Skip to content
All case studies
Case study
E-commerce & retailAI customer service

Answers in minutes, at any hour

We build an assistant that resolves routine requests around the clock and helps your team with the rest.

Our plan for an online store with 6,000 conversations a month

from kickoff to first value
5–7 weeks
estimated payback*
Month 2
of team capacity a year, estimated*
$90k

* Estimates from the illustrative maths in chapter 07; we confirm your real figures in a discovery call.

Industry
E-commerce & retail
Solution
AI customer service
What we build
AI assistant
People in control
Every decision
First value
5–7 weeks
Payback
Month 2*

The problem

The same questions, all day

Where is my order, how do I return this, can I change my booking: most requests repeat. Customers wait in queues, your team answers the same thing a hundred times, and growth means hiring more people for the same work.

  • Long queues at busy times
  • Most tickets ask the same questions
  • No answers after hours
  • New staff take months to get fast
Two colleagues looking at a laptop in a bright office

Why we know it works

The approach is proven at scale

We build on methods that have already worked for large companies, and size them for a growing business. The evidence we draw on, as reported by Klarna:

Klarna, and a Stanford–MIT studyCompany-reported and peer-reviewed

In its first month, Klarna’s AI assistant handled 2.3 million conversations, two-thirds of its customer chats, with satisfaction on par with human agents and resolution times down from 11 minutes to under 2. A peer-reviewed study of 5,172 support agents found that an AI copilot made agents 14% more productive on average and 34% more productive for new and less-experienced staff.

of chats handled by Klarna’s assistant
⅔
to resolve, down from 11 minutes
<0 min
productivity for newer agents (peer-reviewed)
+0%

Klarna’s figures are its own (2024) and some were questioned by the press; the study is Brynjolfsson, Li & Raymond, Quarterly Journal of Economics (2025).

See it work

A conversation, start to finish

An illustrative chat with an assistant we build. It answers from your order data, and hands over to a person, with the whole conversation, when it should.

Illustrative demo

Support assistant

Online · answers from your own data

Our plan · agent hours a month*

Today600 h
With the assistant300 h

50%*

of conversations resolved by the assistant, our planning assumption for 6,000 a month.

24/7

answers, with a person one message away and the conversation already written up.

* Illustrative estimates; see the savings chapter.

How we solve it

Your data, an AI that understands it, and people in control

The same idea, sized for a growing business: we connect to what you already have, put the AI in the middle, and keep a person in charge of every decision that matters.

  1. 01

    Your data

    • Help-centre articles
    • Order and account data
    • Past conversations
    • Policies
  2. 02

    AI assistant

    A language model grounded in your own knowledge and data: it answers, tracks orders and starts returns, and never invents a policy.

  3. 03

    Your workflow

    On your website chat, email and messaging; anything it can’t resolve goes to a person with the conversation so far.

  4. 04

    Your people

    Agents handle the complex cases, with AI-suggested replies to speed them up.

How we build it

First value in 5–7 weeks

One problem, a fixed scope and a clear plan. You see it working on your own data before it goes live.

  1. DiscoverWeek 1

    We sort three months of your conversations into what an assistant can resolve and what needs a person.

  2. Ground itWeeks 2–3

    We connect your help articles, policies and order data, so every answer comes from your own sources.

  3. Test itWeeks 3–5

    We replay real past questions and check every answer before a customer sees it.

  4. Go liveWeek 6

    We start on one channel, with a clear hand-over to your team.

  5. ImproveWeeks 7–10

    We turn missed questions into new answers every week.

When it hands over

People stay in charge of what matters

The assistant is set up to know its limits. These go to your team, with the context already written up.

  • Anything unusual

    Requests outside your policies, or that it can’t answer from your own sources.

  • Money and exceptions

    Refunds above a limit you set, disputes and complaints.

  • When a customer asks

    A person is always one message away.

Where the savings come from

The maths, line by line

An illustrative example with stated assumptions, so you can put in your own numbers. We confirm the real figures together in a discovery call.

$90k

a year of capacity, and answers at 2 a.m.

An online store with 6,000 conversations a month

Illustrative · assumptions
Minutes per conversation
6
Agent hours a month
600 h
Resolved by the assistant (assumption)
50%
Hours freed a month
300 h
Cost per hour
$25
$90ka year of capacity, and answers at 2 a.m.

The J-curve

A small dip, then compounding growth

Research on new technologies finds the same pattern: value dips while you invest in data, process and training, then climbs well past where you started. We keep the dip small by starting with one problem and paying back fast.

Invest and learnPayback · month 2Compounding returnsBreak-even
Month 0Month 12

* Shape after Brynjolfsson, Rock & Syverson (2021); the payback month is illustrative.

What comes next

Where this is heading

  1. 01

    Assistants that act

    Changing an address, booking a collection or issuing a refund, within limits you set.

  2. 02

    Voice

    The same assistant on the phone, in natural speech, for customers who call.

  3. 03

    Problems solved before they’re asked

    Customers told about a delay before they notice it.

Facing the same problem?

A free consultation maps where this would pay back fastest in your business.

Sources

  1. 1.tech.eu: Klarna’s AI assistant (2024)
  2. 2.Chain Store Age: Klarna resolution times
  3. 3.Fast Company: questions about Klarna’s figures
  4. 4.NBER / QJE: Generative AI at Work
  5. 5.Brynjolfsson, Rock & Syverson, “The Productivity J-Curve”, American Economic Journal: Macroeconomics (2021)

Next case study · Healthcare

Clinicians back with patients, not keyboards

Start a conversation

Let's build what's next for your business.

Tell us what's slowing you down. We'll show you where AI, automation, software and data will make the biggest difference, and what to do first.

  • Free consultation
  • Your workflows reviewed by a specialist
  • Clear next steps, no commitment

Working across time zones · Live

  • ----San Francisco 
  • ----New York 
  • ----London 
  • ----Dubai 
  • ----Sydney