这一页目前只有英文版。
Context
A sportswear manufacturer with thousands of club customers gets the same questions all day: where is my order, how much would a kit cost, can I talk to someone about sponsorship. Sales and support were answering by hand across email, phone and web chat, and leads were being lost between systems.
Constraints
- Answers had to be true. Order status comes from operational systems, not from the model's imagination.
- Existing tools stay. The CRM, the order systems and the email stack were not going to change; the agents had to fit around them.
- No dedicated platform team. Integrations had to be inspectable and editable by non-engineers after I moved on to the next thing.
My role
I designed the conversation flows, built the agents, and built and operated the integration layer between them and the business systems. I also owned the order-form front end whose backend is entirely this integration layer.
Architecture
- Chat: Voiceflow agents on the public sites with a knowledge base, price-list grounding and hand-off to humans. 40K+ messages, 15K+ users.
- Voice: Retell agents answering inbound sales lines. On call end, a webhook delivers transcript and summary; a workflow classifies intent, creates or updates the CRM lead, and routes it to the right salesperson with the summary attached.
- Integration layer: 50+ n8n workflows acting as the API surface for the agents and the web apps. Order-status lookups, sales-call booking, estimate generation, approval notifications and CRM synchronisation are all workflows with webhook triggers.
- Webhook-only backend: the multi-step kit order form is a React app whose every server interaction is a signed webhook into this layer. There is no bespoke API server to maintain.
Three decisions
1. Ground every factual answer in a lookup. The chat agent never answers "where is my order" from memory; it calls a workflow that queries the operational data and returns a templated answer. Hallucinated order statuses went from the main risk to a non-issue.
2. Put the integration logic where the business can see it. Workflows in n8n are visual, versioned and editable by the operations team. A conventional service would have been tidier for me and a black box for them.
3. Route by intent, not by queue. Voice calls are classified (new lead, existing order, sponsorship, other) before they reach a human. Each intent has an owner and a response-time expectation, and the CRM record carries the summary, so the first human touch is informed.
Outcome
- 40K+ chat messages and 15K+ users served without a support hire.
- Every inbound sales call becomes a CRM record with transcript, summary and an owner within seconds of hang-up.
- 50+ production workflows carrying order status, bookings and approvals across systems.
What I'd do differently
Add evaluation earlier. I added a weekly sample review of agent transcripts after launch; starting with a scored sample from week one would have caught two bad flows sooner. I would also centralise webhook signature verification and retries in one library instead of per-workflow (that library is now an open-source project).
Stack
Voiceflow, Retell, Zoho CRM, n8n, Node.js, Stripe, Resend, React.