Anyone who wants a project status has to ask the person in charge and wait.
A German construction and property developer with its own planning subsidiary manages thousands of legacy projects and a live portfolio nobody can hold in their head anymore. The knowledge about building-permit procedures sits in the heads of experienced architects: how a development plan turns into a pre-application, then a building application, then an amendment. None of it is written down.
The documents are scattered across many storage locations: a Teams and SharePoint channel per project, an exchange platform for external partners, and an old network drive nobody ever cleaned up. A single building application is a folder with around 30 PDFs, checked against a requirement list from the building authority with about 50 points.
On top of that come two silent risks. Versions: the newest date is not necessarily the valid one, because a file marked “final”, one marked “submitted final”, a review folder, and an archive all sit next to each other. And deadlines: building permits expire after two to three years, with nobody sending a reminder.
We wrote the process down first, then cut the scope small.
There was no documentation to read, so we produced it. And the customer’s conditions became the specification: nobody should walk off with wrong information, so every answer shows its source. Permissions stay exactly as they are in the file system today. And everything the AI employee formulates anew needs a human release.
- Get the knowledge out of people’s heads: The person who owns the process told the whole permit procedure end to end, from pre-application to amendment, before anyone tried to model it.
- Draw the process and let the customer correct it: We drew the workflow and showed it back. The customer’s corrections, not our draft, became the process map: which document is mandatory when, who signs, which deadline starts running.
- Deliberately small start: The customer put it themselves: the basics first, the rest later. Connecting the authority portals and the BIM integration were postponed in writing instead of being promised.
- Pilot on a live project: The first version covers three things: finding documents, looking up procedural knowledge, answering status questions. One running project served as the pilot, one named person on the customer side made the decisions.
80 percent is done by the AI employee. The human decides at two points.
The AI employee indexes the scattered storage locations and maps every file to a project and a procedural stage. From that it derives the status: in progress, submitted, additional documents requested, or approved. It checks the submitted documents against the roughly 50 points of the authority list and names the gaps. Every answer shows the source, the link to the exact location, and a note on how confident it is.
Two stations stay with the human. First, the release: everything the AI employee formulates anew lands in a separate area and only moves into the released directory after review. Released and not-yet-released knowledge never appear in the same view. Second, the deadlines: the AI employee flags an expiring permit to a named person, but whether it is extended, reapplied for, or the project is stopped is a human decision.
Where do we stand on the building application for this project, and what has to happen next?
One field is flagged for review: the version conflict. The AI employee does not pick a winner between two files, it names both locations and lets the human decide.
That was before, this is today.
This is how project reporting ran before, and this is how it runs today with the AI employee. Conservatively calculated, that saves about 80 percent of the active working time per status question, and the waiting time disappears entirely.
| Before | Today | |
|---|---|---|
| Project status | Ask the person in charge, wait hours to days | Answer in under a minute, with source and link |
| Status meetings | A fixed round so everyone knows the state | Only for decisions, not for status updates |
| Completeness check | Manual against the roughly 50 authority points | Automatic, with the gaps marked |
| Finding a document | Digging across Teams, platform, and network drive | Direct link to the exact storage location |
| Deadlines | Permits expire quietly | Watcher flags them in advance to a named person |
| Process knowledge | Only in the heads of experienced architects | Written knowledge base, released by humans |
* Baseline documented in the process mapping with the specialist team, 2026. Savings conservatively calculated: before, about 3 minutes for the person asking plus about 10 minutes for the person in charge, who interrupts their work and searches several storage locations, so about 13 minutes of active time; today about 1 minute for question and answer plus about 1 minute to verify the source link, so about 2 minutes for one person. The waiting time of hours to days is not counted in. The status-meeting time is stated as a range because it differs per team. The person-years are the customer’s own estimate for a manual work-up of the legacy projects, not a measurement of ours.
How do you build an AI employee like this, technically?
The knowledge from this project to take away, whether you build with us or on your own:
Understanding documents: language models instead of keyword search
A keyword search finds file names, but it does not understand that an acknowledgment of receipt from the building authority proves a procedural stage. Models like Claude from Anthropic or GPT from OpenAI read notices, minutes, and plans in context and map them to a stage. For scanned legacy documents, a vision model runs in front of that.
Indexing: dock onto the existing filing
Through the Microsoft Graph API the AI employee connects to SharePoint and Teams, plus the exchange platform and the old network drive. There is no new platform and no system migration. Important: the index is filtered per user and at query time, otherwise a scattered folder becomes a leak.
Retrieval: link sources instead of copying content
The AI employee does not create a second copy of your documents. It remembers where something sits and links the location. That keeps the filing the single source of truth, permissions and deletions take effect immediately, and nobody maintains two versions in parallel.
Rules instead of AI wherever rules suffice
Deadline logic and mandatory documents are fixed rules, not model decisions, and they differ per federal state because building law in Germany is state law. The specialist team maintains those rules itself. For version conflicts the hardest rule applies: never guess, always ask back.
UX: the interface decides adoption
Three things turned skepticism into approval here. Every answer comes with a link to the exact location and a note on how confident it is. Everything AI-generated sits in its own area until a named person moves it into the released directory. And a deadline dashboard shows at a glance which permit expires when and who owns it.
What does it cost in comparison?
Superkind charges per use case. The price grows with project volume, not with headcount. Here is the honest comparison:
| Status by asking around | PM software | Superkind AI employee | |
|---|---|---|---|
| Cost | No license, but waiting time on every question | 10 to 30 € per user per month, plus the time for upkeep | Price per use case, a fraction of a full-time position |
| What is included | The knowledge in one person’s head | Tasks, dates, and reports, but only if somebody maintains them by hand | Search, mapping, completeness check, deadlines, and an answer with its source |
| Where the status comes from | The memory of individual people | Whatever somebody typed into the tool | The real files in Teams, SharePoint, and the drive |
| Scales with | More staff | More licenses and more upkeep | Project volume, without new positions |
| Exceptions | Left with the person in charge | Never make it into the tool | Flagged and sent to a human |
| Rollout | None, it stays as it is | Weeks to months, plus upkeep discipline in the team | 2 to 3 weeks to the first productive version |
The honest comparison is the full cost of the current state: the follow-up questions, the searching across several storage locations, and the permits that expire because nobody was watching.
What we learned from this project.
The hard part is not finding project documents. The hard part is deriving a status from scattered files that an architect will trust. Trust here does not come from a percentage, it comes from a link: whoever doubts opens the source and sees within seconds whether the statement holds. A confidence number on its own cannot be verified by anyone.
The underrated side effect: the process documentation appears as a by-product. For an AI employee to answer reliably, someone has to state which document is mandatory at which stage and which deadline starts when. That description is exactly the documentation this company never had time to write. It stays valuable even independently of the software.
What it is not suited for: With a handful of parallel projects, the person in charge has the status in their head anyway. Without digital filing the basis is missing, because nobody can search what was never scanned. And if nobody owns the releases, new knowledge piles up and the knowledge base never goes live.
