What happens in the two seconds after you tap “Pay in 4” at checkout? Somewhere in that pause, a model decides whether you are a good risk. Klarna and Affirm both make that decision in near real time, and both lean heavily on automated underwriting. But they are not the same machine underneath, and the gap matters in 2027 as new AI rules, shifting regulators, and very different business models reshape the buy now, pay later market.
This comparison looks at how each company’s underwriting works, what is confirmed versus inferred, and what borrowers and merchants should watch for.
The two companies at a glance
Klarna is a Swedish fintech with a global consumer base. Affirm is a US-listed lender that reports to the SEC. That single difference changes how much the public can actually verify.
| Factor | Klarna | Affirm |
|---|---|---|
| Home base | Sweden | United States |
| Public filings | Limited (private company) | Detailed (SEC 10-K) |
| Underwriting style | Point-of-sale, proprietary models plus bureau data | Transaction-by-transaction, proprietary models plus bureau data |
| Product shape | Revolving and installment options | Installment loans per purchase, no revolving line |
| Late fees | Varies by market | Company states it does not charge late fees |
| Transparency | Press releases and interviews | Regulatory filings and investor materials |
The transparency gap is the first practical lesson. If you want to understand Affirm’s credit losses, you can read its filings. With Klarna, you mostly get company statements.
What “AI underwriting” really means at checkout

Underwriting is the process of estimating whether a borrower will repay. In BNPL, that estimate happens in milliseconds at the point of sale. The model looks at signals like purchase size, merchant category, prior repayment behavior with the platform, and data from credit bureaus.
The word AI gets stretched here. Some of it is machine learning. Some is rule-based scoring dressed up in modern language. Neither company publishes its model architecture, so any claim about specific algorithms is inference, not confirmed fact.
What is confirmed: Klarna’s own February 2024 press release said its OpenAI-powered assistant handled 2.3 million conversations in its first month, equal to the work of about 700 full-time agents. That is customer service, not underwriting, but it shows how aggressively the company pushes automation across operations. Affirm, meanwhile, has consistently described underwriting every transaction individually rather than extending a revolving line. That design choice means each purchase gets its own risk decision.
Why the per-transaction model matters
A revolving credit line lets risk build up quietly. A per-purchase model forces a fresh decision every time. In practice, that tends to mean Affirm approves smaller amounts more conservatively and declines more often at the margins. For borrowers, it can feel stricter. For lenders, it limits exposure.
How each model handles risk signals
Both companies combine their own data with bureau data. The differences are in emphasis.
Klarna has historically underwritten at the point of sale and has publicly discussed using transaction and repayment data across its large consumer base. That data advantage is real, but it is also a concentration risk: models trained on your own customers can drift when the customer mix changes.
Affirm’s disclosures in its SEC filings are the best primary source for its underwriting and credit-loss approach. Its filings are public through EDGAR, and anyone comparing the two companies should read them directly rather than rely on summaries.
A practical tip: when you see a headline claiming one company’s AI “predicts default better,” check whether the claim comes from a filing, a regulator, or a marketing page. Most do not survive that test.
The regulatory picture that changes everything in 2027
This is where the comparison gets genuinely interesting, and where most articles get it wrong.
In the US, the CFPB finalized an interpretive rule in May 2024 treating BNPL lenders like credit card issuers under Regulation Z. A federal court vacated that rule in January 2025. The status as of late 2026 is uncertain, and anyone writing about 2027 should verify the current position before relying on either outcome.
In the UK, the government confirmed in October 2024 that it would legislate to regulate BNPL under the FCA. The timing has slipped repeatedly, which is itself a signal about how hard this product is to classify.
In the EU, the Consumer Credit Directive (2023/2225) must be transposed by 20 November 2025, bringing BNPL into scope. Transposition status varies by member state.
The AI-specific layer is the one to watch. The EU AI Act (Regulation 2024/1689) classifies creditworthiness assessment as high-risk. Obligations for that use case apply from 2 August 2026. That means by 2027, any BNPL lender assessing creditworthiness with AI in the EU faces documented risk-management, data-governance, and human-oversight duties. Both Klarna and Affirm operate in Europe, so both are exposed.
For readers tracking how regulation reshapes AI products more broadly, the same pattern shows up elsewhere: rules arrive late, then land all at once. The comparison between Gemini and ChatGPT on output quality follows a similar arc, where evaluation standards lag behind deployment.
What borrowers should actually check
Underwriting models are invisible, but their outputs are not. A few habits separate informed BNPL users from surprised ones.
- Read the repayment schedule before confirming. Installment plans differ in how missed payments are handled.
- Check whether late fees apply. Affirm states it does not charge them. Klarna’s terms vary by market and product.
- Watch for stacking. Using several BNPL providers at once hides total debt from each individual model.
- Know your dispute path. BNPL providers are not banks, and complaint routes differ.
- Re-check after regulatory changes. Terms that were standard in 2025 may not survive 2027.
The most common mistake is treating approval as a signal of affordability. It is not. A model approves based on probability, not on whether the payment fits your month.
Where each company is stronger
Klarna’s strength is scale and integration. It is embedded in tens of thousands of merchants and has deep consumer familiarity in Europe. Its automation push, evidenced by the 2024 assistant rollout, suggests a company willing to cut operational costs aggressively.
Affirm’s strength is disclosure and product discipline. Its per-transaction model and public filings make it easier to evaluate. Its stated no-late-fees position is a clear consumer-facing commitment.
Neither advantage is permanent. Underwriting quality depends on data freshness, macroeconomic conditions, and how fast each company adapts when default rates move.
What the 2027 comparison really comes down to
Strip away the marketing and three questions remain.
First, who has better data? Klarna’s consumer base is large and global. Affirm’s transaction-level data is deep per purchase. Both are valuable in different ways.
Second, who is more exposed to regulation? Both face the EU AI Act’s high-risk creditworthiness rules from August 2026. Affirm’s US listing means its compliance posture is more visible. Klarna’s private status means less is disclosed.
Third, who adapts faster? Automation cuts costs, but it also removes human judgment from edge cases. A model that handles 90 percent of decisions well can still fail badly on the 10 percent, and those failures are where regulators focus.
If you are comparing tools that make automated judgments, the same logic applies to writing assistants. The Claude versus ChatGPT draft quality comparison shows how output quality depends on context rather than model branding alone. Underwriting works the same way: context decides outcomes.
Practical guidance for merchants
Merchants care about conversion, not underwriting philosophy. But the two are linked. A stricter model declines more customers and lowers conversion. A looser one raises sales and credit losses.
Three practical checks before signing with either provider:
- Ask about approval rates by category. A model tuned for fashion behaves differently on electronics.
- Ask who bears fraud loss. The answer varies by contract, not by company reputation.
- Ask about integration depth. Deeper integration usually means better data sharing, which improves model accuracy over time.
Merchants often discover these details only after a quarter of disappointing conversion. Asking upfront is cheaper.
How to evaluate claims about AI underwriting
A simple framework helps separate signal from noise.
| Claim type | Reliability | What to do |
|---|---|---|
| SEC filing language | High | Read the original document |
| Regulator publications | High | Check the date and current status |
| Company press releases | Medium | Treat as a claim, not an audit |
| Journalist summaries | Medium to low | Trace back to the primary source |
| Vendor marketing pages | Low | Verify independently |
This matters because BNPL underwriting claims are frequently repeated without attribution. A statement like “our AI approves more good borrowers” is unverifiable without data. A filing that discloses credit-loss rates is verifiable.
The same discipline applies to any AI comparison. When evaluating DeepSeek, ChatGPT, and Gemini as output judges, the useful question is not which model sounds confident but which one can be checked against a standard.
The bigger trend: underwriting moves from credit to context
Both companies are part of a wider shift. Lenders increasingly use behavioral and transactional signals alongside traditional credit data. That can widen access for people with thin credit files, which is a genuine benefit. It can also encode bias if the signals correlate with protected characteristics.
The EU AI Act’s high-risk classification exists precisely because of that tension. Creditworthiness assessment affects people’s access to goods and services, so regulators want documentation, testing, and human review. By 2027, expect both Klarna and Affirm to publish more about model governance than they do today, not because they want to, but because they will have to.
There is a useful parallel in infrastructure. Just as WireGuard and OpenVPN protect data only up to the point where their limits begin, AI underwriting improves decisions only up to the point where data quality and regulatory oversight run out. Beyond that, the model is guessing.
Conclusion
Klarna and Affirm both underwrite buy now, pay later purchases with automated models, and both are heading into a 2027 landscape shaped by the EU AI Act’s high-risk creditworthiness rules. The differences are in structure: Affirm underwrites each transaction individually and discloses more through public filings, while Klarna operates at greater scale with less public visibility into its models. Neither company’s underwriting architecture is fully disclosed, so any detailed claim about their algorithms should be treated as inference.
For borrowers, the practical advice is unchanged: read the repayment terms, watch for stacked BNPL plans, and never treat approval as proof of affordability. For merchants, ask about approval rates and fraud liability before signing. For anyone comparing the two, start with primary sources and check dates, because the regulatory picture in the US shifted twice in under a year.
The real 2027 story is not which company has smarter AI. It is which one can prove its AI is fair, documented, and accountable when regulators come asking.
FAQ
Is Klarna’s underwriting fully automated?
Klarna has not published its model architecture, so the exact level of automation is not confirmed. The company underwrites at the point of sale using proprietary models plus bureau data, and it has pushed automation hard in customer service. Treat any claim about a fully automated decision chain as inference.
Does Affirm charge late fees?
Affirm has stated that it does not charge late fees. Because it underwrites each transaction individually rather than extending a revolving line, its risk model differs from a traditional credit card. Always confirm current terms directly, since product details can change.
How does the EU AI Act affect BNPL underwriting?
The EU AI Act classifies creditworthiness assessment as high-risk. Obligations for that use case apply from 2 August 2026, which means by 2027 BNPL lenders operating in the EU must meet documented risk-management, data-governance, and human-oversight requirements.
What happened to the CFPB’s BNPL rule?
The CFPB finalized an interpretive rule in May 2024 treating BNPL lenders like credit card issuers under Regulation Z. A federal court vacated it in January 2025. The status as of late 2026 is uncertain, so verify the current position before relying on either outcome.
Can I use Klarna and Affirm at the same time?
Technically yes, but it is risky. Each provider’s model sees only its own repayment history with you, so stacking plans across providers hides your total debt from every individual model. That is one of the most common ways BNPL users get into trouble.
Which company discloses more about its underwriting?
Affirm discloses more, because it is a US-listed company that files with the SEC. Its filings cover underwriting approach and credit losses in detail. Klarna is private and shares information mainly through press releases and interviews, which are company claims rather than audited disclosures.


