AI Form Lead Extractor for Make with Confidence Routing

Free-text form fields produce garbage: names in lowercase, phone numbers in five formats, “we maybe have budget idk” as a company field. This Make blueprint runs every submission through an AI extraction step that returns clean, typed fields plus a confidence score — and only the confident ones flow into your CRM unattended.

What this scenario does

• Custom webhook receives the raw form submission (works with Typeform, Tally, Jotform, or any form that can POST)
• OpenAI module extracts structured fields: full_name (properly capitalized), email (validated, lowercased), phone (E.164 where inferable), company, intent (demo_request / pricing_question / support / partnership / other), budget_hint, a one-line summary, and an overall confidence from 0 to 1
• Parse JSON makes the result mappable
• Router with two filtered routes:
  • Confidence ≥ 0.8: HubSpot contact created with all fields mapped, then the lead is logged to a Google Sheets “Leads” tab
  • Confidence < 0.8: Slack message to #lead-review showing the extracted fields and the raw submission, so a human fixes it in seconds instead of discovering bad CRM data weeks later

The confidence instruction is the heart of the prompt: the model is told to lower its score whenever fields are ambiguous, contradictory, or missing — which is exactly when you want a human in the loop.

Prerequisites

• Make account
• OpenAI connection (gpt-4o-mini with JSON response format)
• HubSpot connection (or swap module 5 for Pipedrive/Salesforce — the mapping is six fields)
• Google connection and a spreadsheet with a “Leads” tab
• Slack connection for the review channel

How to import

1. Download the blueprint JSON from the download box above.
2. In Make: create a scenario → More (⋯) → Import Blueprint → select the file.
3. Re-link all connection placeholders, create the webhook, and select your spreadsheet.
4. Point your form’s webhook/integration at the Make webhook URL and submit a deliberately messy test entry — all-lowercase name, vague message — to confirm it lands in the Slack review route.

What to customize

• The confidence threshold (0.8 in both route filters) — lower it once you trust the extraction, raise it for regulated industries
• Intent categories to match your sales process stages
• CRM field mapping — add hs_lead_status, owner assignment, or a source property
• The extraction rules — e.g., add country or company_size fields to the JSON schema in the system prompt

This template pairs with the AI form data extraction tutorial, which covers prompt-testing against a sample of your real submissions before going live.

Cost per run

One gpt-4o-mini call per submission, typically 300-800 tokens total — as of mid-2026, far below a cent per lead. Make-side, a confident lead consumes 5 operations and a review-routed one 4, so even 1,000 leads a month stays around 5,000 operations.