Lead Autopilot

Every patient inquiry answered in under a minute, day or night. Routine questions get a reply right away. Emergencies and anything unclear wait for the dentist's approval.

Lead Autopilot is an inquiry-handling system built for a small dental clinic. When someone submits the clinic's inquiry form, AI tags the message, it's logged to Google Sheets, a drafted reply goes out by email, and the owner gets a Discord alert.

The demo client is Pinecrest Dental, a fictional family clinic in Baguio City. All data is made up.

Live demo: lead-autopilot-six.vercel.app · Public inquiry log: /log

How it works

Inquiry form → AI tags it → Saved to Google Sheet → Routine: reply emailed now
                                                  → Urgent or unclear: held for approval
                                                  → Spam: no reply
                                                 ↘ Owner alerted in Discord (every inquiry)

1. A visitor submits the form. It's checked in the browser and on the server (Zod), behind Cloudflare Turnstile, a honeypot field and a rate limit of 5 per hour.
2. classifyLead() asks the AI for one JSON shape: intent, urgency, confidence, a one-line summary and a drafted reply.
3. The inquiry is saved as a row in Google Sheets.
4. If it's routine, the reply is emailed right away via Resend. If it's an emergency, high urgency, or the AI is less than 70% confident, it waits for approval and nothing is sent. Spam gets no reply.
5. The owner gets a Discord alert for every inquiry.
6. In the dashboard the owner can review, edit and Approve and send held replies.

If the AI step fails, the inquiry is still saved, the owner is still alerted and the patient gets a standard "we got your message" email. No inquiry is lost.

Screenshots

What visitors see after submitting	Owner dashboard

Review and approve a held reply	Public, anonymised inquiry log

Results

• 20/20 on a fixed test set of sample inquiries: routine bookings, pricing questions, emergencies, angry patients, Tagalog messages, vague messages and spam.
• About 3 seconds from submit to the reply being sent, including the AI, the sheet and the alert.
• Replies are in the patient's own language, never promise prices, never give medical advice and never invent clinic facts.
• Minutes saved counter on the dashboard: inquiries handled × 5 minutes.

AI design

• Every AI call goes through src/lib/ai/classify-lead.ts using the AI SDK. The model is one environment variable:

  AI_MODEL=google/gemini-3.5-flash-lite        # free Google AI Studio tier
  AI_MODEL=anthropic/claude-sonnet-5.5         # plus ANTHROPIC_API_KEY, no code change
  

• The output is validated against a Zod schema, with a 15-second timeout.
• The prompt is in its own file, src/lib/ai/prompt.ts.
• npm run test:ai runs the 20 sample inquiries in scripts/samples.ts and prints a scorecard. Rerun it after every prompt change.

Running it locally

cp .env.example .env.local   # fill in the keys below
npm install
npm run dev                  # http://localhost:3000

Variable	What it's for
AI_MODEL, GOOGLE_GENERATIVE_AI_API_KEY	AI tagging and drafted replies (Google AI Studio)
ANTHROPIC_API_KEY	Only if AI_MODEL is an anthropic/* model
GOOGLE_SERVICE_ACCOUNT_EMAIL, GOOGLE_SERVICE_ACCOUNT_PRIVATE_KEY, GOOGLE_SHEET_ID	Inquiry log. Share the sheet with the service account as Editor; the tab and headers are created automatically.
RESEND_API_KEY, EMAIL_FROM, EMAIL_TEST_RECIPIENT	Reply emails. EMAIL_TEST_RECIPIENT sends every patient email to one inbox (demo mode).
OWNER_ALERT_WEBHOOK_URL	Discord webhook for owner alerts
NEXT_PUBLIC_TURNSTILE_SITE_KEY, TURNSTILE_SECRET_KEY	Cloudflare Turnstile spam check
ADMIN_PASSWORD	Owner dashboard login at /dashboard

Useful scripts:

Command	What it does
npm run test:ai	Runs the 20-message AI test set
npm run test:sheets	Checks the Google Sheets connection
npm run typecheck / npm run lint	Static checks

Project structure

src/app/page.tsx                  Clinic homepage and inquiry form
src/app/api/inquiry/route.ts      Spam checks → processInquiry()
src/lib/process-inquiry.ts        The pipeline: AI → Sheet → email → alert
src/lib/ai/                       classifyLead(), prompt
src/lib/sheets.ts                 Google Sheets log
src/lib/email.ts, alerts.ts       Resend replies, Discord alerts
src/app/dashboard/                Owner login, inbox, review and approve
src/app/log/page.tsx              Public read-only log
docs/PLAN.md                      Plan, brand and build phases

Tech stack

Next.js 16 (App Router, Cache Components), TypeScript, Tailwind CSS 4, AI SDK 7 (Gemini, swappable to Claude), Zod, Google Sheets API, Resend, Discord webhooks, Cloudflare Turnstile, Vercel.