Before touching n8n, the underlying process is: someone reads every inbound email, decides what kind of request it is, checks if the sender is a known customer, answers routine questions, and escalates the rest. That's four repeatable decisions (classify → check CRM → answer-or-not → log), which is what made it a good candidate for automation rather than a case where "just add an AI" would help. Anything genuinely ambiguous (angry customer, refund disputes, security reports) is deliberately kept as a human task — the workflow's job is to remove the manual triage step, not the human judgment step.
The system prompt is written as an ordered checklist (classify → tool-check → decide → draft-or-not) rather than a vague persona description, because ordered steps are what made the model reliably call the KB/CRM tools before answering instead of skipping straight to a guessed reply. The explicit instruction "never invent policy details" plus grounding tools is the main lever against hallucinated pricing/policy answers.
The urgent-priority carve-out ("do NOT draft a reply yourself") is in the prompt, not left to a downstream filter, because relying on a switch-node to catch a bad draft after the fact still means the draft exists — the safer point to stop the agent is before generation.
memoryBufferWindow keyed on sender email, window of 10 turns. A simple
buffer (not a full vector-memory) was the right call here: this is a
support inbox, not a research assistant — recent turns with the same
person are what matters, not fuzzy recall across the whole customer base.
The KB tool is scoped narrowly (search_company_kb, top-k 4) and described
to the agent as something to check before answering factual questions —
the tool description itself is the RAG-usage instruction, since the agent
decides when to call it.
- Ran representative sample emails (a billing question, a spam email, an angry cancellation request, a first-time inquiry) through the agent node in isolation (n8n's "Test step" on the AI Agent) before wiring the Sheets/CRM/LINE branches, to check the structured output matched schema every time.
- Checked the urgent branch specifically for false negatives — an angry email classified as "standard" is the failure mode that matters most, so that path got the most manual test cases.
- Used n8n's execution log / pinned data to replay a specific past email through the workflow after editing the prompt, instead of waiting for a new live email each time.
- The Sheet is the single place ops needs to check daily — it has everything (category, priority, summary, human-needed flag) without needing to open n8n at all.
- The system prompt and JSON schema are the two things to touch if behavior needs to change (e.g. add a new category) — documented in the README so a future editor doesn't need to trace the whole graph to find them.