How Klarna turned AI into an operating system for support, shopping, marketing, and risk

Klarna is not interesting because it used AI once.

Klarna is interesting because the public evidence shows AI spreading through almost every kind of work the company does: support, shopping, marketing, engineering, legal, underwriting, fraud detection, internal knowledge management, product data, and AI-agent commerce.

That makes it one of the better companies to study if you want to understand what practical AI adoption looks like after the press-release phase. This is not a clean story where every detail is public. The exact prompts, evals, retrieval layers, routing logic, and model-monitoring systems are mostly not disclosed. But the receipts are unusually rich: SEC filings, OpenAI’s customer story, Klarna product announcements, Google Cloud’s partnership announcement, GitHub repositories, product docs, and reprinted company releases all point in the same direction.

Klarna is trying to make AI do three jobs at once.

First, reduce operating work in places where the company already knows what good looks like. Customer support is the obvious example.

Second, increase employee output by putting generative tools into everyday workflows. This shows up in engineering copilots, legal document review, internal knowledge search, copywriting, and team-built GPTs.

Third, turn Klarna’s commerce data into AI-native distribution. This is the more strategic part: conversational shopping, image search, AI product feeds, agentic product protocols, and integrations with ChatGPT, Gemini, Google Search, and Google Pay.

Most companies talk about AI as a productivity layer. Klarna’s public story is sharper than that. It is using AI as a cost lever, a product surface, a marketing engine, a risk system, and a bet on how shopping changes when consumers start asking agents instead of browsing stores.

Here’s what to expect

• ✔️ The 60-second briefing: what Klarna is really doing with AI
• ✔️ AI adoption timeline: how it moved from ChatGPT plugin to AI operating model
• ✔️ Function-by-function breakdown: support, product, marketing, engineering, legal, risk, and strategy
• ✔️ Workflow breakdowns: what changed, what tools show up, what results were claimed, and what to steal
• ✔️ Caveats: what is company-reported, what is missing, and what not to overcopy
• ✔️ Sources and receipts: the public evidence behind the case

The 60-second briefing

Main takeaway: Klarna’s AI strategy is not one workflow. It is a company-wide operating map. The company is automating high-volume support, making generative AI normal inside daily work, embedding AI into the shopping product, and exposing commerce data to AI agents.

Most important workflows:

• Customer support: OpenAI-powered assistant handling a large share of support chats, with reported improvements in resolution time and repeat inquiries. [1]
• Employee productivity: ChatGPT Enterprise and internal tools used across the company, including Kiki, an internal knowledge chatbot. [2][3]
• Engineering: AI copilots connected to engineering work software for code creation and review. [3]
• Marketing: AI copy, image, and campaign workflows used to reduce agency spend and increase campaign output. [6]
• Shopping product: AI assistant, product recommendations, product comparisons, image search, and live product discovery inside chat surfaces. [4][9]
• Risk and compliance: ML underwriting, AI risk disclosures, human monitoring, and Google Cloud graph neural networks for fraud and AML. [3][5]
• Agentic commerce: Klarna is building protocols and product-search infrastructure so AI agents can access its product and price data. [8][9][10]

Biggest result: The support assistant is the cleanest metric-backed workflow. Klarna said that, one month after global launch, its AI assistant had handled 2.3 million conversations, roughly two-thirds of its customer-service chats, with errands resolved in under two minutes compared with eleven minutes previously. [1]

Most copyable idea: Start where the work is repetitive, measurable, and reviewable. Klarna did not begin with vague “AI transformation.” It put AI into customer interactions, internal knowledge lookup, code review, document review, copywriting, product search, and risk scoring — workflows where volume, speed, cost, and quality can be tracked.

Best for: Support leaders, product managers, marketers, engineering managers, risk/compliance leaders, and founders trying to understand what an AI-first operating model actually looks like when it touches real work.

Best metric: “The AI assistant has had 2.3 million conversations, two-thirds of Klarna’s customer service chats.” This is the cleanest public proof point because it ties AI to a real operating workflow: support volume, resolution speed, repeat inquiries, customer satisfaction, and profit impact. [1]

AI adoption timeline

The timeline is useful because it shows progression. Klarna did not stay at “we have a chatbot.” It moved from early LLM integration, to support automation, to internal adoption, to shopping discovery, to risk infrastructure, to agentic commerce distribution.

Timeline

How Klarna’s AI work spread from plugins to agentic commerce

1. March 23, 2023

   Public launch of Klarna’s early ChatGPT shopping integration.

   “one of the first brands to work with OpenAI to use its protocol to build an integrated Plugin for ChatGPT”

   Source: Klarna brings smoooth shopping to ChatGPT

2. August 30, 2023

   Klarna shifted from experimental plugin work to broad internal enterprise AI use.

   “implemented OpenAI’s API which offers higher speeds, better availability and enterprise-level data security”

   Source: Klarna is the only bank among ChatGPT Enterprise launch customers

3. October 11, 2023

   Klarna publicly rolled out computer-vision shopping features.

   “New AI-powered image-search allows users to snap, search and shop anything around them.”

   Source: Klarna taps AI for human shopping experience in a major autumn product launch

4. February 27, 2024

   Klarna’s OpenAI-powered support assistant reached global scale within a month.

   “The AI assistant has had 2.3 million conversations, two-thirds of Klarna's customer service chats”

   Source: Klarna AI assistant handles two-thirds of customer service chats in its first month

5. May 14, 2024

   Public milestone for internal genAI adoption and the Kiki knowledge assistant.

   “9 out of 10 Klarna employees (87%) now harness the power of Generative AI in their daily work.”

   Source: 90% of Klarna Staff Are Using AI Daily - Game Changer for Productivity

6. May 28, 2024

   Reveals unusually broad internal GPT experimentation.

   “Klarna employees across the company have built over 300 GPTs for internal use.”

   Source: AI helps Klarna cut marketing agency spend by 25% and run more campaigns

7. May 30, 2024

   Klarna tied AI adoption directly to lower operating expenses in earnings commentary.

   “90% of Klarna employees have integrated AI into their daily workflows”

   Source: Klarna Announces Profitable Start to 2024 as It Sets the Stage for a Year of Innovation and Growth

8. September 19, 2024

   One of the clearest public descriptions of Klarna’s shopping-AI architecture.

   “We’ve combined the intelligence of LLMs, the knowledge of Pricerunner and the Klarna UX”

   Source: Shopping made smarter: Klarna adds more AI features to its assistant powered by OpenAI

9. March 2025

   SEC filing formalized Klarna’s internal AI-copilot usage and governance disclosures.

   “85% of our engineers connected to their work software an AI copilot that can create and review code”

   Source: Klarna Group plc Form F-1

10. May 19, 2025

    AI became central to Klarna’s investor-facing operating narrative.

    “Our AI-first strategy is driving exceptional returns”

    Source: Klarna accelerates global momentum in Q1 2025 and unlocks large gains from AI innovation

11. October 9, 2025

    Klarna publicly broadened beyond OpenAI into Google Cloud models and AI infrastructure.

    “Klarna will leverage Google Cloud's complete AI stack, from infrastructure, to platform, to models”

    Source: Klarna and Google Cloud Enter Strategic AI Partnership to Bring More Creative and Engaging Shopping Experiences to Millions of Klarna Users Worldwide

12. December 15, 2025

    Klarna moved from AI features to agentic-commerce infrastructure and standards.

    “Klarna’s Agentic Product Protocol defines a common language”

    Source: Klarna launches Agentic Product Protocol: The open standard that makes 100M+ products instantly discoverable by AI agents

13. February 2, 2026

    Klarna publicly aligned with emerging open protocols for AI commerce interoperability.

    “UCP enables consumers to shop seamlessly in AI conversations”

    Source: Klarna backs Google’s Universal Commerce Protocol (UCP) to enable agentic commerce across platforms

14. May 12, 2026

    Klarna extended AI activity from discovery to checkout inside conversational search surfaces.

    “bring its flexible payment options to Google’s Gemini app and Google Search, including AI Mode”

    Source: Klarna’s Flexible Payments Are Coming to Google Search and the Gemini App Within Google Pay

15. May 20, 2026

    Klarna explicitly shipped an MCP-backed ChatGPT commerce integration.

    “Powering the experience is Klarna's Product Search MCP server”

    Source: Klarna launches AI-powered Shopping Search app in ChatGPT

How Klarna uses AI by function

Customer support

Workflow: AI support assistant for high-volume customer service

Evidence snapshot: 2.3M conversations in the first month; two-thirds of customer-service chats; errands reportedly fell from 11 minutes to under 2 minutes.

What they’re doing

Klarna’s support assistant is the clearest example of AI moving from experiment to production work. It handles common customer-service jobs: refunds, returns, payment issues, cancellations, disputes, invoice problems, balance updates, payment schedules, and purchase-power explanations. [1]

Before AI, this work sat with human support agents and support systems. After AI, the assistant became the first-line interface for a large share of interactions, with live agents still available for customers who prefer them or likely need escalation. [1]

This is important because it is not just “answer FAQs.” It is support tied to account, order, payment, and policy workflows. That is where AI starts to change the operating model: customers get faster answers, the support queue shrinks, and humans can focus on edge cases.

Setup

The public setup is partially visible. The assistant is powered by OpenAI, lives in the Klarna app, is global, works across 23 markets, and supports more than 35 languages. [1]

The exact architecture is not public. The evidence does not reveal the retrieval layer, routing logic, eval process, confidence thresholds, or escalation design. But the task list implies integration with Klarna’s customer, payment, order, refund, and purchase-power systems.

Result

Klarna reported that the assistant handled 2.3 million conversations in its first month, representing about two-thirds of customer-service chats. It also reported a 25% drop in repeat inquiries, faster resolution, customer satisfaction on par with human agents, and an estimated $40 million profit improvement in 2024. [1]

The most useful operational metric is the time compression: from eleven minutes to under two minutes for errands. [1]

What to steal

Do not start with the most complex customer problem. Start with the highest-volume support errands where you already have clear policies, clear data, and clear success metrics.

A practical version for another company:

1. Pull the top 50 support intents by volume.
2. Mark which ones require account data, policy lookup, workflow action, or human judgment.
3. Start with the intents that are repetitive and easy to audit.
4. Track repeat contact rate, resolution time, customer satisfaction, and escalation rate.
5. Keep a human path visible.

The biggest lesson is that AI support should be measured as an operations workflow, not as a demo.

Caveat

The strongest numbers are company-reported. Also, a support assistant that can answer payment and refund questions requires data access, permissions, monitoring, and escalation rules. A smaller company can copy the workflow pattern, but not the full global deployment on day one.

Product and shopping experience

Workflow: turning shopping search into conversation

Evidence snapshot: Klarna says the assistant combines LLM intelligence, PriceRunner knowledge, and Klarna UX.

What they’re doing

Klarna’s product AI is not just support dressed up as shopping. It is a different workflow: help consumers discover, compare, and evaluate products inside a conversational interface.

In September 2024, Klarna described an assistant that could provide personalized product recommendations, product comparisons, expert advice, customer reviews, product insights, price information, stock, delivery, and cashback data. [4]

That matters because shopping search is usually fragmented. A shopper jumps between search results, product pages, reviews, discount pages, delivery info, and checkout. Klarna’s bet is that an AI assistant can collapse those steps into one guided interaction.

Setup

Klarna said it combined LLM intelligence, PriceRunner knowledge, and Klarna UX. [4]

That one sentence is the product strategy. LLMs make the interface conversational. PriceRunner gives it product and price intelligence. Klarna’s UX and payments network connect the experience to actual commerce.

The visible setup includes:

• a chat interface;
• product and brand search;
• product recommendations;
• product comparisons;
• customer reviews and product insights;
• price, stock, delivery, and cashback data;
• later, ChatGPT and MCP-based shopping search. [4][9]

Result

The public sources give more product-description evidence than conversion evidence for this specific assistant. The better measurable claim appears later in the Google Cloud partnership, where early pilots around AI-driven creative concepts and personalized product campaigns reportedly boosted time spent in the app by 15% and orders by 50%. [5]

For the shopping assistant itself, the result is strategic: Klarna is trying to move from a payment button to a shopping interface.

What to steal

If your company has valuable product, customer, pricing, inventory, or comparison data, do not just ask, “Can we add a chatbot?”

Ask this instead:

What high-intent decision does our customer already make with five tabs open?

Then build the AI workflow around that decision.

For Klarna, the decision is: what should I buy, where is it available, what does it cost, can I trust it, and how should I pay?

For another company, it might be: which plan should I choose, which vendor should I compare, which policy applies, which SKU fits my use case, or which next step should I take.

Caveat

Public evidence does not disclose whether Klarna’s assistant uses RAG, vector search, reranking, tool-calling, or a custom orchestration layer behind the scenes. We can see the product surface and data dependencies. We cannot fully reconstruct the architecture.

Workflow: image search as a product-discovery shortcut

Evidence snapshot: Klarna announced AI-powered image search that lets users snap, search, and shop objects around them.

What they’re doing

Klarna also used AI to let users search from images. In the simplest version, the user sees something, snaps it, and Klarna turns the image into a shopping search. [13]

This is a small but important pattern: AI is not only generating text. It is translating messy real-world input into structured commercial intent.

Setup

The public description is straightforward: computer vision converts an image into a search term or product-discovery query, then Klarna connects that to products. [13]

Result

The public source gives product-launch evidence, not a detailed ROI number. Still, it shows Klarna applying AI at the very top of the shopping funnel, before the user knows the right words to search.

What to steal

Look for moments where users know what they want visually or contextually but do not know the exact query.

Examples:

• ecommerce: photo to product search;
• B2B SaaS: screenshot to help article or ticket category;
• finance: transaction screenshot to explanation;
• healthcare admin: document photo to form guidance;
• internal ops: whiteboard photo to project plan.

The workflow is input translation: turn fuzzy user context into the right structured next step.

Caveat

Image search only works if the downstream product data is good. The AI can identify a jacket, lamp, or sneaker style, but the value comes from matching that intent to accurate catalog, pricing, merchant, and availability data.

Marketing and creative

Workflow: AI copy and campaign production

Evidence snapshot: Public evidence says AI was used for 80% of copywriting and helped cut marketing agency spend by 25%.

What they’re doing

Klarna’s marketing use case is one of the most copyable parts of the whole case. Public reprints of Klarna’s company release describe a Copy Assistant used for a large share of copywriting, hundreds of internal GPTs, more than a dozen AI-driven marketing projects, and over 100 AI projects across the organization. [6]

The business logic is simple: marketing is full of repeatable production work. Copy variants, translations, product descriptions, campaign concepts, image generation, landing-page drafts, CRM copy, audience-specific messages, and creative testing all benefit from faster first drafts.

Klarna reportedly used AI to reduce marketing agency spend by 25% while running more campaigns. [6]

Setup

The public evidence points to:

• internal GPTs built by employees;
• a Copy Assistant for copywriting;
• AI image-generation workflows;
• multiple marketing AI projects;
• later, Google Cloud generative media models, including Veo and Gemini image models, for lookbooks and personalized campaigns. [5][6]

This looks less like one centralized marketing chatbot and more like an internal tool portfolio.

Result

The clearest marketing result is the reported 25% cut in agency spend while increasing campaign output. [6]

The later Google Cloud partnership adds a more customer-facing metric: early AI creative pilots reportedly increased time spent in the app by 15% and orders by 50%. [5]

Those are strong claims, but they should be treated as company/vendor-reported until independently validated.

What to steal

Do not ask AI to “do marketing.” Pick one repeated production bottleneck.

A practical workflow:

1. Choose one recurring content format: lifecycle emails, paid social variants, product copy, weekly promo pages, sales one-pagers.
2. Create a source packet: audience, offer, product facts, claims allowed, claims forbidden, examples of good copy, brand voice.
3. Ask AI for variants, not final copy.
4. Add review gates for legal, brand, regulated claims, or pricing.
5. Measure cycle time, number of variants shipped, approval rate, and campaign performance.

Klarna’s lesson is not “fire the agency and let AI cook.” It is: if your team can standardize the inputs and review criteria, AI can compress the first-draft and variant-production layer.

Caveat

Marketing is where AI can quietly create legal and brand risk. A model can generate plausible claims faster than your review process can catch them. The workflow needs claim libraries, approved proof points, and review stages.

Engineering

Workflow: coding copilots connected to daily software work

Evidence snapshot: Klarna disclosed that 85% of engineers connected an AI copilot to work software for creating and reviewing code.

What they’re doing

Klarna’s F-1 says that, as of October 2024, 85% of its engineers had connected an AI copilot to their work software for creating and reviewing code. [3]

That is a meaningful signal. Many companies have engineers experimenting with coding tools. Klarna disclosed copilot adoption in a public filing as part of its productivity story.

The work changed from “engineers write and review everything from scratch” toward “engineers use AI for draft code, review assistance, and likely scaffolding.” The human still owns architecture, judgment, merge decisions, and production responsibility.

Setup

The filing does not name the specific coding copilot. It describes the function: create and review code. [3]

The more important setup detail is organizational: AI copilots were not a side experiment. They were connected to the work software used by most engineers.

Result

Klarna ties AI use to productivity and revenue-per-employee gains, but does not isolate the exact ROI of engineering copilots. [3]

That is normal. Engineering copilot ROI is hard to measure cleanly because the output is not just lines of code. The more useful metric may be cycle time on small changes, test scaffolding, PR review speed, documentation speed, and reduced toil.

What to steal

If you manage engineers, do not measure AI coding tools only by “how much code did it write?”

Use a workflow lens:

• boilerplate generation;
• test scaffolding;
• migration scripts;
• code explanation;
• PR review checklists;
• debugging hypotheses;
• documentation drafts;
• incident-summary drafts.

Then keep senior engineers responsible for architecture and review. AI should compress the tedious parts of engineering, not lower the standard for production code.

Caveat

Public sources do not show Klarna’s engineering eval framework, security review, coding standards, or incident guardrails for AI-generated code. Copy the adoption pattern, not the unstated governance.

Internal operations and knowledge work

Workflow: Kiki and internal knowledge retrieval

Evidence snapshot: Kiki had answered 250,000+ inquiries, around 2,000 per day, in public company-reported evidence.

What they’re doing

Klarna operates an internal knowledge chatbot called Kiki. The F-1 says Kiki helps employees find information in real time across internal systems. [3]

That sounds boring. It is not.

A huge amount of workplace time disappears into questions like:

• Where is the latest policy?
• What does this acronym mean?
• Who owns this process?
• What did we decide last quarter?
• Which template do I use?
• What is the current approval rule?

If Kiki answers even a slice of those questions, it is an internal operating system, not a novelty chatbot.

A reprinted company release said Kiki had responded to more than 250,000 inquiries, around 2,000 per day. [7]

Setup

The public description says Kiki searches internal systems in real time. [3]

The exact connectors are not public. But the implied setup is the common enterprise AI knowledge workflow: connect approved internal sources, answer in a conversational interface, and make the answer easier to find than asking a coworker.

Result

Klarna links internal AI applications to higher average revenue per employee in its filing. [3]

We should be careful here. Kiki is one piece of a broader AI push; the filing does not prove Kiki alone caused revenue-per-employee gains. But the pattern is still useful: internal knowledge search is one of the easiest ways to make AI broadly useful without asking every team to invent a workflow from scratch.

What to steal

Start with the questions employees already ask in Slack, Teams, Notion, Google Drive, Confluence, or email.

A simple version:

1. Export the top repeated internal questions.
2. Identify the source of truth for each answer.
3. Clean the docs before building the bot.
4. Require citations or links back to the source.
5. Log unanswered questions as documentation gaps.

The real value is not only faster answers. It is finding out which parts of the company are undocumented.

Caveat

An internal chatbot is only as good as the source material and permissions behind it. If the docs are stale, the bot scales confusion. If permissions are sloppy, it creates security risk.

Workflow: legal document review

Evidence snapshot: Klarna’s F-1 says legal teams use AI to expedite document review.

What they’re doing

Klarna’s F-1 says its legal teams use AI to expedite document review. [3]

There is not much more public detail than that. Still, the workflow is obvious enough to be useful: legal teams spend large amounts of time reviewing drafts, contracts, policies, disclosures, and document changes. AI can summarize, compare, extract clauses, flag missing terms, and draft review notes.

Setup

The exact tool is not disclosed. OpenAI’s customer story says Legal had high adoption of generative AI tools, and Klarna’s F-1 separately names legal document review. [2][3]

Result

No specific legal-team time-saved number is public in the research. The value should be treated as directional.

What to steal

Use AI for legal triage and review prep, not final judgment.

A practical pattern:

• summarize document changes;
• identify clauses that differ from standard language;
• create a risk checklist;
• draft questions for counsel;
• route anything ambiguous to a lawyer.

Caveat

Do not let AI become the legal decision-maker. In regulated or contractual contexts, the safe role is reviewer-assist, not autonomous approval.

Risk, underwriting, fraud, and compliance

Workflow: ML-enhanced real-time underwriting

Evidence snapshot: Klarna’s F-1 says ML supports high-frequency, large-scale, real-time underwriting.

What they’re doing

Klarna is a payments and credit business, so AI in risk is not a side project. Its F-1 says ML supports high-frequency, large-scale, real-time underwriting, helping drive conversion and minimize credit losses. [3]

This is a different kind of AI from the generative assistant. It is decisioning infrastructure. Every transaction can carry risk signals: user history, merchant context, purchase amount, device data, repayment patterns, fraud signals, and affordability constraints.

The workflow is not “ask a chatbot if this customer should get credit.” It is machine learning inside a real-time underwriting system.

Setup

The F-1 points to a single cloud-based technology platform, proprietary data, SKU-level data, and billions of data points. [3]

That data layer is the moat. The model matters, but the bigger advantage is the live transaction, merchant, consumer, product, and repayment context Klarna can use.

Result

The public filing connects ML underwriting to conversion and credit-loss goals. It does not give a clean isolated metric for the ML contribution. [3]

What to steal

For risk teams, the copyable idea is not “use ML to approve things automatically.” It is:

• identify high-volume decisions;
• define the cost of false positives and false negatives;
• use models to score or prioritize;
• keep audit trails;
• monitor for bias, drift, and disparate impact;
• preserve human escalation for edge cases.

Caveat

Klarna explicitly discloses fair-lending risk, including the possibility that ML underwriting could create disparate impact on protected groups. [3]

That caveat is the point. AI in risk can improve speed and precision, but it also increases the need for monitoring, documentation, and governance.

Workflow: graph neural networks for fraud and AML

Evidence snapshot: Google Cloud says Klarna will train and deploy graph neural networks for fraud and anti-money-laundering detection.

What they’re doing

The Google Cloud partnership adds a more technical risk clue. Klarna said it would use Google Cloud AI hardware and expertise to train and deploy graph neural networks for fraud and anti-money-laundering work. [5]

That makes sense. Fraud is relational. One transaction may look normal by itself, but suspicious when connected to devices, accounts, merchants, cards, addresses, and behavior patterns.

Graph models are designed for that kind of relationship analysis.

Setup

The visible setup is Google Cloud’s AI stack plus Klarna’s platform data. The stated goal is to detect anomalies and suspicious patterns across users, transactions, and devices. [5]

Result

The source describes the intended security use case, not a measured fraud-reduction result.

What to steal

If your risk problem depends on relationships, do not only model rows in a spreadsheet. Map the network.

Examples:

• fraud rings;
• duplicate accounts;
• unusual vendor relationships;
• collusive behavior;
• account takeovers;
• suspicious transaction paths.

The AI workflow is not just classification. It is relationship detection.

Caveat

Graph models can be powerful and hard to explain. In regulated contexts, the monitoring and explanation layer matters as much as the detection layer.

Executive strategy and organization design

Workflow: company-wide AI adoption as operating leverage

Evidence snapshot: Klarna reported that 96% of employees used generative AI daily as of August 31, 2024.

What they’re doing

Klarna’s most important AI move may be cultural and organizational: it made AI usage widespread enough that the company could talk about it in investor-facing language.

The F-1 says that as of August 31, 2024, 96% of employees used generative AI daily based on internal data from OpenAI and internal tools. [3]

OpenAI’s customer story separately said 90% of Klarna employees were using OpenAI-powered generative AI tools daily, with high adoption in communications, marketing, and legal. [2]

This matters because most companies have isolated pilots. Klarna appears to have pushed AI into the default work environment.

Setup

The ingredients visible from public evidence:

• ChatGPT Enterprise broadly available;
• OpenAI API usage;
• internal knowledge chatbot Kiki;
• employee-built GPTs;
• engineering copilots;
• marketing assistants;
• product assistants;
• AI in legal and operations;
• leadership framing AI as central to productivity and operating leverage. [2][3][6]

Result

Klarna publicly ties AI to internal efficiency, vendor-cost reduction, support cost reduction, and revenue per employee. [1][3][6][14]

The org-design clue is especially interesting: Klarna said it raised tech employees from 36% of its workforce in 2022 to 52% in Q1 2025. [14]

That suggests AI was not just layered on top of the old org. Klarna is describing a leaner, more technical operating model.

What to steal

A company-wide AI rollout needs both permission and patterns.

Permission: employees need approved tools, security boundaries, and leadership saying usage is expected.

Patterns: teams need examples of what good AI use looks like in their function.

A useful internal rollout could look like this:

1. Give every team access to approved AI tools.
2. Create function-specific example workflows.
3. Ask teams to publish their best internal prompts or automations.
4. Track adoption by function.
5. Tie usage to real metrics, not enthusiasm.
6. Keep human review where the output affects customers, contracts, code, money, or compliance.

Caveat

AI adoption metrics can become vanity metrics. “People used the tool” is not the same as “the work improved.” Klarna is stronger than most because it pairs adoption with support, cost, marketing, and investor-facing productivity claims. But any company copying this should measure workflow outcomes, not login counts.

Agentic commerce

Workflow: making Klarna’s product network readable by AI agents

Evidence snapshot: Klarna says its Agentic Product Protocol standardizes 100M+ products and 400M prices across 13 markets.

What they’re doing

This is the part of Klarna’s strategy that may matter most over the next few years.

If consumers shop through AI assistants, the old ecommerce funnel changes. A user may not search Google, click a merchant page, browse categories, compare products manually, then check out. They may ask an AI assistant: find me the best running shoes under $150 that ship this week and let me pay over time.

Klarna is preparing for that world by making product and price data available to AI systems.

Its Agentic Product Protocol aims to give AI systems a live, structured feed across a large product and price catalog. [8][15]

Its ChatGPT Shopping Search app is powered by Klarna’s Product Search MCP server, which brings real-time product discovery into ChatGPT conversations. [9]

Its Google partnership brings Klarna payments into Google Search, AI Mode, and Gemini through Google Pay. [10]

This is not just “Klarna has an AI assistant.” This is Klarna trying to remain a commerce layer when the shopping interface moves somewhere else.

Setup

The visible setup includes:

• a structured product protocol;
• Klarna’s production catalog;
• live product and price feeds;
• MCP server infrastructure;
• ChatGPT app integration;
• Google Search, AI Mode, Gemini, and Google Pay integrations. [8][9][10][15][16]

Result

It is too early to judge commercial results from these 2025 and 2026 moves. But strategically, they show Klarna trying to own the rails between AI agents, product data, merchants, and payment.

What to steal

If your company has data that agents will need, start turning it into a structured interface now.

Ask:

• What facts do agents need from us?
• What data must be live?
• What actions should agents be allowed to take?
• What should require user confirmation?
• What should require human approval?
• What is our version of an MCP server, API, or structured feed?

The agentic-commerce lesson is bigger than ecommerce. In every industry, the companies with clean, permissioned, structured data interfaces will be easier for AI agents to use.

Caveat

Agentic commerce is still early. Standards may change. Consumer behavior may be slower than vendors expect. But Klarna’s direction is rational: if AI agents become a major discovery surface, payment and product networks need to be agent-readable.

The biggest pattern

Klarna’s AI strategy has a pattern:

Put AI where the company already has data, volume, and measurable outcomes.

That is why support was such a strong first public proof point. Klarna had millions of customer interactions, clear intents, clear resolution metrics, multilingual needs, and a visible cost base. AI could be measured against response time, repeat inquiries, customer satisfaction, and cost.

That is why marketing also makes sense. The company has products, offers, audiences, brand rules, and constant creative demand. AI can draft, vary, translate, resize, and personalize.

That is why underwriting and fraud matter. Klarna has transaction data, repayment data, user signals, merchant data, and real-time decisions. ML can improve speed and risk scoring, but also needs governance.

That is why agentic product infrastructure matters. Klarna has a large product and price graph. If shopping becomes conversational, that data becomes a distribution asset.

The through-line is not “Klarna likes AI.” The through-line is operating leverage:

• reduce repeated support work;
• let employees produce more;
• make shopping more personalized;
• turn product data into agent-readable infrastructure;
• use ML where risk decisions happen at scale;
• shift the workforce toward technical leverage.

For ambitious workers, the career lesson is uncomfortable but useful. The workflows becoming table stakes are not abstract. They are the recurring parts of real jobs: support triage, campaign drafts, code review, document review, internal search, product comparison, risk scoring, and customer follow-up.

If your work includes one of those loops, someone at another company is already figuring out how to redesign it.

What to steal by role

If you’re in customer support: Map your top support intents by volume and repeat-contact rate. Pick one where policies are clear and outcomes are measurable. Use AI to draft or resolve, but keep escalation paths visible.

If you’re in marketing: Build a source packet for one repeatable campaign format. Include product facts, approved claims, banned claims, tone examples, audience segments, and review rules. Use AI for variants and first drafts, not unchecked publishing.

If you’re an engineer: Do not just “use a copilot.” Choose specific engineering loops: tests, migrations, PR review, documentation, incident summaries, debugging notes. Measure cycle-time improvements and review quality.

If you’re in legal or compliance: Use AI to accelerate document review and policy lookup, but keep judgment with humans. Your highest-value workflow is often summarization plus risk-flagging, not autonomous approval.

If you’re in risk or finance: Look for high-volume decisions where models can score, rank, or flag cases. Build monitoring for bias, drift, and edge cases from the beginning.

If you’re a product manager: Ask where your product data can become conversational. What does the user currently compare manually? What information do they need before deciding? What live data would make an assistant genuinely useful?

If you manage a team: Do what Klarna appears to have done internally: make approved tools available, collect team-specific workflows, and tie AI usage to business outcomes. Do not celebrate usage alone.

If you’re a founder/operator: The Klarna pattern is to combine cost-side AI and revenue-side AI. Automating support is useful. Using AI to create new discovery, conversion, and distribution surfaces is more strategic.

What not to overcopy

Klarna is unusually positioned. It has a huge consumer base, hundreds of thousands of merchant relationships, payments data, product data, risk data, and enough scale for small efficiency gains to be worth real money.

A smaller company should not copy the scope. Copy the sequencing.

Start with:

1. a repeated workflow;
2. a measurable bottleneck;
3. approved data sources;
4. a human review path;
5. a metric that proves whether AI made the work better.

The most dangerous imitation is copying the ambition without the control system.

Klarna’s own disclosures point to that. It talks about human involvement in AI training and monitoring, and it discloses risks around ML underwriting and disparate impact. [3]

That is the adult version of AI adoption. The more real the workflow, the more real the governance needs to be.

Sources and receipts

Core support automation

[1] Klarna / PRNewswire, Klarna AI assistant handles two-thirds of customer service chats in its first month, February 27, 2024.
https://www.prnewswire.com/news-releases/klarna-ai-assistant-handles-two-thirds-of-customer-service-chats-in-its-first-month-302072740.html

[2] OpenAI, Klarna’s AI assistant does the work of 700 full-time agents, customer story.
https://openai.com/index/klarna/

SEC / investor evidence

[3] Klarna Group plc, Form F-1 registration statement, March 14, 2025.
https://www.sec.gov/Archives/edgar/data/2003292/000162828025012824/klarnagroupplcf-1.htm

[14] Klarna, Klarna accelerates global momentum in Q1 2025 and unlocks large gains from AI innovation, May 19, 2025.
https://www.klarna.com/international/press/klarna-accelerates-global-momentum-in-q1-2025-and-unlocks-large-gains-from/

Product and shopping AI

[4] Klarna, Shopping made smarter: Klarna adds more AI features to its assistant powered by OpenAI, September 19, 2024.
https://www.klarna.com/international/press/shopping-made-smarter-klarna-adds-more-ai-features-to-its-assistant-powered-by-openai/

[11] Klarna, Klarna brings smoooth shopping to ChatGPT, March 23, 2023.
https://www.klarna.com/international/press/klarna-brings-smoooth-shopping-to-chatgpt/

[12] Klarna, Klarna is the only bank among ChatGPT Enterprise launch customers, August 30, 2023.
https://www.klarna.com/international/press/klarna-is-the-only-bank-among-chatgpt-enterprise-launch-customers/

[13] Klarna, Klarna taps AI for human shopping experience in a major autumn product launch, October 11, 2023.
https://www.klarna.com/international/press/klarna-taps-ai-for-human-shopping-experience-in-a-major-autumn-product-launch/

Marketing and internal adoption

[6] Raptor Group reprint, AI helps Klarna cut marketing agency spend by 25% and run more campaigns, May 28, 2024.
https://www.raptorgroup.com/news/ai-helps-klarna-cut-marketing-agency-spend-by-25-and-run-more-campaigns/

[7] Financial IT reprint, 90% of Klarna Staff Are Using AI Daily - Game Changer for Productivity, May 14, 2024.
https://financialit.net/news/artificial-intelligence/90-klarna-staff-are-using-ai-daily-game-changer-productivity

Google Cloud, fraud, and creative AI

[5] Google Cloud, Klarna and Google Cloud Enter Strategic AI Partnership to Bring More Creative and Engaging Shopping Experiences to Millions of Klarna Users Worldwide, October 9, 2025.
https://www.googlecloudpresscorner.com/2025-10-09-Klarna-and-Google-Cloud-Enter-Strategic-AI-Partnership-to-Bring-More-Creative-and-Engaging-Shopping-Experiences-to-Millions-of-Klarna-Users-Worldwide

Agentic commerce and AI shopping infrastructure

[8] Klarna, Klarna launches Agentic Product Protocol: The open standard that makes 100M+ products instantly discoverable by AI agents, December 15, 2025.
https://www.klarna.com/international/press/klarna-launches-agentic-product-protocol-the-open-standard-that-makes-100m/

[9] Klarna, Klarna launches AI-powered Shopping Search app in ChatGPT, May 20, 2026.
https://www.klarna.com/international/press/klarna-launches-ai-powered-shopping-search-app-in-chatgpt/

[10] Klarna Investors, Klarna’s Flexible Payments Are Coming to Google Search and the Gemini App Within Google Pay, May 12, 2026.
https://investors.klarna.com/News—Events/news/news-details/2026/Klarnas-Flexible-Payments-Are-Coming-to-Google-Search-and-the-Gemini-App-Within-Google-Pay/default.aspx

[15] Klarna Docs, Klarna Agentic product protocol.
https://docs.klarna.com/acquirer/klarna/other-products/klarna-search/klarna-agentic-product-protocol/

[16] Klarna, Klarna backs Google’s Universal Commerce Protocol (UCP) to enable agentic commerce across platforms, February 2, 2026.
https://www.klarna.com/international/press/klarna-backs-googles-universal-commerce-protocol-ucp-to-enable-agentic/

Technical clues

[17] Klarna Incubator GitHub, Agentic Product Protocol (APP).
https://github.com/klarna-incubator/agentic-product-protocol

[18] Klarna GitHub, The Klarna Product Page Dataset.
https://github.com/klarna/product-page-dataset

[19] Klarna Incubator GitHub, mleko: Streamlining Machine Learning Pipelines in Python.
https://github.com/klarna-incubator/mleko

[20] Klarna Incubator GitHub, Gram.
https://github.com/klarna-incubator/gram

Evidence gaps

The public record is strong enough to show Klarna’s AI operating map. It is not strong enough to fully reconstruct the system architecture.

Important missing pieces:

• no clear public breakdown of RAG, vector search, embeddings, or retrieval design;
• no detailed model-evaluation or evals framework;
• no prompt-engineering playbook;
• no full customer-support escalation logic;
• no detailed cost model for AI infrastructure;
• no independent audit of support, marketing, or app-engagement claims;
• limited public job-posting evidence on AI org structure;
• limited disclosure of how internal GPTs are approved, monitored, or retired.

Those gaps do not weaken the main lesson. They make the case study more useful: what we can copy is the workflow pattern, not Klarna’s hidden implementation.

The practical takeaway

If you only steal one idea from Klarna, steal this:

Find the repeated work where your company already has the data, the rules, and the metrics. Put AI there first.

That might be a support queue, a weekly report, a campaign workflow, a code-review loop, a contract-review process, a risk-scoring queue, or a product-search experience.

The companies getting ahead are not waiting for AI to become magical. They are finding the parts of work that are already structured enough for AI to help — then redesigning the workflow around that leverage before everyone else treats it as table stakes.