Recommendation
Use a Zap triggered by HubSpot (new contact or form submission) → send lead fields to ChatGPT with a strict scoring rubric that returns machine-readable JSON → validate/clamp the score in Zapier → update the HubSpot contact property. This gives real-time scores you can act on, while keeping parsing and errors simple.

Decision criteria (choose approach)
- Run real-time per submission if your team needs immediate routing and budget allows (OpenAI tokens cost per call). Good for small teams that act quickly.
- Batch-score (daily/weekly) if you want to conserve budget or re-score enriched leads after enrichment. Better for larger datasets or exploratory models.
- Use high-trust scoring (detailed rubric + JSON output) when you need reliable, auditable scores; use lightweight heuristics (Zapier Formatter) if latency/cost is critical.

Best-for / Avoid-if
- Best for: small marketing teams that need an explainable numeric lead score and automatic HubSpot updates.
- Avoid if: you must guarantee zero Hallucination (use deterministic business rules) or you cannot accept per-call costs.

Practical checklist (Zap setup)
1) Trigger: HubSpot — New Contact or Form Submission (map fields like job title, company size, industry, email domain, engagement metrics).
2) Optional: Formatter/Filter — drop leads missing required fields (email, company name) to avoid bad prompts.
3) Action: OpenAI / ChatGPT action in Zapier — send a system message (instructions) and a user message (lead data + rubric). Ask for JSON-only output.
4) Action: Code by Zapier (JS/Python) or Formatter — validate response, extract numeric score, clamp to 0–100, and produce fallback if parsing fails.
5) Action: HubSpot — Update Contact property with numeric score and optionally update a text property with reasons.
6) Add: Error path — send failed calls to Slack/email and increment a retry counter or route to manual review.

Prompt templates (use exact JSON output)
System: "You are an expert B2B sales lead scorer. Always return ONLY JSON with keys: score (0-100), reasons (array of short strings), category (hot/warm/cold). Do not include extra text."
User: "Lead data: {name}, title: {title}, company_size: {company_size}, industry: {industry}, email: {email}, recent_engagement: {engagement_details}. Scoring rules: +30 for decision-maker titles, +25 for company_size >=50, +20 for matching target industry, +15 for recent high-engagement actions, -10 for free email domains. Use these to produce score 0-100. Output only JSON."
Example expected output:
{"score":78,"reasons":["VP title","SaaS industry match","recent demo request"],"category":"hot"}

Error handling tips
- Require strict JSON output in the system prompt; in Zapier, treat any non-JSON as a parsing error.
- Validate: if score missing or not numeric, fallback to default score (e.g., 0) and tag contact for manual review.
- Retries: implement exponential backoff for transient API errors; log requests and responses for debugging.
- Rate limits: batch requests during busy times or use lower model tier to reduce cost.
- Sanitize inputs: strip long free-text fields to avoid prompt bloat; limit total token usage.

Implementation notes
- Use ChatGPT/OpenAI action in Zapier for the model call (maps well to this workflow).
- Map the validated numeric field to a HubSpot custom property (integer). Also write reasons to a separate text property for auditability.

If you want, I can draft the exact Zapier OpenAI action payload and a Code by Zapier validator snippet to paste into your Zap.

Compare Zapier and Make