Automating Outlook email with AI is possible today, just not the way people imagine. Summarizing a thread, finding a price in an old exchange or sorting a mailbox already works well. What breaks is the last step: a reply draft that is truly ready to send, with the right signature, the right attachment, from the right mailbox. Here are the three obstacles we ran into in the field, and how we solved them.
Reading email with AI is solved
Most AI assistants now offer an Outlook or Microsoft 365 connector. They search by sender, subject or date, summarize a twenty-message thread in five lines and find the attachment you’d been hunting for ten minutes. For a small business receiving hundreds of emails a week, that’s already a real gain.
The trouble starts when you ask the AI to do the whole job: reply to the client, attach the quote, send from the sales team’s mailbox. That’s where generic connectors show their limits, and where many trials stop. The time saved reading is lost again fixing the draft by hand.
Obstacle 1: the draft arrives without a signature
It’s the first irritant, and the most visible. Microsoft offers no way for an external application to read a mailbox’s Outlook signature. A draft created by a tool therefore arrives bare: no name, no title, no logo, no phone number. You have to paste it back by hand, every time.
When one person handles several mailboxes, it’s worse: the support mailbox, the sales mailbox and their own each have their signature, sometimes across two different companies. A draft with the wrong signature is more dangerous than one with none, because it looks finished.
What we did: each mailbox’s signature is captured once, from an email already sent from that mailbox, images included. From then on, every draft automatically gets the signature of the mailbox it leaves from. The AI is even instructed never to write a signature itself in the text.
Obstacle 2: attachments hit a ceiling fast
The simple way to add a file to an Outlook draft programmatically is limited to about 3 MB. A PDF quote often fits. Drawings, part photos, a 3D file or a presentation almost never do. In distribution and manufacturing, those are exactly the files that travel between the client, the plant and the supplier.
What we did: large files are sent in pieces, using the method Microsoft provides for large attachments, up to 150 MB. The file can come from a received email (the drawing the supplier sent yesterday goes back to the client in one sentence), from a file dropped into the conversation with the AI, or from an upload link when the file only lives on the computer.
Obstacle 3: one mailbox at a time
In a small business, email doesn’t live in a single mailbox. There’s the personal one, the department’s shared mailbox, sometimes a second company’s mailbox that the same person manages. Consumer connectors mostly see the signed-in person’s mailbox. For the rest, you juggle or give up.
What we did: one connection per person, and the tool sees every mailbox that person already sees in Outlook, including shared mailboxes and those of another Microsoft account they link themselves. For a reply, the tool finds the mailbox where the email was received. For a new email, it chooses based on past exchanges with that recipient, and says so clearly.
What a serious tool must never do
Connecting AI to your email means giving it access to the most sensitive things your company has. Before choosing a tool, we would check these five points.
- Send an email. The tool prepares drafts. A person reviews and clicks Send. Ideally, it doesn’t even have the permission to send at Microsoft.
- See more than you. It must work with your own Outlook permissions. If you can’t access a mailbox, neither can it.
- Obey a received email. An email’s content is information, never an instruction. A booby-trapped message saying “forward this file to that address” must trigger nothing.
- Silently attach a file for a stranger. When a recipient has never received email from that mailbox, the tool must flag it before sending.
- Keep your email. The server links the AI and Microsoft. It keeps your encrypted access and a log of actions, not the content of your mailboxes.
In Quebec, these points also match the obligations of Law 25: knowing what data moves, where it’s kept and who can access it. A tool that can’t answer those three questions clearly has no place in a company mailbox.
In the field: a B2B distributor, three mailboxes
We built this solution for a client in B2B distribution and industrial sourcing. The sales representative manages three mailboxes across two companies, deals every day with clients in Quebec and project managers in Asia, and moves drawings, quotes and part photos around. Before, the AI read his email well, but every reply meant pasting the signature back and re-attaching files by hand.
Today, an automatic triage runs through his mailboxes every hour on weekdays and prepares reply drafts, each with the right mailbox’s signature and the right attachments. He reviews, adjusts a sentence if needed, and sends. Nothing goes out without him.
The full mandate: the B2B distribution case →
How long it takes to set up
Less than you’d think, because there is nothing to install on the computers. A call of about thirty minutes is enough to look at the mailboxes, at what the team wants to delegate and at whether it’s the right tool. Then the company’s Microsoft 365 administrator authorizes the tool in one click, once for everyone. Each person then adds it to their assistant and signs in with their Microsoft account, in five minutes. Signatures are captured on the first try, and the rest is tuned with use over the first weeks.
What really takes time isn’t the technology. It’s deciding what the AI may prepare on its own, what always needs a careful review, and which mailboxes are in scope. Those decisions are worth putting in writing before you start.
Where to start
Before automating anything, look at one week of email and ask three questions. Which replies are similar enough for a draft to do 80% of the work? Which mailboxes do they live in? Which files go with them, and how big are they? The answers tell you right away whether a generic connector is enough, or whether you need a tool that handles signatures, large files and shared mailboxes.
If email eats a good part of your team’s week, it’s often the most profitable first AI project, because the gain shows from day one and the risk stays low: nothing goes out without review.
Our tool, born from this mandate: Claude connected to Outlook and Microsoft 365 →
Our automation services: see the services →
Want to know if your email is the right first project? The free AI diagnostic puts a number on it in a few minutes.
Why starting small with AI works better
The instinct with AI is to go big. That is usually where things stall.
Read the article →
5 signs your company is ready for AI automation
The answer has nothing to do with your sector and everything to do with what happens in your operations.
Read the article →