Answers for IT, data protection and management.
Applications are full of personal data. So aktklar runs on your side, the AI only makes proposals, and people decide.
The application, text recognition and the optional AI run on your server. Nothing goes to outside services.
The AI shows where it got something from. aktklar checks numbers, dates and names against the case.
The activity log shows who changed what and when. A second person can sign off if you want.
Details by topic
The AI functions are off until the administration switches them on. They find passages and make proposals. People decide.
How we use AI
The AI in aktklar prepares and runs locally. Text recognition runs locally. The optional AI functions use a language model on your server or a computer in your LAN: the assistant with actions, inbox assignment, questions about your data, forecasts and simulations, and topic analysis. The application calls no cloud service, and we do not train on your data.
The model only phrases text from verified case data. Every number, date and name is checked against that data before a text appears; documents are only quoted, and a click leads to the source. If the check fails, a fixed text appears. A selection can still be wrong or incomplete, and an original can contain errors, so always check the document.
What the AI functions do and do not do
| They do | They do not |
|---|---|
| Find passages and quote them with their location | Assess eligibility, benefit amounts or persons |
| Propose actions, drafts and assignments that you confirm or reject | Recommend decisions, rank applications or rate employees |
| Select key figures whose numbers come from fixed calculations | Settle conflicts between sources legally or change cases without your confirmation |
Only the administration can switch the AI on, after confirming a notice, and off again at any time. Switching is logged.
EU AI Act: classification and duties
By our documented classification, aktklar is not a high-risk system. The AI functions prepare: finding passages, proposing assignments and drafts, selecting key figures. They give no assessment of eligibility of their own and do not assess people. That's a classification, not a guarantee.
The question comes up because of the topic: assessing eligibility for essential public benefits is on the list of high-risk areas. Systems that only do preparatory tasks and do not materially influence the outcome are exempt; profiling of natural persons is always high-risk. A human making the final call is not enough as a reason on its own, so the AI itself is limited to preparation. Our technical limits support this reading but do not guarantee it.
Status today: the classification is documented. Its review and sign-off and the registration are not complete yet and are planned before the first pilot with real case files. Whether the exemption applies in your specific deployment, we assess with you before delivery.
People should know when they work with an AI, so aktklar labels its AI functions in the interface. Dates follow the consolidated text of 27 July 2026: transparency duties from 2 August 2026, duties for high-risk systems under Annex III from 2 December 2027.
Source: EU AI Act, Annex III point 5(a), Articles 4, 6(3), 6(4), 49(2), 50 and 113. EUR-Lex, consolidated text of 27 July 2026
Is aktklar a high-risk AI system?
No, not according to our documented classification. The AI only prepares: it assigns documents, drafts text and shows summary figures. It does not decide on eligibility and does not evaluate individuals. That's our classification, not a guarantee. You get the classification and its supporting documents in the documents for data protection and IT. Registration is planned before our first pilot with real case files.
Source: EU AI Act, Annex III point 5(a), Articles 6(3), 6(4) and 49(2)
You are the controller. We provide the templates, you check and adapt them.
Who is responsible
That's you, not us. Your authority decides on the purposes and means of processing. Without remote access set up, we have no access to the installation. Maintenance or hosting with data access needs a separate agreement.
Source: GDPR, Articles 4(7) and 28
Where data is stored
On your own server. Case documents are files, metadata and logs in an SQLite database on your server. The optional model server also sits in your network. Backups and downloaded exports are stored where your IT decides.
Retention and deletion
A function shows cases that have reached the configured retention period. The administration confirms the deletion run; locked cases are skipped. There is no automatic scheduled deletion. Backups, server logs and downloaded exports are not deleted with it and need their own retention periods.
Training and impact assessment
No training on your data: aktklar does not use your case documents to train a model. An impact assessment under Article 35 GDPR is required for processing likely to result in a high risk. Whether your procedure falls under it, you check with your data protection advisor. Local operation does not replace that check.
The standard installation enables sign-in. Host hardening, TLS, backups and updates are the job of your IT.
What aktklar secures itself
- Login protection. Passwords are stored as scrypt hashes with their own salt, at least 12 characters and not on the local list of the 10,000 most common passwords. After 5 failed attempts within 10 minutes aktklar locks the user name and address. The login form is protected against CSRF.
- Two-factor sign-in. One-time codes under RFC 6238 from an authenticator app, set up per account or required for everyone. aktklar creates the QR code locally. Each code works once; the ten recovery codes are stored only as hashes.
- Session length. Sign-out after 8 hours idle and after 12 hours at the latest, both adjustable. A password change, a role change or a reset by the administration ends all sessions of the account.
- Roles. Case worker, head and admin, checked on the server.
- Second-person sign-off. Can be switched on per installation. Whoever last edited a case or letter cannot release the request or close the case. The server enforces this, not only the interface.
- Activity log. Every change in a case with the person from the session, the time and, where required, a reason. Deleting a case leaves a record that it happened and when.
- No cloud. aktklar runs on a computer in your network. The content security policy allows only its own origin; fonts, libraries, text recognition and QR code generation are local. No telemetry, no external AI.
Operation in your office: who does what
Your IT
- A server or VM in the office network (Linux or macOS) with storage for scans
- Disk encryption on the host
- A backup target for the daily backup of database and case folders
- HTTPS through a reverse proxy or directly, as you require
Your office
- Data protection impact assessment and record of processing, with our documents for data protection and IT as a template
- User accounts and roles, and the decision on second-person sign-off
- Rules and text blocks, set up together with us
We
- Installation package with install, backup and restore scripts
- Setup and training, operations and security documentation
- Updates as an installation package, applied on a date you choose
- Vulnerability checks of dependencies and a bill of materials (SBOM)
For data protection and IT we provide templates based on the actual implementation.
What the package contains
Record of processing under Article 30 GDPR, building blocks for the impact assessment, technical and organisational measures (TOM) with proof in the code, roles and permissions, a deletion concept and the operator role. Your office remains responsible; the templates are not legal advice.
Vulnerability checks and SBOM
A script compares the Python packages and the bundled JavaScript with public vulnerability databases and creates a bill of materials (SBOM) in CycloneDX format. The run for version 1.0 found no known vulnerabilities. It runs on our build machine; your server needs no internet for it.
Questions from IT or data protection?
In a call we show roles, sign-in and the activity log with sample data.