Three shared mailboxes, and nobody really owns the morning round.
Several hundred mails per week land in info@, service@, and accounting@. Everything arrives at once: customer requests, damage reports, supplier invoices, appointment confirmations, complaints, and now and then a deadline-bound letter as a PDF attachment.
Distribution happens by hand. In the morning someone opens the mailbox, skims the subject lines, forwards, and marks a few mails as unread again for later. Two people forward the same mail because neither sees that the other already took it. Other mails sit for days because the subject line does not say who is responsible. A mail nobody picks up is a mail nobody answers.
When the person who has known the mailbox for years is on vacation, that knowledge leaves with them. The stand-in sees the same list of subject lines but not the logic behind it: which supplier belongs to which cost center, which request goes to the technical team, which goes to accounting. The mailbox does not stop, the sorting quality does.
We wrote down the distribution knowledge first, then built the software.
The start was not a software workshop but a real week of mail at the mailbox, together with the people who sort today. And their conditions became the specification: replies only go out as drafts, the team maintains the responsibility matrix itself, and no mail disappears into a folder unseen. A common objection in the first conversation is: the AI cannot possibly know who is responsible for what here. True, which is why it does not do it alone.
- Walk through a real week of mail: Which requests actually occur, who is responsible per request, what makes a mail urgent, and which cases are genuinely ambiguous.
- Responsibilities as an editable matrix: Sender and request map to responsibility and stand-in, in a table the specialist team maintains itself. Head knowledge becomes something a new colleague can look up.
- One mailbox first, drafts only: The first scope is a single mailbox, usually the one with the highest volume. Every reply is a draft, which keeps the risk at zero while trust is built.
- Go-live only after the team’s okay: Every week the proposals are compared with what the team actually did. Every correction becomes a rule. It went live once the clerks said: this really saves us time.
80 percent is done by the AI employee. The human decides at two points.
The AI employee reads every incoming mail in full, meaning body text, thread history, and attachments, and recognizes the request against your own categories: inquiry, damage report, invoice, complaint, appointment, contract change. It also derives the priority and assigns the mail to the matching case, property, or customer record, using reference numbers, sender history, and the content of the thread.
Responsibility comes from the matrix the team maintains itself, supported by the history of comparable cases. For standard requests the AI employee writes the reply draft directly and places it in the responsible person’s Outlook, so the answer goes out from a real mailbox. Attachments move to SharePoint with metadata and a link to the case instead of staying buried in the thread. If it is unsure, it flags exactly the disputed field and explains why.
A tenant writes to info@ that water has been running down the bathroom wall since the morning, asks when someone will come, and attaches two photos.
One field is flagged for review: the responsible technician is on vacation and the matrix offers two stand-ins. Exactly this decision stays with the human.
That was before, this is today.
This is how the shared inbox ran before, and this is how it runs today with the AI employee. Conservatively calculated, that saves about 80 percent of the time per mail.
| Before | Today | |
|---|---|---|
| Distribution | Skimmed and forwarded by hand every morning | Classified and assigned on arrival |
| Duplicate handling | Two people work the same mail | One responsibility per request, visible to everyone |
| Mails left sitting | Unclear mails sit for days | Every mail has an owner and a deadline from minute one |
| Vacation cover | Sees subject lines, not the distribution logic | Works from the same editable matrix |
| Attachments | Stay in the thread and get searched for later | Filed in SharePoint with metadata |
* Savings conservatively calculated: before, a mail takes 2 to 3 minutes for reading, assigning, and forwarding, today about 20 seconds of checking, which is about 80 percent less time per mail. At around 700 mails per week across three shared mailboxes, about 29 hours per week drop to about 4 hours, so roughly 25 hours per week are freed up, a good 5 hours per working day. That corresponds to about 60 percent of a full-time position, so at 55,000 to 75,000 euros of employer cost per year, roughly 35,000 to 45,000 euros.
How do you build an AI employee like this, technically?
The knowledge from this scenario to take away, whether you build with us or on your own:
Classification: language models instead of keyword lists
A folder rule is a text match and moves a mail without knowing what is inside it. Language models like GPT from OpenAI or Claude from Anthropic read body text, thread history, and attachments and map the mail to your own request types. A reworded subject line no longer breaks the distribution.
Rules or AI: the responsibility matrix stays a rule set
Who is responsible for which request does not belong in the model but in a fixed table: request and sender determine responsibility and stand-in. The AI decides what it is about, the rule decides who gets it. Nothing is guessed, and a change is one line in the table, not a software release.
Several requests in one mail: split instead of squeeze
In a shared mailbox the mix is the normal case: one message reports a defect, questions an invoice, and announces new bank details. The AI employee splits it into sub-requests with their own responsibility and keeps them linked to the original thread. Then only one part is unclear, not the whole mail.
Integration: the existing systems stay
The AI employee connects through the Microsoft Graph API to Outlook and the shared mailboxes, attachments go to SharePoint. There is no new platform and no system migration. Sending runs through the responsible person’s personal Outlook account, so replies land where they belong.
UX: the interface decides adoption
A review screen shows the mail on the left and the recognized fields on the right, with exactly one flagged field instead of a hidden confidence score. The reply opens as an editable draft, nothing goes out unseen. And the responsibility matrix has its own view where the team changes it itself. Exactly these three things turned skepticism into approval.
What does it cost in comparison?
Superkind charges per use case. The price grows with mail volume, not with headcount. Here is the honest comparison:
| Manual distribution | Shared-inbox tool | Superkind AI employee | |
|---|---|---|---|
| Cost | 55,000 to 75,000 € per year per full-time position | 10 to 30 € per user per month | Price per use case, a fraction of a full-time position |
| What is included | The whole process, by hand | Organizing and assigning threads, the tool does not understand the content | Reading, classifying, assigning, reply draft, and filing |
| Scales with | More staff | Number of users | Mail volume, without new positions |
| Exceptions | Human does everything | Sit until somebody assigns them | Flagged with a reason and sent to a human |
| Rollout | Recruiting and onboarding | Create an account, distribution stays manual | 2 to 3 weeks to the first productive version |
The honest comparison is the full cost of manual distribution: the morning round itself, the duplicate handling, the mails left sitting, and the drop in quality as soon as the person who knows the mailbox is missing.
What we learned about shared mailboxes.
The hard part is not the reading. The models understand mails well by now. The hard part is that the sorting logic of a shared mailbox is written down almost nowhere. It lives in the person who has run info@ for six years and can tell from the sender alone whether something is urgent. Whoever skips that knowledge produces confident misroutings, and that is worse than slow but correct sorting by hand.
That is why we consider the draft the honest starting point. A team that has to release every reply knows after two weeks whether the AI employee is any good, and can prove it on their own mails instead of on a demo. Direct sending is something a request earns, it is not a launch feature.
What it is not suited for: If a mailbox only gets a handful of mails per day, sorting is not your bottleneck. In mailboxes where almost every mail is a one-off, there is too little repetition to learn from. And if nobody owns the responsibilities, that should be resolved first.
