AI Workflow Automation / Case study
AI News Agency
An n8n editorial workflow with source checks, human approval, and controlled dispatch.
Workflow
- Ingest sources and normalize records
- Deduplicate and filter for relevance
- Plan, draft, and review against source material
- Send the draft for human approval in Telegram
- Dispatch approved content
Context
An editorial workflow has to handle overlapping sources, material that falls outside its coverage, and drafts whose claims may outrun the evidence. Automating the preparation of drafts creates a useful handoff only when reviewers can see what the draft relies on and can control what gets dispatched.
Approach
The n8n workflow separates source intake, editorial decisions, and delivery. Deduplication and relevance filtering narrow the material before drafting. LLM-assisted planning, drafting, and review operate on that filtered material. A draft then goes to Telegram for human approval before the dispatch step.
Engineering decisions
- Ground claims in collected material. Keep source references available for review so an editor can check where a statement came from.
- Keep review and delivery separate. Approval is an explicit workflow stage. A draft waiting for a decision stays on the review side of that boundary.
- Distinguish failure states. A run with no relevant material has a different next step from a run with a draft needing revision or a failed dispatch. Treating those outcomes separately helps the operator decide what to do next.
- Preserve traceability across handoffs. The source material, draft, review decision, and dispatch outcome are distinct checkpoints for diagnosing a run.
What the work surfaced
The most consequential decisions sit between stages. A reviewer needs enough source context to judge a draft. A delivery step needs a clear approval state. A failed dispatch warrants checking the delivery outcome before retrying so an operator can avoid sending the same item twice.
Current status
This case study describes real workflow work. For a general explanation of the reliability patterns, see Reliable AI workflows need more than a good prompt.