Recommendation (short):
Build a two‑Zap pattern: 1) capture form responses into a durable queue (Google Sheets or Zapier Storage), and 2) scheduled Zap that batches N rows, sends a single ChatGPT/OpenAI request returning compact JSON scores, and updates Salesforce. This minimizes API calls, supports retries, and keeps costs down.

Why this works
- Batching reduces per‑lead token overhead and number of API calls.
- A scheduled Zap gives predictable cost and easy error handling.
- JSON-only responses keep parsing deterministic for Salesforce updates.

Decision criteria (choose by need):
- Low latency / real‑time scoring: call ChatGPT per submission (higher cost, simpler).
- Cost-sensitive / high volume: batch (e.g., 10–50 leads/req).
- High accuracy / complex reasoning: use a stronger model and shorter batches.
- Small dev team: use Zapier’s OpenAI app or Webhooks instead of building custom services.

Concrete Zap flow (step‑by‑step checklist)
1) Trigger Zap A: Form submitted (Typeform/Webflow/Formstack).
2) Action: Append row to Google Sheets or write to Zapier Storage with a unique ID and timestamp.
3) Zap B (Schedule by Zapier): run every X minutes (X = tradeoff between latency and cost).
- Lookup rows where scored=false and created 3, push to a dead‑letter sheet and alert the team.
- If parsing fails, archive raw model output for debugging and retry with a smaller batch.

Cost minimization tips
- Batch N items per request. Experiment to find the N that balances tokens and latency.
- Trim prompts to only necessary fields. Use concise system instructions and ask for compact output.
- Use a lighter model for straightforward scoring; upgrade model only if you need deeper reasoning.

Best‑for / Avoid‑if
- Best for: mid/high volume lead flows that can accept minutes of latency and need predictable cost.
- Avoid if: you need instant real‑time scoring or have very strict SLA (then do per‑lead calls or a dedicated scoring microservice).

If you want, I can: (a) draft the exact Zapier prompt + example JSON; or (b) give a sample Google Sheets schema and a Code by Zapier snippet to validate model output.

Compare Zapier and Make