Recommendation (short): Use Zapier to catch Gmail and Intercom events, call ChatGPT to classify + draft replies, then route by priority using Zapier Paths/Filters. Start by auto-tagging and auto-responding only for low-priority messages; escalate mid/high-priority to human via Slack/Intercom assignment. Iterate after 1–2 weeks.

Why this works: Zapier handles integrations (Gmail, Intercom, Slack, Asana), ChatGPT does NLP and draft responses. This splits triage (automatable) from judgment/escalation (human-in-loop) so you can handle ~200 messages/day safely.

Step-by-step setup (two Zaps, minimal):
1) Zap A — Gmail auto-triage
- Trigger: New Email in Gmail (unread) with label or inbox.
- Action: Send message content to ChatGPT (classification prompt — see prompt snippet below).
- Action: Parse ChatGPT output (JSON with tags, priority, suggested_reply, confidence).
- Path/Filter: If priority = high -> Create Slack alert + Create task in Asana (assign to on-call); Add Intercom ticket if needed.
- Else if priority = medium -> Add Gmail label "Needs human" + Create Asana/Ticket.
- Else (low) -> Use ChatGPT’s suggested_reply to send an automated Gmail reply (toggle “From support”); add label "Auto-replied" and log row in Google Sheets.

2) Zap B — Intercom auto-triage
- Trigger: New Conversation in Intercom.
- Action: Send conversation text + metadata to ChatGPT for classification and reply draft.
- Action: Use Intercom API via Zapier to add tags (type:billing/bug/feature, priority:P0/P1/P2) and to post the drafted reply when priority=low.
- If priority >= medium -> Assign conversation to agent/team via Intercom assignment action and post a Slack alert (include suggested reply and confidence).

ChatGPT prompt (compact):
- System: "You are a support triage assistant. Output a JSON object with: tags (array), priority (high/medium/low), suggested_reply (1–2 sentences), confidence (0–1), escalate_reason (optional)."
- User: include customer message, subject, product context, last 3 messages.

Decision criteria / tradeoffs
- Budget: Zapier multi-step + Paths requires paid plan; expect API costs for ChatGPT/OpenAI tokens. If budget is tight, start with classification-only zaps and manual reply.
- Skill level: Intermediate (Zapier + prompt engineering). Use templates and Zapier community zaps to accelerate.
- Team size & stage: Small teams (1–5) should auto-reply only low-priority. Larger teams can push more automated replies.
- Output quality: Improve by iterating prompts, logging false positives, and adding guardrails (confidence threshold).

Practical checklist before going live
- Define priority rules and sample tags.
- Build and test ChatGPT prompt; require JSON output and validate in Zapier.
- Set confidence threshold (e.g., 80% in live testing. Use Zapier Path branching and ChatGPT JSON output to keep the workflow robust.

Compare Zapier and Make