- 240,000 emails read and sorted by AI
- 352 part numbers found with their exchanges
- 20 years of email moved, zero downtime
- 210 drafts prepared by Claude in October 2026
Where it started
Every file goes through several hands: the client in Quebec, the Canadian team, the project manager in China, the factory, the carrier, customs. Everything happens by email, and many of those emails look alike: acknowledging a request, relaying a technical answer, following up with a quiet client, asking for a delivery date.
Every quote meant redoing the exchange rate by hand, applying the freight markup, picking the right pricing rules for the region, filling in a purchase order for the factory, then writing to a carrier. And every month, invoicing was rebuilt line by line in a spreadsheet.
Everyone worked in their own inbox, with their memory as the only tracking system. On Fridays, the ten or so project managers each sent a summary of their files, and nobody had time to compare it with the previous week’s to see what had disappeared without explanation.
And the company’s history, 20 years of email, sat on an old server that no AI tool could connect to.
Email received
Sorted by role
Draft in Outlook
A person reviews and sends
Project CRM
The challenge
Give the whole team the same memory and the same follow-up, with replies ready to send in each person’s voice, and tool up the full sales cycle without changing its rules: same prices, same templates, same suppliers. All without taking the team out of Outlook or making them learn one more piece of software.
What we built
We started with the history, because nothing could be plugged into what was there. Seventeen domains were brought up to date, and the 20 years of email were moved to Microsoft 365 without a minute of downtime for the team. Three security flaws spotted along the way, outside the mandate, were reported and fixed.
Then AI read and sorted close to 240,000 emails. The lists of clients, parts and prices were rebuilt from that history: 352 part numbers can now be found with their exchanges.
On top of that, we connected Claude, Anthropic’s AI, to the team’s mailboxes. It all runs on Claude Team: a connector built by Avenor links Claude to the Outlook and Microsoft 365 mailboxes, and a custom Claude plugin teaches it the clients, the parts, the project managers and how each person writes. Several times a day, incoming emails are sorted: who they belong to, what needs a reply, what should not go unanswered. The reply to the client is prepared in French, the acknowledgement to the project manager in English. Customs holds, container tracking, air versus sea calls, non-conformities: each recurring case has its own template. Nothing goes out on its own: each draft waits in Outlook for a person to review and send it.
Drafts are written in the voice of the person who will sign them, based on their own emails, and they use the client’s words as is. When a client talks about a pin or a flange hole, the reply talks about a pin or a flange hole.
The sales cycle follows the same logic. The business rules were encoded once: currency conversion off the official Bank of Canada rate, freight markup, pricing rules by delivery region, taxes depending on which entity invoices. The quote comes out laid out, ready to send. The factory purchase order fills itself in the template the supplier expects, and the email goes to the right carrier for the destination. Every month, the reports are read and the invoice comes out with its detailed recap.
The project CRM became the shared memory: every file, who has the ball, the dated history, the files and the deadlines. The team dictates what it learns on the phone, and an assistant powered by Claude answers “what happened with this file?”, with the history, the files and the last price, whichever mailbox the email was in. On Fridays, the project managers’ summaries are compared with the previous week’s. An open line that disappears without explanation is flagged the same day.
Around it, the multi-division website was designed and put online without an agency, with two ad campaigns on the French-speaking and English-speaking markets. And every Monday, a prospecting agent reviews Quebec business news, assesses at least 30 companies and prepares a full profile for the best ones. The team decides: no first contact exists before its yes.
Nothing is sent without a person reading it first. AI prepares, the team decides. No internal rate, margin or supplier price is shown here: the calculation rules live in the system, not in the documents sent to the client.
The email side of this engagement is detailed in the article on automating Outlook email with AI. To plug Claude straight into Outlook and Microsoft 365, see how to connect Claude to Outlook.
How it unfolded
The memory
20 years of email moved with no downtime, then read and sorted by AI.
Sorting and drafts
Each email goes to the right person, with the reply already drafted in Outlook.
The sales cycle
Quote, factory purchase order, carrier email and invoice, from one source of rules.
Follow-up
A shared project CRM, and Friday summaries compared week after week.
Growth
A multi-division site, two campaigns, and an agent that spots companies to approach.
The results
Every morning, the inbox is sorted and replies are already drafted in Outlook, in the right thread, with the signature and attachments. Reviewing them takes a few seconds, where they used to have to be written.
Measured from October 1 to 28, 2026: Claude prepared 210 drafts in three mailboxes. Of the drafts the team decided on, 77% were sent, 54% exactly as Claude wrote them. In the same month, the team asked the CRM assistant 246 questions instead of digging through email.
The biggest gain is not the writing, it is no longer searching. A complex email used to take an hour or more, the time to find the history of a part, a purchase order, a price or a shipment in someone else’s mailbox and in hundreds of project files. We estimate about 13 hours a week given back to the Canadian team, around 640 hours a year. This is an estimate, built on the measured drafts and questions, with assumptions on the minutes saved per search.
A question like “what did we promise this client on this part?” is settled in seconds, with the emails that prove it. What the team knows no longer depends on one person’s memory.
Manual calculation errors are gone from quotes: two quotes prepared two weeks apart use exactly the same rules. The monthly invoice, which used to be rebuilt line by line, now comes out in one pass with its recap.
When a project manager drops a file from their summary without closing it, the team knows that same Friday, not three weeks later.
Handling of requests coming through the site is in development. It aims to bring a request down from about fifteen minutes to two or three: classified, in the right language, with a draft reply already in context.
The email migrations and the email server built for Group NDL now serve other clients, with their own rules.
Replying to a client email
The plan, delivered and ahead
What makes this case different from the others is that it was not built in one go. Each solution came out of a task that recurred too often, and got tooled up when the volume justified it. The result is a coherent ecosystem, assembled piece by piece, in production.
It is also the clearest demonstration of the method: start with the most repetitive task, make it reliable, then move to the next one reusing what already exists.
One memory for the whole team, and the sales cycle tooled up from price request to invoice
- Email sorting and drafts in OutlookEach email goes to the right person, with a reply written in their voice and in the recipient’s language.
- Quotes and price gridsCurrency conversion on the official rate, freight markup, pricing rules by delivery region. Internal rates stay inside the system.
- Purchase orders and carriersThe template the factory expects fills itself, and the request goes to the right carrier for the destination.
- Monthly invoicingReports are read, freight lines excluded by a precise rule, and the invoice comes out with its recap.
- Project CRM and Friday summariesEvery file with its dated history, and project managers’ summaries compared from one week to the next.
- Weekly prospectingAt least 30 companies assessed every Monday, a full profile for the best ones, no contact without the team’s yes.
Close the loop at intake, on requests coming from the site
- Automatic intake of site requestsEvery form submission is picked up at send time, with no hand-watched inbox in the middle.
- Classification by divisionThe request is routed to the right division by its content, rather than sorted on reading.
- Reply in the requester’s languageThe reply goes out in the language of the request, without the question coming up.
- Draft reply in contextA reply already written with the file’s context. All that is left is to approve, adjust and send it.
An ecosystem of solutions around one memory
Each solution handles one specific task and reuses what the others already know. That sharing is what keeps a price or policy change from having to be applied in five different places.
From the price request to the invoice, with no break
The path is tooled end to end. Each step produces the document the next one expects, in the format the recipient already looks for.
The next piece is already mapped: requests coming from the site, classified and pre-drafted before they even land in an inbox. The rest of the path is already waiting for them.
What the company gets out of it
Nothing falls between two inboxes
A client waiting, a project manager running late, a project going quiet: the team sees it before it costs anything.
Everyone replies in their own voice, in the right language
Drafts start from the emails of the person who signs and use the client’s words. French to the client, English to the factory.
One source for every pricing rule
Currency, freight, pricing rules and taxes are computed in one place. Change a rule and it changes everywhere at once.
Invoicing that no longer gets rebuilt each month
Reports go in, the invoice and the recap come out. Traceability sits in the file, not in someone’s memory.
One team memory
The CRM keeps the date and the words of every decision. Six weeks later, the answer is there.
Clients found before competitors
The prospecting agent reads the news every week. The team chooses who to approach.
The thinking behind this mandate is laid out here: Connecting AI to your email, your ERP and your documents.

