ChatGPT is already in the building.
Show it outlines only.
In many companies, customer emails already pass through ChatGPT, without anyone ever deciding that. Umbra masks personal data before the text leaves the building: the AI sees the outlines of the case, the real names stay in the core shadow.
![Editorial illustration: a letter crosses a core-shadow band in front of a copper-glowing eclipse corona; inside the shadow its lines are mere silhouettes, outside they carry monospace placeholders like [[PERSON_1]], [[EMAIL_1]] and [[PHONE_1]]; below sits a small vault box marked with a key.](/img/projects/umbra/hero.png)
![Editorial illustration: a letter crosses a core-shadow band in front of a copper-glowing eclipse corona; inside the shadow its lines are mere silhouettes, outside they carry monospace placeholders like [[PERSON_1]], [[EMAIL_1]] and [[PHONE_1]]; below sits a small vault box marked with a key.](/img/projects/umbra/hero-light.png)
Shadow AI: the quiet routine
A customer email pasted into the chat window, a usable answer back: that is what everyday AI use looks like in many companies. The customer's name, email address and contract details travel along, to a server outside the building, without anyone ever having decided that.
A ban changes little. The tool is too useful, and what is useful gets used, via a private phone if necessary. Ending the quiet routine takes a path that is safe and still convenient.
Redaction is not that path. Redacted text is safe and worthless: the AI answers around black bars, and nobody can work with the result. The case has to stay readable, just not the person.
The AI sees outlines, not faces
Worth knowing: as long as the mapping exists, the data can be restored. That is by design, otherwise the AI's answer could not be translated back. Which is exactly why the mapping lives briefly, in RAM only, and only with you. After two hours it is gone; on request, immediately.
What it can’t do
No tool deserves trust whose limits you only discover in an emergency. These three belong on the table up front:
It does not detect everything. The detection model is trained mostly on English text. It finds salutations and signatures reliably; German first names in mid-sentence can slip through. That is what the deny-list is for: terms placed there, such as customer names or internal project names, are always masked, independent of the model.
Pseudonymous does not mean anonymous. As long as the mapping list exists, the data is personal within the meaning of the GDPR. That is why the list is built to offer as little attack surface as possible: RAM only, never on disk, gone by itself after two hours.
It has no login of its own. Umbra belongs behind your own network’s access control, not on the open internet. That is a deliberate decision: a building block for your own environment, not a public service.
Mask
Every person, email address and number gets a stable placeholder: [[PERSON_1]] instead of the real name. This happens on your own premises, before the text leaves them.
The AI works
The language model receives only the masked text. It sees the structure of the case and can answer precisely, because [[PERSON_1]] and [[PERSON_2]] stay distinguishable.
Translate back
The answer comes back and is restored to real names in-house. The mapping never left your own machine.
Frequently asked questions
Is this anonymisation?
No, and deliberately so. Umbra pseudonymises under Art. 4(5) GDPR: the data remains personal as long as the mapping exists, and that is exactly why the AI's answer can be translated back to real names. Pseudonymisation is a recognised safeguard (Art. 25 and 32 GDPR). Tools that promise you "anonymisation" at this point promise more than the GDPR allows.
Where does my data go?
Not to Aeterna Labs. Umbra runs entirely on your premises, on your hardware or in your environment. Only the masked text goes to the AI, and which AI that is remains your choice. Neither your texts nor your mappings ever leave your company towards Aeterna.
What happens after two hours?
The mapping between placeholders and real names lives exclusively in RAM and expires by itself after two hours; the period is configurable, and it can be deleted immediately at any time. None of it is ever written to disk. The mapping does not survive a restart, and that too is by design.
Can I buy this?
Umbra is a workshop piece, not a product with a price list. It is developed under the Apache-2.0 licence; publication of the source code is in preparation. If you want something like this in your company, adapted to your data and systems, the way there is a conversation.
Use AI without shipping customer data?
Umbra is a workshop piece, not off-the-shelf software. If your company needs something like this, adapted to your data and systems, a conversation is worth it.
Send an email →