Customer Refund AI Agent

01

Problem

Refunds are where support automation usually breaks: a pure LLM agent will eventually approve a refund it shouldn't, or cite a policy that doesn't exist. The business needs conversational speed without giving up control of decisions that cost money. This system — built under the internal codename RefundPilot — solves that split.

02

Agent flow

The conversation runs through a LangGraph state machine: understand the request, collect order details, evaluate against policy, then respond or escalate. The graph keeps the agent on rails — it cannot skip the policy check or invent an outcome.

• Intent and order-detail extraction from the conversation
• Policy evaluation as an explicit graph node
• Respond, deny with explanation, or escalate — never improvise

03

Policy decision logic

Eligibility is computed by a deterministic rules engine — return window, item category, refund caps, prior history. The LLM explains the decision in plain language; it never makes the decision. The same conversation always produces the same outcome.

04

Human escalation

Anything outside clear policy — disputed claims, high-value orders, ambiguous evidence — routes to a human queue with the full conversation and the agent's partial assessment attached, so the reviewer starts with context.

05

Admin dashboard and reasoning logs

A Next.js dashboard lists every conversation with its outcome, the policy rules that fired, and the agent's reasoning trail. Support leads can audit any decision after the fact — the system is designed to be questioned.