Werbetechniker-Shop: managing daily email workflows with ChatGPT and Mailbox MCP
Werbetechniker-Shop, a B2B wholesaler of signmaking supplies, connected its hosted Microsoft Exchange mailbox to ChatGPT with Mailbox MCP on 1 October 2026. Five weeks later, its assistant reads, searches, files, drafts and sends inside that mailbox as part of the shop's daily work, and managing director Sven Kassmann starts each morning with a summary of what needs him.
Sven Kassmann
Geschäftsführer (Managing Director), Werbetechniker-Shop
Technical actions recorded by Mailbox MCP, 1 October to 4 November 2026
8,081 technical actions in the mailbox in 35 days, on 31 of those days
About 1,900 technical actions a week over the last four weeks
These are technical actions recorded by Mailbox MCP. They are not counts of unique customer cases, and they do not measure time saved or efficiency.
99.3% of technical actions completed without a reported error
The 99.3% is the proportion of technical actions completed without a reported error. It is not a measure of the business accuracy of replies or of completed case handling.
One hosted Microsoft Exchange mailbox, connected to ChatGPT with permission to send
- Emails read 3,414
- Searches 1,874
- Messages filed 401
- Emails sent 285
- Drafts written 227
- Attachments read 177
The challenge: Customer email at a B2B wholesaler
Previously, I had to review, organise and follow up on most customer enquiries, supplier confirmations, shipping updates and payment questions myself. Reconstructing the current status of longer-running cases was particularly time-consuming.
With Mailbox MCP: What it does for them now
The assistant uses Mailbox MCP to review the inbox, read related messages, file routine correspondence and prepare replies. Clearly authorised standard replies are sent automatically, while quotations, complaints and exceptions are still decided by people.
Each morning I receive a summary of pending decisions, cases waiting for a response and upcoming follow-ups.
Mailbox MCP is the email layer of a wider setup: the assistant also checks progress against the shop system and the company's knowledge base.
Why it works for them in the Exchange mailbox: The work stays visible
What I find especially useful is that the work remains visible in our existing mailbox. I can review drafts and see the messages that were actually sent.
Sven Kassmann, Geschäftsführer (Managing Director), Werbetechniker-Shop
Nothing moves to a separate tool. Drafts appear in the mailbox's own Drafts folder, sent replies in its Sent folder, and filed messages in the folders the team already uses, so Sven can check the assistant's work in his normal mail client at any time.
For the email workflow: The practical benefit
In Sven's words, the practical benefit is "spending less time searching, sorting and going over the same cases repeatedly." The shop has not measured time or cost saved, and it is still working on reducing unnecessary repeat checks and overall workflow consumption.
The shop's email workflow: What's next
Sven's wish list is about making a busy workflow leaner still: checking only new or changed messages, and recognising cases that are already closed. Both are on our list.
About 80% reading and finding: what AI assistants do in a mailbox
Across all Mailbox MCP customer mailboxes from 1 October to 4 November 2026, about 80% of technical actions by AI assistants were reading and finding, and 2.5% were sending, by Mailbox MCP's own records.
From here on, this is ours and not Sven's. By Mark McNeece, Founder of Mailbox MCP. Sven approved the story and his figures above; what follows is my reading of them.
One wholesaler beside all customer mailboxes: About 12 emails read for each one sent
People picture an AI in their email as the thing that sends, because that's the part you can see. Our records say sending is the small end of the job, and this one Exchange mailbox says the same on its own.
That's close, not identical: the columns below group the tools slightly differently, and the second row sits further apart. Either way, a sent email is the visible end of a long stretch of looking things up.
| Measure | This shop | All customers |
|---|---|---|
| Emails read for each one sent | about 12.0 | about 12.9 |
| Read or search actions for each one sent | about 19.2 | about 26.2 |
| Sending, as a share of all actions | 3.5% | 2.5% |
Emails read for one sent
This shop about 12.0
All customers about 12.9
Email permissions and prompt injection: What this means
Looking and sending are separate decisions. Every signed-in connection carries one of 3 permission levels. Of this shop's named actions, 5,465 would run on Read only, 628 need Draft and file and 285 need Send and delete. That's no verdict on the shop's setting: sending authorised replies is part of the job. But only the last level puts something in front of a stranger.
The protocol these connectors run on says where a person belongs:
For trust & safety and security, there SHOULD always be a human in the loop with the ability to deny tool invocations.
That asks for the ability to deny, rather than someone reading every call. Here the person has moved: quotations, complaints and exceptions are decided by people, and what went out can be read in Sent.
The shop's named actions, by level
Read only 5,465
Draft and file 628
Send and delete 285
Reading has a risk of its own. A mailbox is private data, incoming mail is written by outsiders, and a send can carry things out. Simon Willison calls that combination the lethal trifecta:
If your agent combines these three features, an attacker can easily trick it into accessing your private data and sending it to that attacker.
An email connector with permission to send has all 3 by design, ours included. These figures don't measure that risk and don't suggest it arose here. So judge a connector less on sending, and more on whether it leaves the unread flag alone, reads inside attachments and can be stopped at reading. There's more in is it safe to give AI your email? and can ChatGPT read my email?
data
to send
Limits and method: What this ChatGPT case study can't tell you
It's one shop over 5 weeks: a case study, not a benchmark. Actions are not outcomes, and nobody has measured a saving. The all-customer shares describe what assistants did and imply nothing about time, cost, efficiency or accuracy.
The shares come from our own records, read on 5 November 2026 for the same dates: free and paid mailboxes together, our own accounts left out. A call is one use of one tool: a technical action, and not a task or a customer case. Reading means opening emails, threads and attachments and listing mail; searching includes finding a contact; organising means moving, flagging, marking, deleting and folders.
This customer story: Questions people ask
What do AI assistants mostly do in a mailbox?
Mostly reading and finding. Across all Mailbox MCP customer mailboxes from 1 October to 4 November 2026, about 80% of technical actions were reading and finding, and 2.5% were sending, according to our own records. It describes what assistants did. It is not a benchmark and says nothing about time, cost, efficiency or accuracy.
What counts as a technical action in these figures?
One technical action is one use of one tool by an assistant through Mailbox MCP, such as reading an email, running a search or filing a message. They are not counts of unique customer cases, and they do not measure time saved or efficiency.
Does the 99.3% measure how accurate the replies were?
No. The 99.3% is the proportion of technical actions completed without a reported error. It is not a measure of the business accuracy of replies or of completed case handling.
Did Werbetechniker-Shop measure time or money saved?
No. The shop has not measured time or cost saved, and it is still working on reducing unnecessary repeat checks and overall workflow consumption.
Which mailbox and AI assistant does Werbetechniker-Shop use?
One hosted Microsoft Exchange mailbox, connected to ChatGPT with permission to send. The shop connected it with Mailbox MCP on 1 October 2026.
Is the shop system part of Mailbox MCP?
No. Mailbox MCP is the email layer of a wider setup: the assistant also checks progress against the shop system and the company's knowledge base.
Connect the mailbox you already have
Your mail stays where it is, and the work shows up in your own folders. 5 calls a day on the free tier, with no card.
Related: connect Outlook or Microsoft 365, add it to ChatGPT and pricing.
Sources
- Werbetechniker-Shop: the story and quotations are Sven Kassmann's, approved with the figures in writing on 5 November 2026.
-
Mailbox MCP's records, the shop's figures: first-party records for this one account, published with
his approval and his caveats. Emails read come from
read_email, messages filed frommove_email(one call can carry more than one message), drafts fromdraft_replyanddraft_email, and emails sent fromsend_draft,reply_emailandsend_email. -
Mailbox MCP's records, the all-customer shares: read-only query of our audit log, 5 November 2026.
Reading:
read_email,read_thread,read_attachment,list_emails,catch_up,owed_replies,waiting_on. Searching:search_emails,select_emails,find_contact. Sending:send_email,reply_email,forward_email,send_draft,approve_and_send. Shares are rounded. - Model Context Protocol specification, version 2026-07-28: Tools, read 5 November 2026.
- Simon Willison: The lethal trifecta for AI agents, 16 June 2025, read 5 November 2026. Neither quotation is Sven's view.