Ami qualifies every lead around the clock before a manager sees it
An AI assistant that greets every visitor on Telegram and the site's chat widget, asks the right qualifying questions and hands the warm leads to a manager.
1Sonnet writes the qualifying follow-up live in the conversation
2Visitor types freely, no menu, no forced script
Mid-dialogue: Ami follows a stated pain point with a qualifying question.
1Sonnet writes the qualifying follow-up live in the conversation
2Visitor types freely, no menu, no forced script
Where AI does the work
01 / QUALIFY
Live qualifying conversation
Claude Sonnet writes every qualifying reply live in the conversation, referencing what the lead already said instead of restarting the interview. The same logic runs identically on Telegram and the web widget.
02 / EXTRACT
Background lead extraction
A background Haiku pass rereads recent messages every five messages and pulls out business type, the real problem and tools in use. It only fills fields still empty.
03 / ROUTE
Warm handoff to a manager
Once a lead's score crosses the threshold, Ami requests a handoff: a Telegram alert reaches a manager and an email goes out over Gmail SMTP. Crisis language skips the model entirely.
Built with Claude Code in 5+ months
Before
Every inbound message needed a person to read it and judge what mattered.
After
Ami qualifies each lead in the conversation and pings a manager once it's warm.
[→]/More screens
Widget open: language pick and Ami's opening greeting.
Under the hood
01Context
A flood of messages, most not a real lead
Our own marketing site draws the same kind of inbound we build automations for clients: a chat widget, a Telegram bot and a contact form, all open around the clock. Most of it is not a real lead, a student's question, a competitor checking prices, someone just browsing. Reading every message by hand and deciding what deserves a reply does not scale to a small team also doing client work, so Ami runs the same qualifying logic on both channels.
02How it works
Node and Express, two models split by job
The backend is Node.js and TypeScript on Express, with Supabase Postgres storing every conversation, message, lead and handoff. An incoming message runs through regex and keyword heuristics first, which classify most traffic for free; crisis language is caught immediately and answered with a fixed safety response that skips the model. Claude Sonnet writes the actual conversation; Haiku handles the cheaper utility work, intent classification and the periodic lead extraction. A channel abstraction lets Telegram and the web widget share the same assistant logic without knowing about each other.
03Outcome
In production since July 2026
Ami has been answering on Telegram and anmflow.com around the clock since July 2026. Every conversation is stored, every lead gets a score and every warm one reaches a manager through a Telegram alert and an email, with no dashboard to check to catch a good lead landing at 2 a.m.
ANM SOLUTIONS / CONTACT US
Want something like this?
First 30-minute call is free. We scope it, you decide.
