Examples of work we've built.

Trigger: New quote sent (Xero / QuickBooks webhook)
   ↓
Wait 24 hours
   ↓
Check: Has client opened/replied? (Email-tracking webhook)
   ↓ NO
Send personalised follow-up email
   ↓
Wait 48 hours
   ↓
Check: Still no response?
   ↓ YES
Send second email + log task in CRM
   ↓
Wait 4 days
   ↓
Final follow-up via SMS (Twilio)
   ↓
Mark in CRM as "needs phone call"
   ↓
Notify owner via Slack

You are the assistant for [Practice Name], a physio clinic in [Suburb].

# What you can answer
- Services offered (sports physio, women's health, dry needling, hydrotherapy)
- Fees: $95 first consult, $75 follow-up, $60 hydro group session
- Bulk billing: only for chronic-disease management plan referrals
- Locations: [Address 1], [Address 2]
- Hours: Mon-Fri 7am–7pm, Sat 8am–1pm

# What to do for booking requests
Direct visitors to the online booking page: [URL]
Don't take bookings yourself — the booking system has more accurate
real-time availability than you do.

# What to escalate to reception
- Anything medical-advice-shaped ("does this sound like a torn meniscus")
- Insurance queries (we don't bulk-bill private health on the spot)
- Anything outside your knowledge

# Tone
Warm, plainspoken Australian English. Brief. Don't say "I'm an AI".
If pressed, say "I'm a digital assistant for the practice".

1. Client photographs receipt → uploads via Dext mobile app
2. Dext OCR extracts: vendor, date, total, GST amount
3. Make.com webhook receives the parsed data
4. Lookup: which client account does this belong to?
5. AI step (Gemini): based on vendor + line items, predict the
   correct chart-of-accounts code using the client's prior
   coding history. Returns confidence score.
6. If confidence ≥ 0.85: post to Xero as a draft, ready for review
   If confidence < 0.85: post to Xero AND flag for human review

Result: bookkeeper review time per receipt drops from ~90 sec
to ~10 sec.

// app/api/lead-route/route.ts
import { GoogleGenerativeAI } from "@google/generative-ai";

export async function POST(req: Request) {
  const { name, email, message, listingId } = await req.json();

  // Use a small model to classify intent
  const ai = new GoogleGenerativeAI(process.env.GEMINI_API_KEY!);
  const model = ai.getGenerativeModel({ model: "gemini-2.0-flash" });

  const { response } = await model.generateContent(
    `Classify this property enquiry into: BUYER_HOT, BUYER_BROWSING,
TENANT, or AGENT. Reply with only the category.

Message: ${message}`
  );
  const intent = response.text().trim();

  // Route based on intent
  if (intent === "BUYER_HOT") {
    await notifyAgentImmediately({ listingId, name, email, message });
  } else {
    await addToCRMNurtureFlow({ listingId, name, email, message, intent });
  }

  return Response.json({ ok: true });
}