Short answer
Use a Zap triggered by a new lead (HubSpot form, webhook or new contact), send a tightly-structured prompt to ChatGPT that returns a small JSON payload (intent + confidence + tags), parse that JSON in Zapier, then update the HubSpot contact with tags. Add a predictable fallback path that flags low-confidence or API errors for manual review.

Recommended step-by-step Zap (practical)
1) Trigger: New lead in HubSpot (or webhook/form). Pull name, email, message, page, and any metadata.
2) Action — Formatter (Text): sanitize and truncate the lead message to ~500 characters (reduce token use).
3) Action — ChatGPT / OpenAI: call the model with a deterministic prompt. Use temperature=0 and a low max_tokens for short JSON output. Provide 2–3 few-shot examples and instruct: “Reply ONLY in JSON: {"intent":..., "confidence":0.0, "tags":[...] }”.
4) Action — Formatter / Code by Zapier (or JSON Parse): parse the response into fields intent, confidence, tags.
5) Action — Filter/Paths: if confidence >= threshold (e.g., 0.75), proceed to HubSpot update; otherwise go to the review path.
6a) Action (high confidence): HubSpot — Update Contact (append tags or set a custom property). Use contact ID for idempotent updates.
6b) Action (low confidence): HubSpot — Update Contact (set property = needs_review) + Slack or email notify a reviewer.
7) Error handling path: on API/network error create a log row in Google Sheets and send urgent Slack notify. Optionally implement a retry step with delay.

Prompt template (compact, deterministic)
System: You are a lead-intent classifier. Follow instructions exactly. User: "Classify this lead. Return only JSON with keys: intent (one of: buying, trial, researching, support, other), confidence (0-1), tags (array). Examples: [give 2 short examples]." Then include the sanitized lead text.

Cost per call (how to control and estimate)
- To estimate cost: tokens = prompt tokens + response tokens. Multiply tokens/1000 by model rate. In practice keep prompts small (≤300 tokens) and response target = 60–150 tokens.
- For low-cost high-volume: use a cheaper model (gpt-3.5 family). For higher accuracy/complex language use gpt-4 variant but expect 5–10x cost.
- To minimize cost: truncate text, use few-shot minimal examples, force short JSON responses, and cache repeated queries.

Error handling and reliability checklist
- Use temperature=0 and exact JSON instruction for easy parsing.
- Parse with JSON step; if parse fails, route to manual review and log the raw output.
- Implement Zapier Paths: success -> update HubSpot; low confidence or parse fail -> needs_review + notify.
- Add retry/backoff on transient errors, log all failures to Google Sheets, monitor Zap history and set SLA alerts.
- Make HubSpot updates idempotent (use contact ID), and batch/tag updates to avoid race conditions.

Decision criteria (pick model & thresholds)
- If budget is tight and volume high: gpt-3.5, threshold 0.6, heavier manual sampling. Best-for automation pilots and small teams.
- If you need high accuracy and can absorb cost: gpt-4 variant, threshold 0.75–0.85, smaller manual review queue. Best-for revenue teams and high-value leads.

Best-for / Avoid-if
- Best for: automating initial triage, enrichment at scale, flagging hot leads.
- Avoid if: you need legally auditable decisions or 100% accuracy without human review — keep humans in loop for edge cases.

Quick checklist before you turn on Zap
- Build deterministic JSON prompt + examples
- Set temperature=0 and small max_tokens
- Add Formatter truncate step
- Configure parse + confidence threshold path
- Add logging (Sheets), notify (Slack), and retry logic
- Monitor cost (track tokens or usage report) and sample-check predictions weekly

If you want, I can draft the exact prompt + two few-shot examples and the JSON parsing expressions for Zapier’s Formatter/Code step.

Compare Zapier and Make