Inbound payment calls look routine,

Inbound payment calls look routine, but every one carries risk: the caller has to be verified, a legal disclosure has to be read, and any offer has to follow policy. I designed and built an AI phone agent that handles the whole call, with a human one sentence away.

What I designed:

• A 30-node conversation flow: look up the caller by phone, confirm identity, read and log the required disclosure, then negotiate and take payment by keypad.

• Guardrails that fire from any step. Attorney mentions, cease and desist requests, disputes and requests for a person each route to the right place and get logged.

• A six-rung negotiation ladder: full balance, six payments, 80% settlement, 70% final offer, hardship plan, final probe. Plan math always sums to the balance, restated offers are paraphrased, one offer per turn.

• Six agents that pass callers to each other during a call: English and Spanish inbound, declined-payment recovery in both languages, callbacks and identity verification.

• An operator dashboard for live calls, outcomes, results by agent and compliance checks.

How changes ship: every change is cloned into a sandbox, run through scripted test conversations that assert which exit fired, checked against named rules on the live flow, then published.

Stack: Retell conversation flows, Python, Next.js, Supabase, Recharts. Dashboard screenshots use demo data.