DeutschlandGPT

Voicemails land overnight in a shared service mailbox. The agent runs Transcribe Audio on each recording, passing German as the spoken language and seeding the prompt with product names and the account terminology the caller is likely to use, which is where most transcription accuracy in a support context is actually won. Create Chat Completion then classifies the request, extracts the customer reference, and drafts a reply in German. The case, its transcript and the draft are written to the service desk and posted to the team channel. The team opens their queue and sees classified cases with drafts attached instead of a row of audio files.

An internal knowledge flow embeds policy documents, contracts and process notes with Create Embedding, batching several texts per call, and stores the vectors in Qdrant or pgvector alongside the source references. When an employee asks a question, the agent embeds the question, retrieves the matching passages, and answers with Create Chat Completion grounded in those passages. The documents, the question and the answer all move through the same German endpoint, so the retrieval layer inherits the residency posture of the generation layer instead of introducing a second one.

A product release is approved in Notion. The agent drafts the German announcement copy with Create Chat Completion and produces the accompanying visual with Generate Image. It then renders the spoken version with Generate Speech, which lands the audio in FlowRunner file storage and hands back a URL later steps can use. The text goes to the site, the audio goes to the internal channel, and every asset was produced through the same connection rather than assembled from three vendors with three different data-processing agreements.