Klarna: the RiskMandate X-ray

a company · prepared · sources read on 2026-11-07 · the deck · how this X-ray was made · X-ray another name

What happened: how this X-ray was made →

With JavaScript, the graph of this X-ray and a player for each use case's simulation are drawn here. Everything they show is written out below.

What it does

Klarna is a payments and shopping company (it describes itself as a global digital bank and flexible payments provider) whose app lets consumers pay later, manage payments and shop. Since February 2024 its OpenAI-powered AI assistant has answered customer-service chats in the app, handling refunds, returns, payment issues, cancellations, disputes and invoice inaccuracies; Klarna says it handled 80% of customer-service chats in 2025. Customers can still ask for a human, and in May 2025 the CEO said Klarna would recruit human agents so that a person is always available. s1 s6 s3 s8 s9

Its AI agents

AI assistant (service)

The customer-service chatbot in the Klarna app, launched globally in February 2024 with OpenAI, available 24/7 in 23 markets and more than 35 languages. Klarna says it had 31 million conversations since launch and handled 80% of customer-service chats in 2025, the equivalent of over 850 full-time agents. It also gives customers their balances, payment schedules and spending limits. LangChain's customer story says it is built on LangGraph and tested with LangSmith. s1 s6 s3 s4 s15

What it can do:

Answers customer-service chats in the app, in more than 35 languages
Handles refunds, returns, payment-related issues, cancellations, disputes and invoice inaccuracies
Shows outstanding balances, upcoming payment schedules and purchase power, and explains spending limits
Customers can choose a human agent instead; Klarna tells users they are talking to a chatbot

AI assistant (shopping)

Chat-based shopping features added to the same assistant in September 2024: product search, personalised recommendations, comparisons and pricing, integrated with Klarna's price database of 5.6 million products from thousands of stores (Pricerunner). Users can buy directly in the app with Klarna's flexible payment methods. s2

What it can do:

Searches products and brands, with reviews and price history
Recommends products from the user's shopping preferences and habits
Compares products with an AI-generated pros-and-cons summary
Reads price, stock, delivery and cashback information across merchants
Leads to a purchase in the app with Klarna's payment methods

Kiki (internal)

Klarna's internal knowledge chatbot on OpenAI models, launched June 2023. Klarna said in May 2024 it had answered over 250,000 employee questions (about 2,000 a day) and that 85% of employees used it; the 2025 annual report says it helps employees find information across internal systems. s5 s6

What it can do:

Answers employee questions in 1 to 5 seconds
Personalises answers to the asker (for example, what my team is working on)
Draws on information across Klarna's internal systems

Use cases: reach, mandate and gap

For each use case: what the agent can technically reach (from the sources), what a deployer would plausibly authorise (the mandate: illustrative, our judgement), the gap between the two (authority nobody accepted), and the controls that publicly stand in the way.

2024 to 2025: AI-first customer service, then humans back

The agent: AI assistant (service) · who uses it: Klarna's consumers using the in-app chat (Klarna's release cites over 114 million active users), and the outsourced and freelance human agents behind it.

Reach	Kind	Sources
Customer-service chats in the Klarna app: two-thirds of chats in the first month (2.3 million conversations), 80% of chats in 2025, Klarna says	people	s1 s6
Refunds, returns, payment-related issues, cancellations, disputes and invoice inaccuracies	money	s1 s14
Each customer's outstanding balances, upcoming payment schedules, purchase power and spending limits	data	s1
The hand-off to a person: customers can choose a live agent, and the US help page offers human support or a dedicated agent	people	s1 s3 s4
The customer-service workforce: about 3,000 agents, mostly at outsourcing partners, which a spokesperson said would fall to about 2,300	people	s13 s14
Klarna's public and regulatory statements about the assistant: 700 then 853 agents' equivalent work, $39 million (2024) and about $59 million (2025) in cost savings, satisfaction on par with humans, all from Klarna's own measurements	outside	s1 s6 s7 s11

Mandate (illustrative)
Klarna's stated policy: every user is told they are talking to a chatbot and can choose to speak with a human instead (Klarna, December 2024)
As a Klarna spokesperson put it in March 2024: more experienced and senior staff are still needed, with specialised training for complex or sensitive cases
Illustrative: the assistant resolves routine errands end to end; a change to a customer's money or payment terms is recorded with who decided it

Gap: reach beyond the mandate	From the reach
Disputes, refunds and invoice corrections sit in the same chat reach as routine questions; nothing public shows where the line to senior human staff for complex or sensitive cases was drawn.	Refunds, returns, payment-related issues, cancellations, disputes and invoice inaccuracies; Each customer's outstanding balances, upcoming payment schedules, purchase power and spending limits
The human path that contains the agent (the choice of a person) depended on a workforce that was shrinking while the assistant took more chats.	The hand-off to a person: customers can choose a live agent, and the US help page offers human support or a dedicated agent; The customer-service workforce: about 3,000 agents, mostly at outsourcing partners, which a spokesperson said would fall to about 2,300
The evidence of success is Klarna's own: share of chats, resolution time and satisfaction from its chat logs and internal surveys, with no independent measurement in the public record.	Customer-service chats in the Klarna app: two-thirds of chats in the first month (2.3 million conversations), 80% of chats in 2025, Klarna says; Klarna's public and regulatory statements about the assistant: 700 then 853 agents' equivalent work, $39 million (2024) and about $59 million (2025) in cost savings, satisfaction on par with humans, all from Klarna's own measurements

Controls that publicly stand in the way	Sources
Users are told they are talking to a chatbot and can choose a human	s3 s1 s4
Klarna's 2025 annual report describes processes and controls for generative AI risk, including human involvement in training and monitoring its AI tools	s6
A 'dual-track' approach: scalable AI in customer service combined with high-quality human support, offered to all customers (added to Klarna's disclosures after an SEC comment)	s7 s6
From May 2025, recruitment of human agents in a remote, freelance 'Uber type' pilot so customers can always reach a person	s8 s9 s10
The assistant's critical use cases tested and evaluated with LangSmith (LangChain's account)	s15

The simulation, each action placed against the mandate (RiskMandate is never in the request path: read-only by design: it places, it does not stop):

December 2023: Klarna announces a hiring freeze except for engineering roles as it turns to AI (person, reaches The customer-service workforce: about 3,000 agents, mostly at outsourcing partners, which a spokesperson said would fall to about 2,300): inside the mandate
February 2024: the assistant, live globally for a month, has had 2.3 million conversations, two-thirds of all customer-service chats; Klarna says that is the work of 700 full-time agents (agent, reaches Customer-service chats in the Klarna app: two-thirds of chats in the first month (2.3 million conversations), 80% of chats in 2025, Klarna says): inside the mandate
A spokesperson says the outsourced customer-service department of about 3,000 will shrink to about 2,300 because of the assistant (person, reaches The customer-service workforce: about 3,000 agents, mostly at outsourcing partners, which a spokesperson said would fall to about 2,300): inside the mandate
(illustrative) Handles a customer's dispute over an invoice inaccuracy inside the chat, as the launch release says it can, without a senior human looking at it (agent, reaches Refunds, returns, payment-related issues, cancellations, disputes and invoice inaccuracies): outside the mandate: in the gap. What RiskMandate's record shows: RiskMandate's record shows disputes and refunds (money) inside the same reach as routine questions, against a stated mandate that complex or sensitive cases need experienced staff, and no public evidence of where that line is enforced.
Customers complain of generic answers and of questions too complicated for it, a Forrester analyst later tells CX Dive; an engineer who tested it said it recites the docs and passes him to human support fast (agent, reaches The hand-off to a person: customers can choose a live agent, and the US help page offers human support or a dedicated agent): outside the mandate: in the gap. What RiskMandate's record shows: RiskMandate's record shows the hand-off to a person as the control that contains the agent, and the human capacity behind that hand-off being reduced at the same time. Klarna's own record says satisfaction did not drop.
8 May 2025: the CEO tells Bloomberg the cost-cutting in customer service has gone too far, that cost was 'a too predominant evaluation factor' and the result 'lower quality', and starts recruiting human agents so customers always have a person (person, reaches The hand-off to a person: customers can choose a live agent, and the US help page offers human support or a dedicated agent): inside the mandate
16 July 2025: answering an SEC comment, Klarna says the remark was about the quality of its human customer service, not AI-driven service, was misreported, and adds its 'dual-track' disclosure (person, reaches Klarna's public and regulatory statements about the assistant: 700 then 853 agents' equivalent work, $39 million (2024) and about $59 million (2025) in cost savings, satisfaction on par with humans, all from Klarna's own measurements): inside the mandate
2025: Klarna says the assistant handled 80% of chats, did the work of over 850 full-time agents and saved about $59 million (agent, reaches Customer-service chats in the Klarna app: two-thirds of chats in the first month (2.3 million conversations), 80% of chats in 2025, Klarna says): inside the mandate

Shopping assistant: from recommendation to purchase

The agent: AI assistant (shopping) · who uses it: Klarna app users looking for products, and the merchants whose listings and prices it reads.

Reach	Kind	Sources
Klarna's price database of 5.6 million products from thousands of stores, with stock, delivery and cashback information	data	s2
The user's shopping preferences and habits (what they buy and when, brands they prefer), used for recommendations	data	s2
A purchase made directly in the app with Klarna's flexible payment methods (Pay Later and Fair Financing are consumer credit)	money	s2 s6
Merchants' listings and customer reviews, summarised into an 'unbiased, AI-generated' pros and cons	outside	s2

Mandate (illustrative)
Illustrative: help the user find and compare products with a neutral summary
Illustrative: the purchase, and any credit used for it, is the user's own decision, made outside the recommendation

Gap: reach beyond the mandate	From the reach
Recommendations built from the user's habits run straight into a purchase, possibly on credit, in the same chat flow.	The user's shopping preferences and habits (what they buy and when, brands they prefer), used for recommendations; A purchase made directly in the app with Klarna's flexible payment methods (Pay Later and Fair Financing are consumer credit)
Klarna calls the summaries unbiased, but nothing public shows how the ranking is evidenced to be independent of Klarna's merchant relationships.	Klarna's price database of 5.6 million products from thousands of stores, with stock, delivery and cashback information; Merchants' listings and customer reviews, summarised into an 'unbiased, AI-generated' pros and cons

Controls that publicly stand in the way	Sources
Klarna says it does not use generative AI in credit underwriting; credit decisions use established machine-learning models	s6 s7

The simulation, each action placed against the mandate (RiskMandate is never in the request path: read-only by design: it places, it does not stop):

(illustrative) A user asks for the best tablet under $200 (person, reaches Klarna's price database of 5.6 million products from thousands of stores, with stock, delivery and cashback information): inside the mandate
(illustrative) Reads the price database and merchant listings and suggests three tablets with a pros-and-cons summary (agent, reaches Merchants' listings and customer reviews, summarised into an 'unbiased, AI-generated' pros and cons): inside the mandate
(illustrative) Ranks them using the user's past purchases and preferred brands (agent, reaches The user's shopping preferences and habits (what they buy and when, brands they prefer), used for recommendations): inside the mandate
(illustrative) Presents the top pick with a buy button and Klarna's flexible payment options in the chat (agent, reaches A purchase made directly in the app with Klarna's flexible payment methods (Pay Later and Fair Financing are consumer credit)): outside the mandate: in the gap. What RiskMandate's record shows: RiskMandate's record shows the step from recommendation to a credit-backed purchase inside one agent flow, while the mandate puts that decision with the user; the credit decision itself sits with Klarna's underwriting models, not the assistant.
(illustrative) The user buys the tablet and chooses to pay later (person, reaches A purchase made directly in the app with Klarna's flexible payment methods (Pay Later and Fair Financing are consumer credit)): inside the mandate

Kiki: internal answers across internal systems

The agent: Kiki (internal) · who uses it: Klarna employees across teams, including Communications, Marketing and Legal.

Reach	Kind	Sources
Information across Klarna's internal systems	data	s6
Who is asking: answers are personalised to the person and their team	people	s5
Over 250,000 employee questions and Kiki's answers (about 2,000 a day)	data	s5
Third-party AI tools employees use, where Klarna's own risk factors say confidential information could be disclosed	outside	s6

Mandate (illustrative)
Illustrative: answer an employee from information that employee is already allowed to see
Illustrative: no confidential information leaves Klarna's systems through an AI tool

Gap: reach beyond the mandate	From the reach
Kiki answers across internal systems and personalises by asker; nothing public shows how its answers follow each employee's own access.	Information across Klarna's internal systems; Who is asking: answers are personalised to the person and their team

Controls that publicly stand in the way	Sources
Klarna's annual report describes processes and controls for generative AI risk, including human involvement in training and monitoring its AI tools	s6

The simulation, each action placed against the mandate (RiskMandate is never in the request path: read-only by design: it places, it does not stop):

(illustrative) An employee asks Kiki what their team is working on (person, reaches Who is asking: answers are personalised to the person and their team): inside the mandate
(illustrative) Answers from internal systems in a few seconds, personalised to the asker (agent, reaches Information across Klarna's internal systems): inside the mandate
(illustrative) A different employee asks about another team's contract terms (person, reaches Who is asking: answers are personalised to the person and their team): inside the mandate
(illustrative) Answers from documents held in a system the asker does not normally use (agent, reaches Information across Klarna's internal systems): outside the mandate: in the gap. What RiskMandate's record shows: RiskMandate's record shows Kiki's reach spanning internal systems while the mandate is the asker's own access, with no public evidence of the control that ties one to the other.
(illustrative) The employee pastes the answer into an outside AI tool to summarise it (person, reaches Third-party AI tools employees use, where Klarna's own risk factors say confidential information could be disclosed): outside the mandate: in the gap. What RiskMandate's record shows: RiskMandate's record shows confidential content able to reach a third-party tool, a risk Klarna's own annual report names.

Incidents, as reported

2025-05: Klarna's CEO told Bloomberg that cost-cutting in customer service had gone too far and produced 'lower quality', and that Klarna would recruit human agents (a freelance, remote pilot) so customers always have a person. A Klarna spokesperson said the company was still 'very much AI-first'. In July 2025 Klarna told the SEC the remark concerned human customer service, not AI-driven service, and had been misreported. s8 s9 s10 s7
2026-02: Documented risk: Klarna's annual report says generative AI, including in its customer-service chatbot, may create inaccurate, incomplete or misleading output that may not be easily detectable. s6

The run

The search agents that looked for leads on 2026-11-07T19:08:54.696Z, for a total cost of $0.0563. What a model says is a lead, not a fact.

Agent	Model	Time	Citations	Leads
Perplexity search	`perplexity/sonar`	6.8 s	15	11
OpenAI search	`openai/gpt-5.4-nano:online`	9.5 s	1	5
Gemini search	`google/gemini-3.1-flash-lite:online`	2.4 s	1	6
OpenRouter web search	`anthropic/claude-haiku-5.5`	25.7 s	9	17

Leads: what the models claimed, and what we found

Model	Claim	Status	Source
perplexity	Klarna's AI assistant is powered by OpenAI technology.	confirmed	s1
perplexity	The assistant is available in the Klarna app.	confirmed	s1
perplexity	The assistant manages multilingual customer service, refunds and returns.	confirmed	s1
perplexity	Klarna's US help page says the AI assistant gives 24/7 support in any language, with real human support available.	confirmed	s4
perplexity	Klarna says users are clearly told they are talking to a chatbot and can choose a human.	confirmed	s3
perplexity	The assistant handles a large share of customer-service chats with major efficiency gains.	confirmed	s1
perplexity	Klarna added chat-based shopping features to the assistant.	confirmed	s2
perplexity	Public reporting noted complaints that the AI gave generic answers and struggled with nuanced questions (November 2025).	confirmed	s11
perplexity	Klarna says the assistant is not a full replacement for human support and keeps a dual-track approach.	confirmed	s6
perplexity	The assistant can escalate to a dedicated human agent.	confirmed	s4
perplexity	The assistant assists with shopping guidance and finding products.	confirmed	s2
openai	Klarna launched its AI assistant in February 2024 in partnership with OpenAI.	confirmed	s6
openai	Klarna reports the assistant handled 80% of customer-service chats in the year ended 31 December 2025.	confirmed	s6
openai	The assistant is available 24/7 in 23 markets in more than 35 languages.	confirmed	s6
openai	The assistant has had 31 million conversations since launch.	confirmed	s6
openai	The assistant manages multilingual customer service and refunds and returns.	confirmed	s6
gemini	In its first month the assistant handled 2.3 million conversations, the equivalent of 700 full-time agents.	confirmed	s1
gemini	The assistant works in 35 languages and led to a 25% drop in repeat inquiries.	confirmed	s1
gemini	The assistant handles refunds, returns, payment issues and disputes.	confirmed	s1
gemini	The assistant accesses customer account data (balances, payment schedules).	confirmed	s1
gemini	The assistant adjusts payment due dates.	not-found	s1
gemini	The assistant integrates with Klarna's internal CRM and payment systems.	not-found	s1
openrouter	Klarna announced the assistant on 27 February 2024; in its first month it handled 2.3 million conversations, two-thirds of chats, equal to 700 full-time agents.	confirmed	s1
openrouter	Available in 23 markets, 24/7, in more than 35 languages, with satisfaction on par with human agents.	confirmed	s1
openrouter	Resolution time fell from 11 minutes to under 2, and repeat inquiries dropped 25%.	confirmed	s1
openrouter	Klarna's consumer base is about 150 million consumers worldwide, in 23 markets.	confirmed	s1
openrouter	Klarna's Q3 2025 update put the assistant at the work of 853 agents and about $60 million in savings.	confirmed	s11
openrouter	In May 2025 Klarna said it was 'not reversing on AI' and the assistant did the work of over 800 full-time roles.	unverified
openrouter	The assistant accesses customer data, tracks ticket state, issues returns and manages payment plans, with a 100-person specialist team for complex queries.	unverified
openrouter	Outsourced external human agents fell from about 3,000 to 2,300.	confirmed	s13
openrouter	Human support is now a freelance, on-demand model paying agents per shift, with AI for routine queries.	confirmed	s9
openrouter	The CEO said cost had been too dominant an evaluation factor, producing lower quality, and Klarna began rehiring human agents (May 2025).	confirmed	s9
openrouter	Customers complained of generic answers; Klarna said it saw no drop in consumer satisfaction.	confirmed	s11
openrouter	Critics said the 700 framing was misleading because it described avoided hiring, not layoffs.	confirmed	s16
openrouter	The assistant sometimes gave confident but wrong answers on policy, fees or payment terms.	confirmed	s16
openrouter	A hands-on critique found it mainly automates Level-1 support and documented prompt-injection vulnerabilities.	confirmed	s18
openrouter	AI handling disputes and account closures raised compliance concerns.	confirmed	s17
openrouter	The assistant was built on GPT-4-class models with integrations to account and transaction systems.	confirmed	s17
openrouter	The assistant routes complex cases such as disputes and fraud to human agents with context attached.	not-found	s1

The questions we would ask

We have not asked them yet.

Which errands can the AI assistant complete on its own (refunds, dispute outcomes, invoice corrections, payment changes), and which must go to a person before money or payment terms change?
How is a dispute or other complex or sensitive case detected in the chat and handed to experienced staff, and is that hand-off recorded?
The satisfaction and accuracy figures come from your own chat logs and surveys: are they broken down by case type, and has anyone outside Klarna measured them?
When the shopping assistant leads to a purchase on credit, what separates the recommendation from the credit offer, and how is the 'unbiased' summary evidenced?
How do Kiki's answers respect each employee's own access to internal systems?

Sources

s1 Klarna AI assistant handles two-thirds of customer service chats in its first month (27 February 2024), Klarna International press (own-site, read 2026-11-07)
https://www.klarna.com/international/press/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month/
s2 Shopping made smarter: Klarna adds more AI features to its assistant powered by OpenAI (19 September 2024), Klarna International press (own-site, read 2026-11-07)
https://www.klarna.com/international/press/shopping-made-smarter-klarna-adds-more-ai-features-to-its-assistant-powered-by-openai/
s3 Ensure AI Benefits Everyone (4 December 2024), Klarna International (own-site, read 2026-11-07)
https://www.klarna.com/international/press/ensure-ai-benefits-everyone/
s4 Customer service | Klarna US, Klarna (own-site, read 2026-11-07)
https://www.klarna.com/us/help/
s5 90% of Klarna staff are using AI daily: game changer for productivity (14 May 2024), Klarna International press (own-site, read 2026-11-07)
https://www.klarna.com/international/press/90-of-klarna-staff-are-using-ai-daily-game-changer-for-productivity/
s6 Klarna Group plc annual report on Form 20-F for the year ended 31 December 2025 (filed 26 February 2026; listed on the EDGAR company page), US SEC EDGAR (Klarna Group plc filing) (filing, read 2026-11-07)
https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=2003292&type=20-F
s7 Klarna Group plc response letter to SEC staff comments on its Form F-1 registration statement (16 July 2025; listed on the EDGAR company page), US SEC EDGAR (Klarna Group plc correspondence) (filing, read 2026-11-07)
https://www.sec.gov/cgi-bin/browse-edgar?action=getcompany&CIK=2003292&type=CORRESP
s8 Klarna Slows AI-Driven Job Cuts With Call for Real People (8 May 2025; opening paragraphs only, rest paywalled), Bloomberg Law (press, read 2026-11-07)
https://news.bloomberglaw.com/artificial-intelligence/klarna-rethinks-ai-cost-cutting-plan-with-call-for-real-people
s9 Klarna plans to hire humans again, as new landmark survey reveals most AI projects fail to deliver (9 May 2025), Fortune (press, read 2026-11-07)
https://fortune.com/2025/05/09/klarna-ai-humans-return-on-investment/
s10 Klarna Is Hiring Customer Service Agents After AI Couldn't Cut It on Calls, According to the Company's CEO (9 May 2025), Entrepreneur (press, read 2026-11-07)
https://www.entrepreneur.com/business-news/klarna-ceo-reverses-course-by-hiring-more-humans-not-ai/491396
s11 Klarna says its AI agent is doing the work of 853 employees (20 November 2025), CX Dive (press, read 2026-11-07)
https://www.customerexperiencedive.com/news/klarna-says-ai-agent-work-853-employees/805987/
s12 Klarna sees its AI assistant as a success. Is it too early to tell? (12 March 2024), CIO Dive (press, read 2026-11-07)
https://www.ciodive.com/news/klarna-generative-ai-customer-service-call-center/709916/
s13 Klarna's AI bot is doing the work of 700 employees. What will happen to their jobs? (28 February 2024), Euronews (press, read 2026-11-07)
https://www.euronews.com/2024/02/28/klarnas-ai-bot-is-doing-the-work-of-700-employees-what-will-happen-to-their-jobs
s14 Klarna's AI assistant is doing the work of 700 people, and it paints a bleak picture for the future of work (29 February 2024), ITPro (press, read 2026-11-07)
https://www.itpro.com/technology/artificial-intelligence/klarnas-ai-assistant-is-doing-the-work-of-700-people-and-it-paints-a-bleak-picture-for-the-future-of-work
s15 How Klarna's AI assistant redefined customer support at scale for 85 million active users (12 February 2025), LangChain (vendor customer story) (other, read 2026-11-07)
https://www.langchain.com/blog/customers-klarna
s16 What Klarna's AI Did in 30 Days, And What Broke (18 March 2026), Twig (vendor blog) (other, read 2026-11-07)
https://www.twig.so/blog/how-klarna-is-revolutionizing-customer-support-with-ai
s17 Klarna AI Saved $40M on Support, Then Walked It Back (18 March 2026), Twig (vendor blog) (other, read 2026-11-07)
https://www.twig.so/blog/klarna-ai-customer-support-efficiency
s18 Klarna AI Assistant: Two-Thirds of Support Automated, Fini (vendor blog) (other, read 2026-11-07)
https://www.usefini.com/blog/klarna-automates-two-thirds-of-customer-service-with-ai-assistant
s19 Klarna AI reversal: mid-market lessons (21 May 2026), IrisAgent (vendor blog) (other, read 2026-11-07)
https://irisagent.com/blog/klarna-ai-reversal-mid-market-lessons/
s20 Klarna's AI customer service deployment | AI Autopsy 002 (12 June 2026), Bigeye (vendor blog) (other, read 2026-11-07)
https://www.bigeye.com/blog/klarnas-ai-customer-service-deployment
s21 Klarna: Enterprise AI Deployment Profile (31 October 2026), Imagine (other, read 2026-11-07)
https://imagine.ai/research/klarna

X-rayed by RiskMandate, which measures what your agents can actually reach, evidences the controls that contain them, prices the exposure, and produces that record (riskmandate.ai/llms.txt). The mandates are illustrative. Built by xray@web-summit.dinis from data/xray/klarna/xray.json.