Workflow: Zapier + ChatGPT for automatic lead enrichment

Recommendation (short):
Build a Zap that triggers on new lead rows (form, ad, or CRM), calls ChatGPT via a Webhook action with a tightly constrained JSON schema prompt, and writes parsed fields back to your CRM. Batch or queue calls for volume to save tasks and control costs.

Concrete Zap pattern (step-by-step):
1) Trigger: New Lead — e.g., “New Form Submission” (Typeform/HubSpot/Google Forms).
2) Action: Formatter (Zapier) — normalize fields (company name, website, lead text, title, source). Build a single prompt payload. 1 Zap task.
3) Action: Webhooks by Zapier — POST to ChatGPT/OpenAI API with model + system prompt requesting JSON output. 1 Zap task.
4) Action: Formatter / Code — parse JSON response into individual fields (company_enrichment, persona, ideal_message_subject, 2-line_message). 1 Zap task.
5) Action: Update CRM — write mapped fields to HubSpot/Salesforce/Close. 1 Zap task.
6) (Optional) Action: Slack / Task — route low-confidence leads to SDRs.

Example prompt instructions to include in webhook body (short):
- System: “You are a lead enrichment assistant. Return ONLY valid JSON with keys: company_data, persona, ICP_flags, outreach_subject, outreach_body, confidence_score.”
- User: include lead inputs (company name, website, job title, form answers). Ask for concise output and examples.
This forces deterministic parsing and reduces downstream errors.

Parsing schema (mandatory):
Return JSON keys with simple scalars or arrays (no explanatory text). Example mapping: company_data.name, company_data.domain, persona.title_level, persona.pain_points[], outreach_subject, outreach_body, confidence_score (0-1).

Rate limits & task accounting:
- Zapier: each action counts as a task. For one-lead enrichment expect ~4 tasks (trigger + webhook + parse + CRM update). Use Formatter consolidation to keep actions minimal.
- OpenAI/ChatGPT: API rate limits depend on model/account. Design for ~1 request/second or lower per API key if volume is high; use batching or multiple API keys if you need parallelism. Implement retries with exponential backoff.

Cost estimate method (quick):
Cost per lead = (tokens_used / 1000) * model_cost_per_1k. Tokens include input + output. Example: if ~500 tokens total and model cost is $0.03 per 1k tokens, cost per lead ≈ $0.015. Multiply by monthly lead volume to estimate spend. Choose cheaper models (3.5-style) for high volume, higher-tier models for richer outputs.

Decision criteria (pick strategy):
- Low volume (50k/mo): batch requests, limit fields, or use a multi-key parallel strategy with queuing.

Best-for / Avoid-if:
- Best for: teams that need consistent structured enrichment and can accept a short structured prompt-output cycle.
- Avoid if: legal/PII-sensitive leads unless you have appropriate data controls and model choices.

Practical checklist before go-live:
- Define JSON schema and test with 20 sample leads.
- Harden prompts to return only JSON.
- Add zap error handling + retry/backoff.
- Cap daily spend and monitor tokens per response.
- Add a human-review queue for <0.6 confidence_score.

If you want, I can draft the exact Webhooks POST body + JSON schema and a 3-sentence system prompt you can paste into Zapier’s Webhooks action.

Compare Zapier and Make