Fraud detection, credit underwriting, and algorithmic trading all use "AI" today — but the maturity, autonomy, and regulatory scrutiny at each stop are wildly different, and the most-hyped win (Klarna's chatbot) had to be walked back.
Every time a Mastercard-branded card is used online or in a store, a machine learning model scores the transaction before the authorization response goes back to the merchant. That model is the descendant of Brighterion, an AI fraud-detection startup Mastercard acquired in 2017 and folded into what it now calls Decision Intelligence. Mastercard says the platform has helped block tens of billions of dollars in confirmed fraud since launch, and issuers like Worldpay have reported sustained year-over-year fraud reductions attributable to it. This is AI in finance at its most mature: a narrow, well-defined prediction problem, running continuously, with a clear dollar-denominated success metric.
That maturity is not evenly distributed. "AI in finance" gets used as a single phrase to cover four very different jobs — catching fraud, deciding who gets a loan, generating trading signals, and answering customer support tickets — and they sit at wildly different points on the curve from experimental to load-bearing. Some of that difference is technical. Most of it is regulatory: the areas where AI is furthest along are the areas where a wrong answer costs money, not a lawsuit.
Fraud scoring was one of the first places banks and payment networks put machine learning into a live decision path, and it remains the deepest deployment. Stripe Radar, PayPal's in-house models, and Mastercard's Decision Intelligence all do roughly the same job: score a transaction in real time using features like device fingerprint, IP geolocation, spending velocity, merchant category, and — increasingly — graph features that link a card, device, or account to other entities that have already been flagged.
The graph piece matters more than the marketing copy suggests. A single suspicious transaction is weak signal; a device that has touched twelve different stolen cards in the last hour is strong signal, and that pattern only shows up once you're modeling the network of relationships between entities, not just scoring transactions independently. Mastercard's more recent "Decision Intelligence Pro" leans on exactly this kind of graph modeling combined with generative techniques to build richer entity profiles.
The limitation isn't accuracy — these systems are good, and have been for years. It's the asymmetry of the cost function. A missed fraud case costs the issuer money; a false positive costs a legitimate customer a declined card at checkout, which is its own kind of expensive (cart abandonment, customer churn) and much harder to see in a model's training loss. Every fraud team is quietly trading off these two failure modes, and the model doesn't know which one your business cares about more this quarter. Fraud patterns also drift fast — the moment a scoring rule becomes effective, fraud rings adapt around it, so these systems need continuous retraining, not a one-time deployment.
Underwriting is structurally similar to fraud scoring — both are binary classification problems on transaction or applicant data — but it operates under the Equal Credit Opportunity Act (ECOA) and Regulation B, which require lenders to give applicants specific, actionable reasons when they're denied credit. That single requirement is why underwriting AI has moved so much more cautiously than fraud AI.
Upstart is the clearest case study here. In September 2017, the Consumer Financial Protection Bureau issued its first-ever no-action letter to Upstart, giving the company limited assurance it wouldn't face ECOA enforcement while using alternative data (education, employment history, and more) in its underwriting model. In 2020 the CFPB issued a second letter, this time describing a model that incorporated over 800 variables trained on more than 9 million repayment events. Then in June 2022, Upstart itself asked the CFPB to terminate that letter after notifying the agency it planned to add a significant number of new variables — changes big enough that Upstart preferred to drop the special regulatory status rather than wait for CFPB review. Zest AI (formerly ZestFinance) has pursued a similar alternative-data approach for other lenders, packaged with explainability tooling specifically built to satisfy adverse-action notice requirements.
The recurring tension is proxy discrimination: alternative data features (which zip code you live in, what school you attended, how you use a smartphone) can correlate with race or other protected characteristics even when the model is never given that attribute directly. A model can be accurate and still be functionally discriminatory, and "the algorithm decided" is not a defense under fair lending law. That's why underwriting AI, unlike fraud AI, comes bundled with a parallel industry of explainability tooling (SHAP values, counterfactual explanations, adverse-action reason codes) that exists purely to satisfy a legal requirement the model architecture doesn't care about on its own.
This is the category most misunderstood by the term "AI in finance." Quantitative trading firms — Renaissance Technologies, Citadel, Two Sigma — have used statistical and machine learning models to generate trading signals for decades, well before "AI" was the industry's preferred label. Rules-based algorithmic execution (breaking a large order into smaller pieces to minimize market impact, routing between venues for best price) is older still, and it's the category that produced the May 6, 2010 Flash Crash, when a large automated sell algorithm interacting with high-frequency trading systems briefly wiped out roughly a trillion dollars in market value in minutes before prices recovered. That event is why U.S. markets now have circuit breakers and limit-up/limit-down rules — a direct regulatory response to automated systems moving faster than human oversight could follow.
What's newer is large language models entering the research and analysis layer rather than execution. Bloomberg trained BloombergGPT, a domain-specific LLM for financial text, and announced it in 2023; the pitch was better sentiment analysis and document summarization over financial filings and news, not autonomous trade decisions. That distinction holds broadly across the industry: LLMs show up in research assistance, report summarization, and signal generation as one input among many, but very few firms let a generative model place trades directly and unsupervised. The actual decision layer at most quant funds is still statistical models with tight risk limits, extensive backtesting, and human sign-off on strategy changes — for the same reason underwriting keeps a human in the loop: the cost of a bad autonomous decision (a flash-crash-style cascade, a wrong-way bet sized too large) is asymmetric and hard to undo.
The honest failure modes here are backtest overfitting (a strategy that looks great on historical data because it accidentally learned the noise, not the signal) and crowding (when many funds converge on similar ML-derived signals, the trade gets less profitable and more correlated with everyone else's risk exactly when you don't want it to be).
The most publicized AI-in-finance story of the last two years is Klarna's customer service assistant, built with OpenAI and announced in February 2024. In its first month, it handled 2.3 million conversations — two-thirds of Klarna's customer service volume — and Klarna described that as equivalent to the output of 700 full-time agents, with an estimated $40 million profit improvement for 2024 and resolution times dropping from about 11 minutes to under 2. By late 2025 Klarna was citing an even larger figure, describing the assistant's output as equivalent to 853 employees.
The part that gets left out of the retelling: in May 2025, CEO Sebastian Siemiatkowski told Bloomberg the cost-cutting had "gone too far," citing inconsistent quality ("some customers get an amazing agent, some a less engaged agent, prompting repeated contacts") and said Klarna was rehiring human agents so customers would always have the option to reach a person. Klarna's headcount reduction over that period is attributed mostly to attrition and a hiring freeze rather than direct AI-driven layoffs, but the walkback itself is the interesting data point: the highest-profile AI win in finance was a support chatbot, not a core financial decision system, and even that had to be partially reversed for quality reasons rather than cost.
Robo-advisors (Betterment, Wealthfront) are the other consumer-facing category people lump into "AI in finance," but most of what they do is rules-based portfolio allocation and tax-loss harvesting against a target allocation — closer to a decision tree than a learned model, and largely unrelated to the LLM-driven wave everyone means when they say "AI" today.
| Use case | Core technique | Decision autonomy | Regulatory scrutiny | Maturity |
|---|---|---|---|---|
| Fraud detection | Supervised ML + graph features, real-time scoring | Fully automated, human review only on edge cases | Card network rules, limited explainability mandate | High — decade-plus of production use |
| Credit underwriting | Supervised ML on alternative + traditional data | Automated scoring, but decisions must be individually explainable | ECOA / Reg B adverse-action requirements, active CFPB oversight | Medium — real deployments, but under continuous legal constraint |
| Algorithmic trading | Statistical/ML signal generation; rules-based execution | Signal generation automated, position sizing and risk limits human-set | SEC/exchange circuit breakers, post-2010 market structure rules | High for execution, low for full autonomy |
| Customer support/advisory | LLM-based conversational agents | Automated for common cases, escalation to humans preserved | Consumer protection, limited financial-advice liability rules | Medium — fast growth, quality issues already forcing correction |
Line these four up and a thesis falls out that's more useful than "AI is transforming finance": AI in finance is most mature exactly where a wrong answer only costs money, and most constrained exactly where a wrong answer creates legal liability. Fraud scoring and trade execution can fail expensively but privately — the cost shows up on a balance sheet, and the firm eats it or tunes the model. Credit underwriting and, increasingly, customer service failures show up as discrimination complaints, regulatory inquiries, and reputational damage that no amount of model accuracy fixes after the fact. That's not a temporary state waiting for better models — it's a structural constraint tied to who bears the cost of an error, and it will keep shaping which parts of finance let AI operate autonomously and which parts keep a human explicitly in the loop.
For anyone building in this space, the practical takeaway is to separate "can the model predict this well" from "can this decision be explained to a regulator or a denied applicant" — they're different engineering problems, and the second one is usually the one that actually gates deployment.