Use buffering + batching: collect messages for 30–60s or until ~50 items, dedupe/compact per thread, then send one summarize+action prompt to ChatGPT. Add per-user debounce (30s) and per-channel debounce (60s). Run a persistent queue (Redis/DB) with a single worker that enforces rate limits and exponential backoff with jitter; fallback to cached recent answers or a cheap classifier when queue grows. Cut tokens via incremental summarization and tight system prompts.

If Zapier's execution model is the blocker, compare Zapier vs Make to find stronger queuing options: Compare Zapier and Make