Recommendation
Use Zapier to collect and normalize inbound leads, then call ChatGPT to compute a reproducible JSON score + reasons. Keep a lightweight rule-based fallback inside Zapier for missing fields or rate-limit cases so lead flow never stalls.

Why this works
- No servers: all steps run inside Zapier and native ChatGPT/OpenAI integration.
- Structured output: force ChatGPT to return strict JSON so downstream Zaps/CRMs can parse automatically.

Step-by-step Zap (practical)
1) Trigger: New form submission / new inbound email / webhook. Prefer direct form app (Typeform, HubSpot) to reduce parsing work.
2) Preflight normalization (Formatter by Zapier)
- Strip HTML, trim, lower-case email/phone, extract U.S. phone format, remove non-UTF chars.
- If message >2000 chars, create a short summary using Formatter (first 800 chars) and save full text to a storage step.
3) Quick rule-based scoring (Formatter + Lookup Table)
- Apply cheap rules for must-haves: company size, industry match, geography. Produce preliminary points.
- If required fields missing, tag as "low_info" and send to manual queue.
4) ChatGPT action (call LLM)
- Send a small JSON payload: normalized fields + preliminary points + 3 examples (few-shot) and an instruction to return EXACTLY this JSON schema.
- Example system instruction: "You are a lead-scoring assistant. Output only valid JSON with: score (0-100), bucket (hot/warm/cold), top_factors (array of short reasons), action (route/manual_review)."
- Example user prompt (compact):
{"company":"Acme Inc","size":"200","industry":"SaaS","role":"CMO","message":"Looking for pricing" ,"pre_points":25}
+ few-shot examples showing inputs → desired JSON outputs.
5) Parse JSON with Zapier’s utilities and route: CRM update, Slack alert, assign owner.
6) Fallback: If ChatGPT fails or returns invalid JSON, use Zapier’s stored rule-based score and flag the record for review.

Example ChatGPT prompt snippet
System: "You are a lead scoring API. Return ONLY valid JSON. No explanation. Schema: {score:int 0-100, bucket:string, top_factors:array, action:string}."
User: Provide the normalized lead + 2 examples. Then: "Score using weight: intent 40%, fit 35%, timing/urgency 25%. If required fields missing, set action to 'manual_review'."

Edge cases & rate limits
- Missing / malformed inputs: preflight normalization, then set 'low_info' and route to manual review.
- Long messages: summarize before sending to ChatGPT; include link to full text stored in Google Drive/Sheet.
- Gibberish / non-English: use a cheap language-detection step; if non-English, either auto-translate or route to human.
- Rate limits: batch or debounce incoming events (Zapier Delay), call ChatGPT only when needed (e.g., high-pre_points threshold), use exponential backoff and a rule-based fallback when OpenAI returns 429. If volume is high, move scoring to periodic batch runs (every 5–15 min).

Decision criteria (when to use LLM vs rules)
- Use ChatGPT when: you need nuanced intent extraction, variable messaging, small-to-medium volume, and you can tolerate minor costs/latency.
- Use rules when: very high volume, low budget, or strict latency needs.

Practical checklist
- [ ] Map inbound fields and required minimums.
- [ ] Build Formatter clean-up steps (HTML, length, phone/email).
- [ ] Create a lightweight rule-based pre-score.
- [ ] Build ChatGPT prompt with strict JSON schema and 2–3 examples.
- [ ] Add fallback Zap route for rate limits/invalid JSON.
- [ ] Log both LLM output and fallback decisions for audit.

Best-for / Avoid-if
- Best for: non-technical marketers, small-to-medium lead volume, nuanced messages.
- Avoid if: extremely high-volume (100k+/month) or strict PII rules—consider server-side batching.

If you want, I can draft the exact Zap steps with the full prompt + sample JSON few-shots tailored to your form fields.

Compare Zapier and Make