Customer Support Copilot

The problem

Support volume scales with sales, but headcount can't

A large share of tickets are the same questions — where's my order, returns, sizing — yet each still waits in a queue. Agents burn out on repetition while genuinely hard cases wait behind them.

This copilot answers the routine instantly with order context, and drafts the rest for a human to approve — so the team spends its time where judgement is needed.

Primary objective

Deflect repetitive tickets and cut agent handle time while keeping humans on anything that matters.

How it works

Ticket in → retrieve context → draft reply → human on edge cases

Ticket in

Email, chat, social

→

Retrieve

Order, policy & FAQ context

→

Draft reply

Grounded, on-brand answer

→

Human gate

Approve sensitive / refunds

What it does

✓ Order-status answers
✓ Returns & policy handling
✓ RAG over help center
✓ Draft-and-approve mode
✓ Multilingual replies
✓ Macro suggestions

Built with

✓ Claude + RAG grounding
✓ Helpdesk (Zendesk/Gorgias) API
✓ Order / OMS integration
✓ Confidence-based routing
✓ Agent approval UI

Guardrails

✓ No refunds without approval
✓ Answers cite a source
✓ Escalation on low confidence
✓ PII handling controls
✓ Full transcript audit

Under the hood

Confidence-gated auto-reply

Only high-confidence, low-risk answers can auto-send; anything about money or with weak grounding is drafted for an agent.

support/route.py PYTHON

def handle(ticket, answer):
    # answer = {text, confidence, cites, touches_money}
    if answer["touches_money"] or not answer["cites"]:
        return queue_for_agent(ticket, answer)     # human decides
    if answer["confidence"] >= 0.85:
        return auto_send(ticket, answer["text"])
    return queue_for_agent(ticket, answer)         # draft, don't send

Outcomes

~40%

Tickets deflected

Faster

First response

24/7

Instant routine answers

Human

On every refund

Cut response times without cutting quality

Pairs with Review Insights to keep your FAQ answers current.

Start a conversation →