AI in Finance Today — FSB Report and the 75% Adoption Rate── AI is established in credit, fraud detection, asset management and customer service, but generative AI remains a thin layer

0175 percent of financial institutions use AI; only 2 percent let it decide autonomously

The Bank of England and Financial Conduct Authority published their third AI survey in November 2024. Of the 118 responding firms, 75 percent reported active AI deployment — up from 58 percent in 2022. The increase is steep, but fully autonomous decision-making accounts for just 2 percent of all use cases. Another 55 percent involve partial automation with a human making the final call, and 24 percent are semi-autonomous with human oversight at critical junctures.

That same month the FSB released its G20 report, "The Financial Stability Implications of Artificial Intelligence," identifying four vulnerability channels: (1) third-party dependence and service-provider concentration, (2) amplified market correlations, (3) cyber risk, and (4) model risk and data governance. AI improves operational efficiency, but it also creates structural risk when many institutions depend on the same models and the same handful of providers.

02In credit scoring, AI has lifted approval rates by 20–30 percent while holding default rates steady

Credit scoring is one of the areas where AI has most firmly taken root. Machine-learning models improve ROC-AUC by roughly 2 percentage points over logistic regression. The number sounds small, but applied across millions of applications it changes who gets approved. Zest AI reports that its clients, using models with more than ten times the conventional number of credit variables, achieved 20–30 percent higher approval rates with no increase in risk.

The effect on financial inclusion has been documented. A study published in MIS Quarterly in 2024 examined a bank with more than 50 million customers that switched from a rule-based model to an AI model. In the previously underserved population, approval rates rose while default rates fell. Separately, CreditVidya extended credit to 25 million individuals with thin credit files and reduced delinquency by 33 percent at the same risk level.

Japan's Financial Services Agency (FSA) flagged the other side of this equation. Its AI Discussion Paper (version 1.0, March 2025) listed explainability and fairness in AI-driven credit decisions as key issues. If a model cannot explain why it declined an applicant, neither complaint handling nor regulatory compliance can function. Accuracy gains and accountability must advance together.

03Fraud detection is AI's oldest financial use case — but fraud itself is evolving with AI

A BioCatch survey of 600 financial-services professionals conducted in January–February 2024 found that 73 percent use AI for fraud detection and 74 percent for financial-crime detection. Regional variation is notable: EMEA leads at 86 percent, North America sits at 71 percent, and Asia-Pacific trails at 67 percent. An additional 22–23 percent plan to adopt within six months, making fraud detection the most widely deployed AI application in finance.

The technology mix is telling: 83 percent of respondents use advanced machine learning, 72 percent use natural language processing, and 67 percent use deep learning. Fraud detection is fundamentally a classification problem — distinguishing fraudulent from legitimate transactions — and supervised learning excels at it. These models are no longer experimental; they are infrastructure.

But adversaries use AI too. Reported identity-fraud and related losses in US financial services reached $12.5 billion in 2024, up 25 percent year on year. Synthetic identities — fabricated personas assembled from real and fictitious data — are a primary driver. Banks now flag 1 in every 20 verification attempts as potentially fraudulent, yet the estimated share of illicit money flows actually caught stands at around 2 percent. AI-powered defence and AI-powered attack are advancing in parallel.

04Generative AI and traditional ML are not competing — they form layers

Discussions of AI in finance often conflate generative AI with traditional machine learning, and the conflation muddles the analysis. Traditional ML is discriminative: it classifies inputs. Fraud or not fraud, default or no default, buy or sell. Generative AI is generative: it produces text, code and scenarios.

In the Bank of England's 2024 survey, foundation models — the core technology behind generative AI — accounted for 17 percent of all AI use cases. The remaining 83 percent comprised regression analysis, decision trees and conventional neural networks. The core of fraud detection and credit scoring still runs on discriminative models. Generative AI sits on top, serving as the layer that explains results to humans and reads unstructured data.

The most effective financial AI deployments in 2026 combine both paradigms: discriminative models for scoring and anomaly detection, generative models for language processing, summarisation and client communication. Neither layer is sufficient on its own.

05Asset management and customer service — the other two domains

In asset management, AI centres on portfolio optimisation and risk management. A BIS working paper published in June 2024 mapped AI's impact across four financial functions: intermediation, insurance, asset management and payments. In asset management, pattern recognition and prediction in market data are the primary ML applications, while generative AI is beginning to automate market commentary and translate scenario analyses into natural language.

In customer service, chatbots are the most visible application, but the deeper shift lies elsewhere. The ECB has reported that European financial institutions use AI most frequently for client profiling and customer support. Generative AI allows institutions to produce individualised responses informed by account data and transaction history, rather than routing queries to template answers. Yet fully autonomous customer-facing AI remains rare. The Bank of England figure — 2 percent fully autonomous — signals that firms are deliberately cautious at the customer interface.

06The same two-layer structure is emerging in pharmaceutical and healthcare applications

The pattern visible across finance's four domains — discriminative ML for judgment, generative AI for communication — also appears in pharmaceutical and healthcare settings. The FDA received more than 500 submissions containing AI components between 2016 and 2023, and in January 2025 it issued draft guidance covering AI in nonclinical, clinical, post-marketing and manufacturing phases. The volume of AI-related clinical-trial publications grew 444 percent from 2019 to 2024, a compound annual growth rate of 40 percent. By end of 2024 more than 75 AI-derived molecules had reached clinical stages.

The structural parallel is clear. Clinical-trial data analysis — patient selection, adverse-event signal detection — is a discriminative-model domain. Drafting regulatory submissions, summarising trial reports and organising safety data are tasks where generative AI is entering the workflow. And both industries face the same demand for explainability: in finance, regulators want to know why a loan was denied; in pharmaceuticals, they want to know how a trial outcome was assessed. The FSB's warning about model risk is not unique to finance. It applies to any regulated industry where AI informs consequential decisions.

07Regulators are tracking adoption, concentration and governance — FSA, FSB, BIS

Japan's FSA published version 1.0 of its AI Discussion Paper in March 2025, drawing on a survey of 130 financial institutions conducted in late 2024. It identified three priority areas: AI in credit decisions, AI in financial-product sales and solicitation, and responses to emerging AI developments. In March 2026 the FSA updated the paper to version 1.1, incorporating findings from its AI Public-Private Forum held between June and December 2025. The FSA's stance emphasises the "risk of not challenging" — the cost of falling behind on AI adoption — while calling for voluntary governance frameworks.

The FSB's October 2025 follow-up report systematised monitoring indicators for AI adoption and examined in detail the risks arising from financial institutions' reliance on a small number of third-party AI providers. The BIS, in its June 2024 Annual Economic Report, noted that AI affects productivity, consumption, investment and labour markets, with direct implications for both price stability and financial stability. It also documented that central banks themselves are beginning to use AI for policy purposes.

08Three actions for institutions that have already deployed AI

First, manage discriminative and generative models separately. The discriminative models that power credit scoring and fraud detection require established model-risk management — back-testing, sensitivity analysis, explainability assurance. Generative models require a different discipline: fact-checking outputs and designing human-review workflows. Lumping both under a single "AI governance" label produces oversight that is too coarse to be effective.

Second, quantify third-party dependence. Inventory which external providers supply the models, data and infrastructure your AI systems rely on, and identify single points of failure. As the FSB's 2025 report underlined, concentration in a handful of AI providers is not a firm-level risk alone; it is a potential source of systemic risk.

Third, treat explainability not as a compliance cost but as a trust-building investment. The ability to explain a credit denial to a customer is a regulatory requirement, but it is also a means of earning trust. The same logic applies in pharmaceuticals, where explaining a trial-outcome assessment builds credibility with regulators and patients alike. Investment in explainability pays returns beyond compliance.

09The question that follows a 75 percent adoption rate

AI adoption in financial services is no longer a frontier initiative. A 75 percent adoption rate means that not using AI is becoming the exception. Yet the fact that only 2 percent of use cases are fully autonomous shows that institutions recognise the weight of delegating judgment to a model.

The issues that follow — algorithmic trading and market stability, regulatory frameworks, model-risk management, systemic risk from model concentration — are the subjects of the remaining articles in this series. Article 2 examines AI in capital markets, article 5 covers regulation, article 6 addresses model-risk governance, and article 9 considers systemic risk. The challenge in financial AI is not the capability of the technology. It is how to contain it.