SirenScale: knowing when the agent should hand over.

Built, not launched

Our role: Designed and built by our team as a product. It was not commercially launched, and we have since retired it.
Why we show it: It's an honest sample of how we design AI agents that know their limits.

A WhatsApp-first assistant that answers sales and booking questions from business context, and hands the conversation to a person when it isn't confident enough to answer.

WhatsApp message
Retrieval
Confidence score
Reply, or hand to a person

What we built

Conversation history that persists across messages and sessions.
Retrieval over business context, so replies come from what the business actually offers.
Confidence scoring on each reply, with a threshold that triggers escalation.
A handoff path that notifies staff and passes on the conversation context.

What we learned

[One paragraph from the engineers: what they would change in the architecture, and what the market taught them]

How this applies to you

If you're building an AI agent, escalation rules and evaluation matter more than the prompt. We can help you design both.

All code is private. We offer code walkthroughs under mutual NDA.

See the architecture audit