Zammad-AI is a GenAI middleware layer for the Zammad ticket system. It supports helpdesk teams with AI-based ticket triage, suggested answers and knowledge-based responses, while the final responsibility always stays with the staff.
Zammad-AI is deliberately built as a separately deployable component next to Zammad rather than part of its core. This decouples GenAI features from Zammad release cycles and keeps the leading system untouched. It picks up incoming tickets over event-driven interfaces, classifies them, routes them based on configurable rules and drafts answers using connected knowledge sources such as the Zammad knowledge base and the Dienstleistungsfinder KI-Suche data set. Answers are always marked as AI suggestions, are reviewed by staff before use and are only automated end-to-end for clearly bounded, low-risk categories.
The AI models are modular and are consumed via an open-source model hosting service (gpt-oss-120b for triage and answers, qwen3-embedding-4b for retrieval). Its own guardrail service additionally checks prompts and answers for personal data and sensitive content.
Technical details
Zammad-AI consists of three Python services operated as containers:
- Workflow engine: FastAPI and FastStream service for REST endpoints and event-driven Kafka processing of ticket events.
- Index pipeline: synchronises knowledge content into the Qdrant vector database for retrieval.
- Guardrail service: independently operated service for content checks of prompts and answers.
- Observability: tracing and prompt management with Langfuse, metrics with Prometheus and Grafana.
More information: https://ki.muenchen.de/ki-systeme/zammad-ai