Case study / E-commerce / AI Agents for Ops & Support

Two thirds of support tickets resolved instantly, with a higher satisfaction score.

Order status, returns, and product questions were burying a small support team. An AI agent connected to the store, the help desk, and the carrier now resolves 68% of tickets on first contact and passes the rest to people with full context.

Industry: E-commerce
Client: Direct-to-consumer homewares brand
Scale: ~3,200 tickets per month · 5 support agents · 2 warehouses
Timeline: 5 weeks from discovery to full coverage
Solution: AI Agents for Ops & Support

9 hrs → instant

first response

68%

tickets resolved without an agent

4.2 → 4.7

CSAT score

−54%

cost per ticket

01 / The challenge

Where the time was going.

Sixty percent of tickets were some version of “where is my order?” or “how do I return this?”. Agents answered them by opening three systems, copying tracking links, and typing the same paragraphs. Response times slipped past nine hours in peak season and reviews started mentioning support.

01Median first response of 9 hours, rising to 2 days during sales
02Agents context-switching between the store, the help desk, and two carrier portals
03Returns initiated inconsistently, with policy applied differently by different agents
04Product questions answered from memory rather than the current catalogue
05Wholesale and VIP customers waiting in the same queue as everyone else

02 / The solution

How the workflow runs.

We built a support agent that can actually do things, not just answer. It looks up orders, starts returns within policy, answers product questions from live catalogue data, and escalates with a summary when a person is needed.

I
Identify and verify
The agent matches the customer by email or order number and verifies with a second detail before disclosing anything about the order.
II
Answer with live data
Order status, carrier tracking, and estimated delivery come straight from Shopify and the carrier APIs, with proactive explanations for known delays.
III
Act within policy
Returns and exchanges are initiated automatically when they fall inside the policy window and condition rules. A label is generated and the refund path is explained.
IV
Answer product questions accurately
Dimensions, materials, care, and compatibility come from the catalogue and approved help articles, never from general knowledge.
V
Escalate with context
Anything outside policy, any sign of frustration, and all wholesale or VIP accounts are routed to a named agent with a summary and suggested reply.

Human-led by design

Guardrails that kept people in control

Refunds above £150 or outside the policy window always require agent approval.
The agent cannot change shipping addresses after dispatch or alter payment details.
Sentiment detection routes upset customers to a person immediately, no matter the topic.
Every automated reply is tagged in the help desk so quality can be sampled and reported.

Stack

Shopify Admin API
Gorgias
Anthropic Claude
Royal Mail & DPD tracking APIs
Klaviyo
Metabase

Timeline

Week 1
Discovery
Clustered 6 months of tickets into intents, documented the returns policy as rules, and set escalation criteria.
Weeks 2–3
Build
Connected store, carriers, and help desk; agent drafted replies for human approval to measure accuracy.
Week 4
Pilot
Auto-resolved order-status tickets only, then added returns after a review of 300 conversations.
Week 5
Full coverage
Product questions and exchanges enabled; weekly quality sampling continues.

03 / The results

Before and after.

Measured over the 90 days after go-live against the 90 days before. Figures are rounded and client details anonymised.

Measure	Before	After
First response time	9 hours	Instant
Tickets resolved without an agent	0%	68%
Average resolution time	1.8 days	3.5 hours
CSAT	4.2 / 5	4.7 / 5
Cost per ticket	£3.90	£1.80

What the team noticed

Agents now spend most of their day on wholesale accounts, complex complaints, and proactive outreach.
Peak-season temp hiring was not needed for the first time in three years.
Product and fulfilment teams receive a weekly digest of the questions customers actually ask.

What comes next