Back to Explore
NewsVibeFix editorialUpdated Nov 18, 2026

Klarna's New AI Assistant Will Spend Your Money: Buy First, Pay When It Arrives

On 17 November 2026 Klarna launched an AI assistant that books tables, finds babysitters and buys things for you — you do not pay until the order arrives. It lives in WhatsApp and iMessage, starts with a small group on a waitlist, and follows the company's public 2025 climbdown from replacing humans with AI.

A young woman on a city street holding a phone and a payment card, shopping bags on her arm

In short: this time it does not just answer questions — it spends your money

On 17 November 2026, Klarna announced the Klarna AI-assistant, a personal assistant that handles tasks on your behalf: it finds products, compares prices and buys them for you, books a table for two, finds a babysitter, keeps track of deliveries and reservations, drafts email replies for you to approve, files receipts, and flags upcoming birthdays with gift ideas. It does not live only inside the Klarna app. It lives in WhatsApp and iMessage — where you already chat every day. Right now only a small group of early users can try it, and everyone else joins a waitlist. Joining is free, but only a few people are let in at a time. The most striking line in the announcement is this: it finds the best product and price, buys it for you, and you do not pay until it arrives. Buy first, pay when it arrives — Klarna is applying its buy-now-pay-later business to the AI assistant.

This assistant did not come out of nowhere. It builds on Klarna's existing shopping assistant, and the numbers behind that are not small: since 2020, Klarna customers have spent more than $34 billion through the app's built-in browser, even at stores that do not offer Klarna. In other words, Klarna stopped being just a payment button at checkout a long time ago. It is already a place where users browse, compare and buy. The new AI assistant upgrades that place from "you browse it yourself" to "it browses it for you". Klarna co-founder and CEO Sebastian Siemiatkowski stretches the timeline back to 2015 in the announcement: back then they imagined a future in which your phone would buzz to tell you how to save on your mortgage, and do the work for you. He calls this assistant a major step towards that vision. A vision is moving, but a vision is not evidence. The evidence is in how it plans to avoid the mistake it made two years ago.

The earnings report published the same day explains why it is in a hurry to tell this story

Klarna published its third-quarter results on the same day, 17 November, and the two announcements make more sense together. In the third quarter, gross merchandise volume reached $36.1 billion, up 10% year over year, with the US up 24%. Revenue was $1.009 billion, up 12%. Transaction margin dollars — the money left after processing, credit and funding costs — came to $371 million, up 32%. Adjusted operating income was $21 million, up $35 million on a year earlier. The net loss was $65 million, $30 million narrower than the $95 million loss in the third quarter of 2025. Active consumers reached 122 million, up 7% year over year, and merchants reached 1.35 million, up 59%. The CEO's line in the release is that this was the fourth straight quarter with revenue above $1 billion and positive adjusted operating income.

But the other side of the report has to be read alongside it: on a statutory basis the company is still loss-making, with a net loss of $65 million. Adjusted operating income is positive only because costs such as share-based payments are stripped out, and the full-year outlook was reiterated rather than raised, at revenue of $4.08 billion to $4.16 billion. In other words, the business is getting better, but not yet good enough to relax. The AI assistant arrives at this moment to carry the "next growth story": payments and instalments are today's business, an assistant that handles errands is tomorrow's. Two more numbers in the report are worth noting: Fair Financing in the US grew 73% to $3.6 billion, Klarna Card volume doubled, and subscription revenue grew 530%. Users are handing Klarna bigger and more everyday spending, and the assistant is being built to take over the chores attached to that spending.

Bar chart: Klarna Q3 2026 versus a year earlier — transaction margin dollars up from $281 million to $371 million, adjusted operating income from a $14 million loss to a $21 million profit, net loss narrowed from $95 million to $65 million

It has already played the "AI replaces people" story once — what is different in this script?

To understand this launch, you have to put Klarna's famous AI episode on the table first. In February 2024, Klarna announced loudly that its AI assistant had handled 2.3 million conversations in its first month, doing the equivalent work of 700 full-time customer service agents, and cutting average resolution time from 11 minutes to 2 minutes. It became one of the most quoted "AI replaces humans" cases in the industry, held up as a template everywhere. Then in May 2025, Siemiatkowski told Bloomberg in an interview that he was walking it back: cost had been too predominant a factor in how the rollout was judged, quality had dropped as a result, and the company was hiring humans again, piloting an Uber-style model with remote agents, students and Klarna's own users.

The lesson of that episode is not "AI does not work". It is "they measured the wrong thing". The 2024 announcement was all cost and speed: how many conversations were handled, how many people were saved, how many minutes were cut. What arrived late was the quality signal — the complicated cases that need judgement were brushed aside, and user frustration only showed up months later in complaints and churn. By the time the company reacted, the money saved had been paid back several times over in reputation. That is why the wording of this new announcement deserves to be read word by word: it repeats that the assistant keeps users "firmly in control", email drafts wait for approval before sending, and the launch is limited to a small group with a waitlist that opens slowly. Behind the wording is a company that has been burned once, signalling that this time the brakes go on before the speed.

But promises are promises, and three things the announcement does not say are exactly what will decide whether this succeeds. First, it does not say whose model powers the assistant, or which decisions it can make alone and which need the user to sign off. Booking a table for two and buying something worth hundreds of dollars are two completely different risk levels, yet the announcement lists them side by side without drawing the line. Second, it does not say in which countries it will be available. Klarna's users span many markets, but "spending money for you" looks different under the payment rules and consumer protections of each one. Third, it says nothing about limits or what happens when things go wrong: if the assistant books the wrong time or buys the wrong model, who handles the refund and the argument? None of this proves the product is flawed. It only proves it is too early to judge — a product open to a small group of users has not reached the point of being tested yet.

"You don't pay until it arrives": that slogan is the real product

Take the most advertising-like sentence in the announcement and look at it on its own, because it is actually the most concrete product definition in the whole release. The Klarna assistant finds the best product and price, buys it for you, and you do not pay until it arrives. What makes that sentence possible is not how smart the assistant is, but Klarna's old business: credit and risk. The earnings report carries a number that underwrites exactly this promise — provisions for credit losses were 0.56% of gross merchandise volume in the third quarter, down from 0.72% a year earlier. Klarna dares to let an assistant spend your money because it has a worked-out way of absorbing the cost of "bought the wrong thing, returned it, did not pay it back". A chatbot that books the wrong table costs an apology. An assistant that buys the wrong thing costs real money. Who absorbs that cost decides whether this kind of assistant can move from a demo into everyday life.

The tasks listed in the announcement also form a ladder of trust, and it is worth reading them in order. At the bottom: filing receipts and flagging birthdays — read-only, no money moves, the cost of a mistake is smallest. One rung up: drafting emails for approval — it writes, but does not send, and the decision stays with you. Higher still: booking a table, finding a babysitter — now it spends your time and other people's time, and a mistake needs an apology. At the top: buying things for you — moving money directly, and money Klarna fronts first. The babysitter example deserves an extra second of thought: it means trusting an assistant with your children's schedule, which is no longer a question of convenience but of trust. Klarna lists all four rungs in one announcement, which amounts to a public promise that its assistant can climb the whole ladder. In practice, most assistants fall off at one rung or another — one that files receipts well does not necessarily book a babysitter well, and one that books a babysitter well does not necessarily buy the right thing. The cost of an error is different at each rung, and the announcement says nothing about whether different confirmation rules apply at each level. That is the detail most worth watching in early user feedback: at which rung does the assistant first make someone think "I will just do this myself"?

This is also where Klarna really stands in the race to build assistants that act for users. Google, Meta and Anthropic are all building assistants that run errands, but none of them is simultaneously the card issuer, the risk manager and the payment collector. Klarna has 122 million active consumers, 3.8 million transactions a day, 1.35 million merchants, and a credit assessment system that has run for years. Its bet is that the moat in the age of agents is not conversation quality but "who pays for the assistant's mistakes". If the assistant buys the wrong thing, Klarna can pay first and absorb it through its system; if it books the wrong trip, Klarna can use its merchant relationships to change it. That ability to stand behind the error is something a pure software company cannot learn quickly. The reverse is equally true: if the cost of standing behind errors spirals out of control, the hole in the accounts will be bigger than the customer service one was — and the 2024 lesson already proved that Klarna is sensitive to cost numbers.

WhatsApp and iMessage: a choice that matters more than the feature list

The easiest sentence to skim past in the announcement is actually the key to the whole product strategy: the assistant lives in WhatsApp and iMessage, not only inside the Klarna app. Klarna has 122 million active users, so its app is hardly short of an audience — yet it still chose to put the assistant in the chat apps people open dozens of times a day. The judgement is the same one Microsoft and Salesforce made by putting the CRM into Copilot and Teams: users will not move house for an assistant; the assistant has to go where users already are. An assistant that books restaurants, tracks deliveries and answers emails will lose to the word "handy" if it first asks you to open a finance app every time.

But living in a chat app also raises the bar for trust. Chatting with the assistant in WhatsApp feels no different from chatting with a friend — except that this "friend" holds your bank card, your shopping history and your calendar. It can see when your birthday is, who you bought gifts for, and which order is being returned. That intimacy is a double-edged sword: used well, it is "the assistant that knows you best"; one mistake — say, spoiling a surprise gift for a partner, or sending an email you meant to write yourself — and the trust lost is worth more than any app crash. Klarna's line about keeping users "firmly in control" carries extra weight in a chat app, because everything there is so smooth that it is easy to forget what you just authorised.

The earnings report adds another set of numbers explaining why Klarna thinks now is the time. Merchants grew 59% in a year to 1.35 million, driven by default-on integrations with the world's largest payment service providers: in October, Clover, Fiserv's small-business platform, went live at 100,000 US in-store locations, and Klarna extended the partnership to Fiserv's Commerce Hub; in Europe, Klarna updated its agreement with Mollie, which serves 350,000 businesses; Southwest Airlines and J.Crew went live in the US, and the Marks & Spencer launch means every major UK department store now offers Klarna. The bigger the merchant network, the wider the range of errands the assistant can run — booking a restaurant, booking a flight, buying clothes, all on the same network. The assistant is not building an ecosystem from zero. It starts on top of a merchant network that is already laid down, adding a layer of "running errands for you" on top. That is its biggest difference from general-purpose assistants that start from nothing: Klarna's assistant is born with a checkout counter attached.

Not your turn yet: a small group, a waitlist, and the things left unsaid

For ordinary readers, the most practical sentence in this story is: it is not your turn yet. The assistant is live with a small group of early users, the waitlist only opened this week, joining is free, and the company says it is letting only a few people in at a time. There is no published timetable for general availability, no list of launch markets, and no word on the model behind it or its operating boundaries. In other words, on 17 November Klarna published a direction and a queue to join, not a product you can download and use today.

Three things are worth watching next. First, whether early users soon post failure stories online: a wrongly booked restaurant, a wrong size, a customer service line nobody answers — the first real feedback will say more about the assistant's level than any launch event. Second, whether the "wait for approval" line gets wider over time: today email drafts need approval; tomorrow, will purchases need approval too? Once the approval step is loosened in the name of a "smooth experience", the risk moves from the assistant to the user. Third, whether Klarna publishes hard numbers: how many tasks the assistant handles each week, how many need human intervention, and what error correction costs. In 2024 it published conversation volumes and headcount saved; this time, if it dares to publish task completion rates and error rates, that will be the real sign it learned the lesson.

One more question deserves to be asked on its own: this assistant is free to join, so how does Klarna make money from it? The answer is in the earnings outlook — full-year revenue is expected to be 2.74% to 2.75% of gross merchandise volume. Every transaction the assistant completes for a user is a transaction Klarna takes a cut of. The busier the assistant gets, and the more everyday spending users hand to it, the faster Klarna's transaction volume grows. It is a clean business model, and it is also an incentive structure to be wary of: an assistant that earns from transaction volume is naturally inclined to make you buy more and buy faster, not buy less and wait. The announcement says the assistant will "find the best product and price" — but who defines "best"? The best deal for you, or the best for Klarna's volume? There is no answer to that yet, but it will decide whether this assistant ends up as a housekeeper that saves you money or a salesperson that is unusually good at making you spend. Once it opens to more users, watch which way its recommendations lean. That will say more than any list of skills.

What this means for people building projects

There are two lessons here for people building AI products. The first is about channels: the assistant does not ask users to download a new app. It moves into WhatsApp and iMessage. That is the same judgement Microsoft and Salesforce made by putting the CRM into Copilot and Teams — users will not move house for an assistant, the assistant has to go where users already are. For developers building their own products, that sentence is worth writing on the wall: however good your assistant is, if it asks users to change their habits, it has already lost half the battle.

The second lesson matters more, and it is about the design of control. Klarna's 2024 story and its 2026 story are two sides of the same coin: in 2024 it replaced people with AI and measured cost; in 2026 it lets AI spend people's money and stresses control. Between the two lies a public admission of error, and the company's product language has visibly changed — from "how much can it do for you" to "it has to ask you first". For any project that gives an agent real actions — spending money, sending messages, changing data — that order is worth copying exactly: put approval, rollback and a backstop into the product first, and only then talk about how capable it is. Otherwise, every penny you save will eventually be paid back several times over on the trust account.

Views 0Comments 0

Comments (0)

ME
0/1000
Loading comments...

Sources

Browse projectsPublish your project

Related articles

A young woman seals a cardboard shipping box with a red tape gun in a bright home office, shelves of shipping boxes behind her, a laptop on the desk
News
The Code Is Done. The Business Is Just Starting: Replit Agent Can Now Buy Supplies and Run Your Ads

On 12 November 2026, Replit gave Replit Agent two new jobs: buying supplies for your business and running your Google Ads. Purchases go through Stripe's Link wallet and Visa's rails; ads go through Google Ads. The common thread: the agent can only prepare the work — money moves only when the founder approves, with checks enforced in code, not left to the model.

Payments & MonetizationProduct LaunchAI Coding
A developer with glasses types on a laptop late at night at a wooden desk, a large monitor behind showing colourful code, blurred city lights outside the window
News
Run the Middle Layers Three Times: Inkling 2 Mini Puts a 26B Coding Model on Your Laptop

On 20 November 2026, Thinking Machines Lab released Inkling 2 Mini: a 26.4B-parameter looped transformer that runs its middle 12 layers three times per token, fits in about 20GB, and runs offline on a laptop. It leads open-weights models of its size on DeepSWE and SWE-bench Verified — with the caveats printed in the same announcement.

Product LaunchModel UpdatesAI Coding
A developer with glasses types on a laptop at a wooden desk, an external monitor to the left showing C code on a dark background, a small plant on the desk
News
Microsoft's Coding Model Stops Playing Second Fiddle: MAI Code 2 Scores 96% of Opus 5.5 for 5% of the Cost — Read the Three Caveats First

On 17 November 2026, Microsoft launched MAI Code 2 and Code 2 Flash: the larger model has 77 billion active parameters and a one-million-token context, averaging 63.2% across four benchmarks — close to GPT-6.1 Sol and Claude Opus 5.5 — at $0.309 per task. Read the three caveats first: promotional pricing, the lowest score of the three, and no long-horizon work.

Developer WorkflowProduct LaunchModel Updates