Halving a support queue with Claude — safely
AI support automation that customers rate 4.6/5 — because the boring engineering came first.
TradeNest's ticket volume doubled every two quarters; the support team couldn't. Off-the-shelf AI chatbots had been tried and rolled back — they hallucinated order details and enraged exactly the customers who mattered most.
We built retrieval before intelligence: a pipeline grounding every draft in the customer's actual orders, the help centre and policy documents. Claude classifies each ticket, drafts a grounded reply, and a confidence gate decides — auto-send, human review, or straight to a specialist with full context attached.
Every automated reply is logged, scored and sampled weekly. The safe half of the queue vanished; the hard half now arrives pre-investigated.
How we got there
Ticket archaeology
Six months of history clustered and labelled to find what was truly automatable.
Grounding pipeline
RAG over orders, policies and help content — with citations in every draft.
Guardrails & evals
Confidence gating, red-team suite and a weekly human evaluation loop.
Gradual rollout
Shadow mode → 10% → 50% of eligible tickets, expanding with the scores.
What changed
for the client
More case studies
How Precision MFG cut order processing time by 3.2×
A 400-person plant running on spreadsheets became a real-time operation — without stopping the line for a single day.
Read case study LogisticsFrom internal tool to 40-customer SaaS in 14 weeks
A dispatch tool commissioned for one courier company became FleetPilot — a multi-tenant SaaS now moving 2.1M parcels a month.
Read case studyLet's build something exceptional together
Tell us about your project and get a detailed proposal — including scope, timeline and investment — within 24 hours.