Use AI without giving your data away.

Coveniq is the multi-tenant AI platform for companies, public authorities and educational institutions. An upstream Privacy Guard detects personal data in every request, replaces it with placeholders and restores it only in the response. Technically enforced, not merely promised.

Request a demo

GDPR-oriented architecture
Fail-closed: no guard, no model call

coveniq-app.aitas.org

Summarize the case of Martina Berger briefly.

1 entity protected

The appeal against the fee notice dated 12 August was justified: the second home had been deregistered on 1 July. The fee paid in excess will be refunded.

Please draft the reply to Martina Berger, m.berger@musterstadt.de, regarding case KD-45678-B. The refund goes to DE89 3704 0044 0532 0130 00. Please keep it factual and brief.

4 entities protected

Dear Ms Berger, regarding case KD-45678-B we can confirm that the refund has been issued. The amount will be credited within the next working days.

Type a message…

Protected content

This content was pseudonymized before being sent to the AI model.

Martina BergerPerson
98 %Was unnecessary
m.berger@musterstadt.deEmail address
99 %Was unnecessary
KD-45678-BInternal ID
87 %Always protected
DE89 3704 0044 0532 0130 00IBAN
Always protected

The starting point

Productivity and data protection are not mutually exclusive.

Many organizations weigh up two routes: allow AI and carry the residual risk, or restrict its use and accept that it happens outside their own systems. Both routes leave the same question open, namely what leaves the building. A protective layer that sits technically in front of the model, and that every request has to pass through, answers it.

In everyday work

Personal data reaches AI tools because the task contains it: names from emails, details from tickets, personnel matters, file numbers. A policy sets out what is permitted. Whether it is followed at the moment of typing is not something the policy itself can ensure.

The evidence

Even when nothing has been disclosed, the question remains how to show it. That requires a layer that records what actually went to an external provider.

Weighing it up

Restrict the use and it moves to private accounts. Permit it and the overview of what leaves the building is missing. In both cases it remains open what a model has seen.

Coveniq separates what a user types from what the model gets to see.

Step by step

What the model sees cannot give your data away.

Five steps, played through once. Two run before the model call, two after, and between them lies the only point at which anything leaves the platform. Click a step to look at it in detail.

Privacy Guardready

Input

Please draft a reply to Martina Berger[PERSON_1] at m.berger@musterstadt.de[EMAIL_1] regarding case KD-45678-B[ID_1]. The refund goes to the account DE89 3704 0044 0532 0130 00[IBAN_1].

This is the text as it was entered. The Privacy Guard has not changed it yet.

Before the model call

After the model call

The guard does not catch every item and sometimes marks too much; you report both in one click, and your administrators sharpen it permanently for your organization.

How the guard works

Calibration

The guard learns your language.

Every organization has its own names, abbreviations and file numbers. Show them to the guard once, and it takes them into account from then on.

Google Maps is not a person, and EMP-1234 very much is an internal identifier. You can teach the guard both in two minutes, with a live test of the search pattern. Your corrections apply to your organization alone.

15 detected categories3 content layers per messageGuard response time under 2 seconds

Create correction

Fixing a false positiveFixing a false negative

Search patternEMP-\d+

CategoryInternal ID

Detection confidence

85%

Live test

Please check the cases EMP-4471 and EMP-2988 before sending.

2 matches in example text

CancelSave

Every pattern is tested against your own example before it is saved.

Knowledge spaces

Upload PDF, Word and text documents. Coveniq extracts, chunks and vectorizes them, and your assistants answer from your own content. Every document is checked for personal data on upload.

Assistants

As many assistants per organization as you need, each with its own system instruction, its own model and its own visibility. For the whole organization, for one department, or just for you.

Chat

Answers appear in real time, keep generating in the background when you switch conversations, and can be copied with formatting into Word or Excel. On request, the assistant produces finished documents as PDF or Word.

Tenants and sub-organizations

Complete data separation between organizations, enforced in the database through row level security. Departments and sites are modelled as sub-organizations with their own assistants and knowledge spaces.

Who it is for

Who are you?

Each area has its own case: what the guard detects there and what the model gets to see of it.

Companies

Make internal knowledge bases searchable with AI
Assistants for HR, legal and customer support
Usage analytics for cost control

HR has an assistant draft a reference letter. The employee name, her personnel number and her date of birth never reach the model. The draft comes back complete and correct.

Public authorities

Internal knowledge assistants without passing data on
File processing with automatic pseudonymization
Tenant separation between public authorities and their sites

A case officer has an appeal decision drafted. File numbers and citizen data stay in the building. What the model saw can be traced afterwards.

Educational institutions

Learning assistants for pupils and students
Institutional knowledge spaces for teaching material
Tenant separation between schools and faculties

A teacher uploads exam papers for review. The names of the learners do not leave the institution. The assessment comes back complete and can be handed to the class without rework.

Timeline

Where we stand and what comes next

We ship demonstrably and fast, and we show it publicly. Every version goes into live operation, not into an announcement.

Running
Live operation
German and EnglishRunning with real customers
- Stable operation under real conditions
- Calibrating the guard to the language of the connected organizations
- Extending the transparency and control features
- Improving response times and cost tracking
Next
Connecting your systems
- Single sign-on and SAML integration
- Compliance exports beyond usage and cost
- Public API for connecting your own systems
- Extended guard tuning and SCIM user provisioning
Later
Expansion
- Enterprise onboarding
- Service level commitments
- Extended billing model
- Further model providers beyond OpenAI and Anthropic

Application, database, file storage, and guard run in Frankfurt. Zero data retention applies to OpenAI and Anthropic: transmitted content is not stored there and not used for training.

Who is behind it

Coveniq is a product of AITAS Technologies.

AITAS Technologies builds software and automation around artificial intelligence and trains organizations in using it practically. The company is based in Freiburg.

Coveniq grew out of that work. In projects with companies, public authorities and educational institutions the same question came up every time: how can a language model be used without personal data leaving the building? Coveniq is the answer, built as a product rather than a one-off solution.

AITAS Technologies AITAS Academy

AITAS Technologies

AI software development and automation from Freiburg.

AI software development

Process automation

Training through AITAS Academy

Coveniq grew out of this work and is developed further by the same team.

Calibrate the guard to your language.

The platform is running, and it runs with real customers. What makes each organization distinct is its language: abbreviations, case numbers, turns of phrase. That is exactly what the guard is calibrated on. Show us yours, and we will show you the platform.