Two-thirds of customer chats resolved by an AI assistant
How consumer businesses use AI assistants for service, and what the best-known deployment teaches about keeping humans in the loop.
Figures: Klarna, 2024. Source and caveats below.
01 The problem
Customer service volume rises with every order, and most of it is the same handful of questions: where is my order, how do I return this, why was I charged. Staffing a contact center for those questions around the clock and in every language is expensive, and customers still wait.
The first generation of chatbots made this worse. They matched keywords, failed on anything unexpected, and trained customers to type "agent" straight away.
02 Where AI fits
Service assistant
An assistant connected to order, payment and returns systems that resolves the request instead of deflecting it: it issues the refund, changes the address, rebooks the delivery.
Agent assist
For conversations that reach a person, the model summarizes the history, suggests the reply and completes the after-call notes.
Product discovery
Conversational search and recommendations that understand "a waterproof jacket for a week in Scotland under $200".
Catalog content
Product descriptions, attributes and translations generated and kept consistent across thousands of SKUs.
Demand forecasting
Store and SKU-level forecasts that reduce both stockouts and markdowns.
03 Reference architecture
The layers we put in place for a typical retail & customer experience deployment, from your systems of record up to the people who review the output.
- L1Commerce, OMS and CRM (Shopify, Salesforce, Zendesk)
- L2Knowledge base and policy retrieval
- L3Service agent with transactional tools
- L4Guardrails and escalation rules
- L5Human agent workspace
- L6Quality scoring and analytics
04 Real-world implementation
Klarna reported that in its first month live, its AI assistant held 2.3 million conversations, two-thirds of all its customer service chats, across 23 markets and more than 35 languages. Klarna said the assistant did work equivalent to 700 full-time agents, matched human agents on customer satisfaction, cut repeat inquiries by 25% and brought average resolution time from 11 minutes to under 2. The company estimated a $40 million profit improvement for 2024.
Source: Klarna press release, February 2024. Klarna is not a 925 Techies client.
What to keep in mind: In 2025 Klarna's chief executive said the company had leaned too far toward cost and would make sure customers can always reach a person. The lesson we take from it: automate the routine, and make the route to a human short and obvious.
05 Business impact
Cost to serve
Routine contacts resolved automatically lower cost per contact and remove the need to staff up for seasonal peaks.
Speed
Customers get an answer in seconds, at any hour and in their own language.
Revenue
Better discovery and faster answers before purchase lift conversion and reduce abandoned carts.
Agent quality
Human agents spend their time on complex or sensitive cases, with the context already assembled.
06 Implementation roadmap
- Weeks 1–2
Assess
Classify a year of tickets by intent. Identify the top intents by volume and how resolvable each is.
- Weeks 3–7
Prove
Assistant handles two or three intents for a small share of traffic. Compare resolution and satisfaction with human handling.
- Weeks 8–16
Build
Add transactional tools, languages and channels. Build the handoff so context carries over to the human agent.
- Ongoing
Operate
Review failed conversations weekly, expand intents, keep satisfaction on par with human agents.
07 Guardrails
- One-step escalation to a human at any point
- Limits on refund and credit amounts
- Answers grounded in current policy only
- Satisfaction tracked separately for AI and human conversations
08 Services involved
Agentic AI & Workflow Automation
Agents that plan, call your tools and complete multi-step work, with a human approving what matters.
RAG & Enterprise Knowledge
Answers grounded in your documents, with a citation on every claim and access rules respected.
LLMOps & Managed AI
Monitoring, evaluation and tuning that keep a live AI system accurate and affordable.
Start here
Exploring AI in retail & customer experience?
We can run a proof of concept on your own data in about four weeks, measured against the KPIs above.
