had lived for nearly 20 years on an old server at a web host, and AI tools do not connect to that kind of server cleanly. We first moved it to Microsoft 365, without a minute of downtime for the team.

What we learned doing these moves for more than one client:

• Inventory before copying. Mailbox by mailbox, folder by folder. We almost always find working folders in the wrong place, sometimes even in the trash.
• Look for forgotten rules. An automatic copy set up ten years ago does not follow the move, and nobody notices until an important email goes missing.
• Copy first, switch later. Until the switch is made, the old server keeps receiving. Nothing is irreversible.
• Use the move for security. Two-factor sign-in and phishing protection, which old servers did not offer.

If your email is already on Microsoft 365 or Google Workspace, this step is done.

How AI connects, without the jargon

For a long time, every connection required hand-written code. Since late 2024, there is a standard plug, the Model Context Protocol (MCP). Anthropic created it, and it is now run by a Linux Foundation project with OpenAI, Microsoft and GitHub at the table.

What matters for you is that we decide exactly what AI is allowed to do: search an email, read an order, prepare a draft. Nothing more, and it can all be cut off at once.

The ERP connects through its API, the door the vendor provides for other software. A tip from a current project: ask what the API really allows before signing anything. At one manufacturer, it could read production but not reschedule it directly. Better to know at the diagnostic than halfway through.

Read, prepare, act: in that order

First reading: AI searches, summarizes and finds. Then drafts: it prepares the reply, with the right signature and attachments, and a person sends it. Direct actions, like creating an order or changing a date, only come once the first two levels have proven themselves.

A draft to fix costs thirty seconds. An email sent on its own to the wrong client can cost a relationship. That is also how our service connecting Claude to Outlook works: Claude prepares, it never sends.

Security, in practice

• An account reserved for AI, with no admin rights. We know what it sees, and we can cut it off without touching employee accounts.
• What does not belong stays out. Management mailboxes, payroll and HR are excluded from the start.
• A Law 25 assessment before connecting. It is required when personal information leaves Quebec, which happens with most AI services. Our Law 25 guide walks through it.
• The team is told. A simple notice: what AI reads, what it does and what it does not do. People get on board better when they do not find out by accident.

What it changes for your team

Before, you search Outlook with three keywords, open twelve emails and ask a colleague whether they remember last year’s price. After, you ask the question and the answer comes with its proof.

And the day an experienced employee retires, what they knew stays in the company.

Where to start

Pick one memory and one question that comes back every week, often “find everything about this client or this part”. We connect read-only, measure search time before and after, then open up drafts. That is our four-step approach, applied to your data. And if your data is not ready yet, we start there: here is why.