Support Agent | Delivered-Not-Received Workflow
I built this so the LLM has one job: turn a messy customer message into structured case data. The rest of the workflow is explicit.
When a customer says a package was marked delivered but never arrived, the system retrieves the evidence, applies policy, checks authority, and either issues an idempotent refund or routes the case to a human.
A missing-package claim can end in a refund. The system needs to make a good decision without giving the model uncontrolled access to money or customer state.
Customer message → structured case → evidence lookup → policy checks → authorization → refund or human review.
Workflow behavior, authorization, duplicate actions, failure recovery, model extraction, tracing, and a staged rollout plan.
In the synthetic base case, the workflow covers 70% of eligible cases. It frees roughly 1,200 hours of support and review work each month and produces about $46K in monthly net value in the model. These are modeled results. I would validate the assumptions with a real customer before expanding the deployment.
The LLM can classify the case. Policy and authority live elsewhere. Money moves only after the workflow passes those checks.
