Skip to guide

Case study guide

AI customer support automation: how companies did it, step by step

Updated 7 Nov 2026

AI customer support automation means an AI agent answers customer questions and does routine tasks such as order status, returns or password resets, and passes the rest to a person. This page walks through three real company examples, step by step.

A customer support agent wearing a headset working at a laptop
76%of help-site questions resolved without a personSalesforce, reported Dec 2024
about 20%true resolution at Kogan.com, up from 6.2%Kogan.com via Salesforce, reported Jul 2026
2.3Mchats in Klarna's first monthKlarna, reported Feb 2024

Quick answers

  • Salesforce Help site: the company reports about 76% of questions resolved without a human, with about 5% handed to a person with a summary.
  • Kogan.com: started with order tracking (about 60% of requests), then added a fixed seven-step return flow. It reports the agent handles about 50% of inquiries.
  • Klarna: the most quoted example, but in 2025 its CEO said cost-cutting had gone too far and the company would make sure customers can reach a person.
  • The pattern: start with the most common question, give the agent the customer's own data, fix the steps where a mistake costs money, and always hand off to a person with a summary.

What the three examples have in common

  • Start with the question asked most. Kogan.com began with "where is my order" because it was about 60% of requests.
  • Give the agent the customer's own data. Salesforce connects case history, purchases and entitlements so answers are about the person asking.
  • Fix the steps where order matters. Kogan.com enforces its seven return steps in order and lets the model handle the rest.
  • Hand off with a summary. Both Salesforce and Kogan.com pass the person a summary so the customer does not repeat the problem.
  • Add one use case at a time and measure real resolution, not just chats handled.

Salesforce Help site: from a keyword bot to an AI agent

Salesforce put an AI agent (Agentforce) on its own help site. The company reports that it resolves about 76% of questions without a person. Source dated 4 December 2024.

Support agents with headsets working at laptops along a long table
Workflow diagram
Before
  1. BeforeCustomer asks the keyword bot
  2. BeforeBot lists resources, no personal answer
  3. BeforeStuck on a password reset: open a ticket
  4. BeforeWait, sometimes days
After (Agentforce)
  1. 1
    CustomerCustomer asks in plain language
  2. 2
    SystemAgent reads case history, entitlements, articles (Data 360, RAG)
  3. 3
    AI agentAgent answers or does the task, e.g. fixes the password reset
  4. 4
    Hand-off to a personNeeds more care (about 5%): ticket opened, live rep joins the same chat with a summary

Before

The help site gets more than 60 million visits a year. An earlier bot (Einstein Bots) handled a fixed list of requests and needed keywords. It could not give personal answers. If a customer got stuck on a password reset, they had to open a support ticket and could wait days.

How the process was automated, step by step

  1. The customer types any question in plain language into the chat on the help site.
  2. The agent looks up the customer's own data: case history, product purchases, entitlements, knowledge articles and product documentation. Salesforce says it combines these into one prompt using retrieval-augmented generation (RAG).
  3. The agent answers using that context. Example from the source: a customer asks about their API limits, and the agent gives the exact answer for their account instead of a list of search results.
  4. The agent does routine tasks itself. Example: if the self-service password reset fails, the agent steps in and fixes it without a ticket.
  5. If the request needs more care, such as raising an API limit, the agent opens a service ticket and connects the customer to a live person in the same chat window. It writes a summary first, so the customer does not repeat themselves.

Tools and systems connected

  • Agentforce
  • Service Cloud
  • Data 360

Agentforce is the AI agent and Service Cloud is the help desk. Data 360 holds case histories, customer IDs, purchases, entitlements, articles and documentation.

Seven languages: English, Spanish, Portuguese, French, Italian, German and Japanese.

Where people stay in the loop

  • About 5% of conversations go to a human support engineer.
  • Human reps focus on the more complex cases. The source says the agent has handled more than 1.7 million conversations to date, while human reps focused on 1.3 million more complex cases.
  • Salesforce says its teams defined the use cases, set guardrails and tested the system before launch. The build took two months using low-code tools.

Reported results (company figures)

  • 76% of inquiries resolved without a human.
  • 65% shorter response time for 90% of users.
  • More than 1.7 million conversations handled to date.

Sources: Salesforce customer story, 4 December 2024

Kogan.com: one agent, built use case by use case

Kogan.com is an Australian online retailer. Salesforce published its story on 10 July 2026. The agent launched in March 2026.

A person packing and labelling cardboard boxes at a desk next to a laptop
Workflow diagram
Before
  1. BeforeWeb forms and email only
  2. BeforeNo live chat, no phone line
  3. BeforeSlow replies at peak sales
Order questions
  1. 1
    CustomerCustomer asks about an order
  2. 2
    SystemNot logged in? A login agent logs them in, then returns them
  3. 3
    AI agentAgent checks status, cancels, or updates the delivery address
TV return: 7 fixed steps in order
  1. 4
    AI agentOrder lookup
  2. 5
    AI agentMarketplace eligibility
  3. 6
    AI agentDuplicate case check
  4. 7
    AI agentTroubleshoot if fault not confirmed
  5. 8
    CustomerPhoto upload
  6. 9
    AI agentReturn eligibility decision
  7. 10
    SystemCase created
When the agent cannot solve it
  1. 11
    SystemAgent checks the live queue
  2. 12
    Hand-off to a personPerson takes over with a conversation summary
  3. 13
    CustomerIf the wait is over 15 minutes: callback case offered

Before

All customer service ran through web forms and email. There was no live chat and no inbound phone line. Responses were slow, especially at peak sales. The most common question, "where is my order", made up about 60% of requests.

How the process was automated, step by step

  1. Order questions come first. The first use case covers the full order life: status checks, cancellations and delivery address changes. If the customer is not logged in, the agent passes them to a small login agent that does only one job: get the customer logged in. Then it returns them to where they were.
  2. Returns come next. The TV return flow has seven fixed steps that are always followed in order: look up the order, check marketplace eligibility, check for a duplicate case, troubleshoot if the fault is not confirmed, upload photos, decide return eligibility, create the case.
  3. If a customer leaves the return halfway to ask something else, the agent clears that part and takes them back to the main chat without losing the session.
  4. Product advice and TV troubleshooting come after that. Product advice uses Kogan.com's own search and product-detail APIs. Troubleshooting uses the same knowledge articles the support team already writes.
  5. If the agent cannot solve it, it checks the live queue. If the wait is over 15 minutes, it offers a callback case. The human gets a conversation summary on every hand-off.

Tools and systems connected

  • Agentforce
  • Salesforce Knowledge
  • Agent Script
  • Kogan.com order, search and product APIs

Agent Script is Salesforce's scripting language for fixed steps. Kogan.com uses it only where the order of steps must not change, such as returns. Order status and product questions use the model's own reasoning.

Where people stay in the loop

  • Human reps take everything the agent cannot solve, with a summary.
  • The customer care team builds and owns the agent, adding one use case at a time. The CTO says the people who know customers best build and own the agent.

Reported results (company figures)

  • The agent now handles about 50% of customer inquiry volume.
  • Kogan.com's own "true resolution" measure (question fully answered with no follow-up) rose from 6.2% at launch to about 20%.
  • Customer preference for the agent rose from 50% to 78%, and satisfaction for agent chats is within four points of human chats.
  • Over a year, sales volume grew 20% while servicing cost fell 10%.
  • Kogan.com's CTO said at Dreamforce that it resolves 67% of AI-handled conversations and 81% are not escalated to a person (reported by Marketing-Interactive, Sep 2026). Salesforce's case study gives a stricter measure, "true resolution": 6.2% at launch to about 20% now. Neither source explains how the two figures relate.

Sources: Salesforce blog on Kogan.com, 10 July 2026, Marketing-Interactive, 21 September 2026

Klarna: the best-known example, and the reversal

Klarna, the payments company, launched an AI assistant in 2024 and became the most quoted example. Its own later statements are part of the story.

A hand holding a smartphone with a payment card on a wooden table
Workflow diagram
What the sources say (order of steps not published)
  1. 1
    CustomerCustomer asks in the Klarna app
  2. 2
    AI agentAssistant on OpenAI models; LangGraph routes requests between tasks
  3. 3
    AI agentRefunds, returns, payment problems, cancellations, disputes, invoice errors, balances
  4. 4
    PersonFrom May 2025: customers can always ask for a person

Steps not published. Klarna has not published the sequence. The boxes show what it has stated.

Before

Klarna quotes an average resolution time of 11 minutes before the assistant.

Tools and systems connected

  • OpenAI models
  • LangGraph
  • LangSmith

OpenAI models power the assistant, per Klarna's press release and OpenAI's case study. LangChain says the assistant was built on LangGraph, which routes requests between different tasks, with prompts tailored to each scenario, and tested with LangSmith.

The assistant handles refunds, returns, payment problems, cancellations, disputes, invoice errors and balance questions, in the Klarna app, in more than 35 languages.

Where people stay in the loop

  • Not stated at launch.
  • In May 2025 a Klarna spokesperson told Customer Experience Dive the company wants customers to always have the option to speak to a human.

Reported results

  • Company-reported, February 2024, after one month: 2.3 million conversations, two-thirds of customer service chats, equal to 700 full-time agents, satisfaction on par with humans, 25% fewer repeat inquiries, resolution in under 2 minutes instead of 11.
  • Company-reported by Klarna, quoted by OpenAI: an estimated US$40 million profit improvement in 2024.
  • Reported by LangChain, February 2025: about 70% of repetitive support tasks automated and 80% faster resolution of customer queries in escalation handling.
  • May 2025, reported by Bloomberg: the CEO said cost-cutting had gone too far, and Klarna planned a recruitment drive so customers can always reach a real person. Fortune, citing Bloomberg, reported that he said the AI approach had led to lower quality.

Klarna has not published the step-by-step flow. It is here because it is the best-known example, and because of what happened next.

Sources: Klarna press release, 27 February 2024, OpenAI case study, LangChain case study, 12 February 2025, Bloomberg, 8 May 2025, Customer Experience Dive, 9 May 2025, Fortune, 9 May 2025

What to watch out for

  • Most figures come from the companies or their vendors, not independent studies.
  • Counting methods differ. Kogan.com's own "true resolution" figure (about 20%) is far lower than its share of inquiries handled (about 50%). Ask which one a vendor means.
  • Cost cutting can go too far. Klarna's CEO said in May 2025 that it had, and the company moved to make human help always available.
  • Keep a clear way to reach a person. Every example here has a hand-off route; the Klarna reversal shows why.
  • Test before launch and after every change. Salesforce set guardrails and tested for two months; Kogan.com ran A/B tests at every release.

Common questions

How do companies automate customer support with AI?
In the cases on this page, the company connects an AI agent to its customer data and knowledge articles, starts with the most common question, and adds use cases one by one. The agent answers or does the task, and passes the rest to a person with a summary.
What are examples of AI customer service automation?
Salesforce's help site (about 76% of questions resolved without a human, company-reported), Kogan.com (order tracking, returns and product advice) and Klarna (an assistant for refunds, returns and payment questions).
How much of customer support can AI handle?
Companies report different shares: Salesforce about 76% of help-site questions, Kogan.com about 50% of inquiries and about 20% true resolution, Klarna two-thirds of chats in its first month. These are company-reported and measured differently.
Do customers still reach a human?
Yes in the Salesforce and Kogan.com setups, where the agent passes the conversation to a person with a summary. Klarna said in 2025 that it wants customers to always have that option.
What AI agents or tools did these companies use?
Salesforce and Kogan.com used Salesforce Agentforce with Service Cloud and Knowledge. Klarna used OpenAI models and, per LangChain, LangGraph and LangSmith.

Sources

Related on Ai Siftly

Browse customer support workflows and case studies. More guides: AI accounts payable and finance automation examples and AI recruiting automation examples.