Add LLM steps to business automations

AI triage automation for Hindi and English support tickets

Build a workflow (n8n or Zapier) that takes a support message in Hindi or English, uses an LLM step to classify category and urgency and pull key fields into structured JSON, drafts a suggested reply, and routes high-urgency tickets to a human-approval step before anything is sent. Feed it real bilingual text from the MASSIVE dataset, which already carries intent labels, so you can check what the LLM step got wrong instead of guessing.

MASSIVE (AmazonScience) hi-IN and en-US utterances as your ticket text — real user requests in both languages, already intent-labelled so you can grade the LLM step

1. Get the workflow running end to end with a hardcoded LLM response · ~1.5h
2. Write the classification prompt until it returns parseable JSON on 10 varied tickets · ~2.5h
3. Add the malformed-output catch and the error log · ~1.5h
4. Add the urgency branch and the human-approval gate before any send · ~2.5h

• LLM step reliably returns valid structured JSON (test with at least 10 varied ticket examples)
• Workflow handles both Hindi and English input correctly
• High-urgency tickets are routed to a human-approval step, not auto-sent
• Malformed LLM output is caught and logged rather than crashing the workflow

Evidence a recruiter can check

• Ten input/output pairs across Hindi and English showing exactly what the LLM step returned, including at least one it got wrong
• The exported workflow with the JSON-parse failure path visible as its own branch in the node graph
• A recording of a high-urgency ticket stopping at the human-approval gate instead of auto-sending
• The prompt itself, with the output schema and the retry you added to stop malformed responses

Added an LLM triage step to a support automation — classifies Hindi and English tickets into structured JSON, catches malformed output instead of crashing the flow, and holds high-urgency replies for human approval.