Google Sheets is a shockingly capable app backend. Pair it with Apps Script, the Twilio WhatsApp Business API, and OpenAI function calling, and you get an autonomous CRM that reads and writes your data over chat — for about $100/month instead of $600/seat. Here's the architecture, the tool design, and the one trick that cuts the AI bill 5–6×.
So we built the opposite: a WhatsApp AI CRM that lives entirely inside Google Sheets, powered by Google Apps Script, the Twilio WhatsApp Business API, and OpenAI's GPT-4. Total cost: around $100/month for 2,000–3,000 conversations. No SaaS subscription. No database migration. Your sales data stays in the spreadsheet your team already knows.
Here's the part most tutorials get wrong.
A chatbot answers. An AI agent acts.
The unlock isn't GPT generating nice text — it's OpenAI function calling (tools). When a customer says "can I book a demo Thursday?", GPT-4 doesn't just reply politely. It recognizes the intent and calls a function inside your Apps Script: it checks the calendar, writes the appointment to a Sheet, and confirms — in seconds.
That single capability turns a spreadsheet into an autonomous CRM.
The architecture that keeps it from blowing up
Naively wiring GPT straight to your data is how you get duplicate charges and corrupted rows. The pattern that works is a layered flow:
Customer → Twilio (WhatsApp API) → Apps Script doPost (webhook)
→ Context loader (reads Customers + Conversations + Orders)
→ AI brain (GPT-4 + tool library)
→ Action executor (the ONLY layer with write access)
The webhook returns 200 fast to dodge the timeout. The AI decides but never writes directly. A single action-executor layer handles all writes — with idempotency keys and an audit log — so the AI can't accidentally refund the same order twice.
The tool library is where the magic hides
You give the model three tiers of tools:
-
Read tools (safe, free to call):
get_customer,get_orders,get_inventory,search_faq -
Write tools (logged, idempotent):
create_lead,update_status,book_appointment -
Escalation tools (human-gated):
request_human,process_refund— never auto-executed
One hard-won lesson: the description field on each tool schema matters more than the code. It's what teaches the model when to reach for each function.
The trick that cuts the AI bill 5–6×
Running every message through GPT-4 is what makes people say "AI is too expensive." The fix is model routing: a tiny GPT-4o-mini classifier (~$0.0001/message) sorts each message into FAQ / order_status / new_lead / complaint / escalation. The cheap 80% never touches GPT-4.
Result: 3,000 messages/month drops from ~$90–150 to ~$15–25.
What's in the full guide
The complete build on the MageSheet blog covers the parts I couldn't fit here: the stateless-memory problem (and the conversation-summarization fix), per-customer profile memory, the real cost breakdown vs. HubSpot, data-compliance notes (SOC 2 / GDPR), and where Apps Script's ~50–100 concurrent-conversation ceiling forces a move to Cloud Run.
Read the full guide with the complete architecture and tool schemas here:
👉 https://magesheet.com/blog/whatsapp-ai-crm-google-sheets
If you've built something similar in Sheets, I'd love to hear how you handled memory — drop a comment.
Top comments (0)