About the project

Tannenhof Berlin-Brandenburg is an addiction therapy provider in Germany. Therapists spent significant time taking paper notes during sessions and re-entering them into the Patfak digital patient file system afterwards – slow, error-prone and exhausting. The documentation had to be formatted for German Pension Insurance (DRV) reporting, adding another layer of manual work.

N3XTCODER built an AI transcription pilot, funded through Civic Coding, that records therapy sessions on work iPhones and iPads, transcribes them, and produces structured reports formatted for DRV reporting – replacing the paper-notes-to-system pipeline entirely.

The challenge

Healthcare transcription is not a standard NLP task. The requirements went beyond accuracy:

• Patient data is life-critical. Audio recordings of therapy sessions are among the most sensitive data categories. EU-hosted infrastructure with hardened access controls was non-negotiable.

• Clinical terminology must be right. Addiction therapy has its own vocabulary. A transcription that gets the clinical terms wrong is worse than no transcription.

• Output must match DRV reporting format. The German Pension Insurance has specific reporting requirements. The AI output had to be structured to match, not just raw text.

• Clinicians must trust it. If therapists do not trust the system, they will not use it. Buy-in from the clinical team had to be built into the process, not assumed.

The approach

We started with a premortem session before writing any code. The team mapped failure modes across three categories: legal change (what if regulations shift mid-pilot?), data security (what if a recording is accessed by the wrong person?), and user acceptance (what if therapists reject the tool?). This shaped every architecture decision that followed.

The prototype records sessions on work iPhones and iPads, transcribes them using speech-to-text models on EU-hosted infrastructure, and produces structured summaries formatted for DRV reporting. We hardened storage bucket security, A/B-tested against the standard Patnova product, and renamed UI terminology from "Monologue" to "Einzel" after clinical-team feedback – a small change that made a big difference to adoption.

The pilot used 32 hours of Civic Coding consulting and was built feature-first: the smallest useful version shipped first, then iterated based on real clinical feedback.

Tech stack

• Transcription: Speech-to-text on EU-hosted infrastructure

• Recording devices: Work iPhones and iPads

• Output format: Structured reports formatted for German Pension Insurance (DRV)

• Patient file integration: Patfak digital patient file system

• Infrastructure: EU-hosted with hardened storage bucket security

• Funded under: Civic Coding (Germany's federal initiative)

The outcome

• Premortem-first approach mapped legal, security and user-acceptance failure modes before any code was written

• A/B-tested against the standard product (Patnova) to validate that AI transcription met clinical quality standards

• Clinical team shaped the UI – terminology changes and workflow adjustments based on therapist feedback

• Reached the milestone where Tannenhof can request internal budget for full 2026 implementation

• Datenschutz, Patientenschutz und Arztgeheimnis built into the architecture from day one – not retrofitted