AI in Financial Services: Turning Trusted Data into Scalable Decisions

Financial services was among the earliest industries to adopt machine learning (ML) — from fraud detection and algorithmic trading to credit risk and portfolio optimization. Generative AI and agentic AI have expanded those capabilities into new areas, including customer service, research, regulatory reporting and software development.

The expansion is already producing measurable returns. Snowflake-sponsored Omdia research found that 68% of financial services respondents had quantified positive returns from generative AI, while 92% said their organizations were training, tuning or augmenting large language models with proprietary data. Access to a capable foundation model is increasingly common. The advantage now comes from grounding it in the organization’s data, business context and expertise.

That shift places new demands on the data, controls and workflows that support AI at scale. Models need timely and reliable information, while the systems around them must preserve lineage, enforce access controls and support decisions that may later be examined by regulators, auditors, employees or customers. In financial services, the ability to scale AI is closely tied to the ability to govern the data and workflows on which it depends.

Watch John Young, VP of Data at Purpose Financial, discuss the company’s journey to transform its approach to fraud detection, automation and self-service, driving growth and improving operational efficiency:

AI supports financial decisions and workflows that involve large volumes of data, rapid analysis or the consistent application of complex criteria. ML remains central to prediction, classification and scoring. Generative AI helps people work with documents and other unstructured information, while AI agents can use models, data and tools to carry out defined sequences of work.

Fraud detection and anti-money laundering

Fraud detection remains one of the most established applications of ML in financial services. Models evaluate signals across transactions, accounts, devices, locations and prior behavior, looking for combinations that indicate elevated risk. Unlike static rules, which detect activity that meets predefined conditions, ML can identify less obvious patterns and adjust risk scores as new information arrives.

Rules still play an important role, however. They capture known schemes, regulatory thresholds and institutional policies, while models add probabilistic analysis and help prioritize the activity most likely to warrant intervention. Within anti-money laundering (AML) programs, AI can also uncover relationships among accounts and entities, rank alerts and assist investigators working through large volumes of transactions. A joint Bank of England and Financial Conduct Authority survey identified AML and fraud prevention among the areas producing the greatest current benefits from AI.

The operational constraint is speed. A fraud score may need to return before a payment is authorized, leaving little room for delayed data or an extended review process. The system must also capture investigation outcomes so teams can determine whether model performance is performing as expected or if it’s deteriorating or responding to patterns that criminals have abandoned.

Credit decisioning and risk assessment

AI can expand the analysis used in credit decisions, helping lenders assess repayment capacity, default risk and the terms appropriate for a particular applicant. Models can draw on conventional credit information alongside other legally permissible data, finding relationships that simpler scorecards might miss.

Beyond individual underwriting, ML can help financial institutions monitor portfolios and identify early signs of changing credit conditions. Other models support market, liquidity and operational risk analysis, although each involves different inputs, time horizons and definitions of an adverse outcome. Treating “risk assessment” as a single technical problem can conceal those differences.

Credit decisioning also illustrates how historical data can carry earlier decisions into a new system. An FCA literature review found that past decision-making, historical exclusion and sampling problems are leading potential sources of bias in supervised ML. Bias can also enter through variable selection, modeling choices and the way people interpret or act on predictions. For this reason, technical mitigation needs to be paired with contextual analysis and human review.

Financial institutions need the ability to see which information influenced a decision, whether performance differs across relevant groups and how the model behaves as economic conditions change. Those requirements affect model development, validation and monitoring, as well as the process used to communicate and review an adverse decision.

Algorithmic trading and investment analytics

Algorithmic trading depends heavily on latency and market synchronization. Trading systems use quantitative models to detect signals, manage execution or respond to market conditions. Their effectiveness depends on current prices, accurate positions and consistent timestamps, so a delay or data mismatch can change the result. Investment analytics generally works over longer time horizons, helping professionals evaluate securities, portfolios, market trends and economic scenarios.

Investment research requires traceable sources, sound financial semantics and a clear distinction between retrieved facts, model-generated analysis and the professional judgment applied to them. Generative AI can assist in investment analytics by making large collections of research easier to navigate. Analysts are now using it to summarize filings and earnings calls, compare reports or retrieve information through natural-language questions. More organizations are also integrating AI agents to coordinate several stages of research, using data retrieval, analytical models and other tools to prepare a result for review.

As these workflows span more sources and systems, the quality of the result depends increasingly on whether the agent can interpret the information in the right financial context. As Jonathan Regenstein and Tony Yiu write in their analysis of agentic AI in financial services, “Agents are only as good as the data and context they are given.” Financial information spans structured market and holdings data, semi-structured filings and unstructured research. A system may retrieve the correct records yet misread their significance when definitions, relationships or business context are inconsistent.

Customer service and personalized financial guidance

Generative AI allows customers and employees to interact with information conversationally. Customer-facing assistants can answer product questions, explain account activity and guide users through routine service processes. Internal assistants can retrieve approved policies, summarize previous interactions and help representatives resolve an inquiry without searching several disconnected systems.

Personalization may draw on customer needs, product holdings, eligibility criteria and prior interactions to recommend relevant information or a suitable next step. The system needs an accurate view of the relationship and current consent choices because a recommendation based on incomplete holdings, outdated eligibility rules or activity belonging to the wrong customer can create significant problems.

Agentic AI extends assistance into multistep service work, such as collecting information, verifying that required documents are present or preparing a case for employee review. With each added action, identity, permissions and approval boundaries become more important.

In their joint survey, the Bank of England and FCA found that 55% of reported financial services AI use cases involved some degree of automated decision-making, while only 2% used fully automated decisions. The difference suggests that many institutions are adopting automation while retaining human involvement for ambiguous or consequential cases.

Financial, liquidity and scenario forecasting

ML can help financial institutions forecast cash flows, liquidity needs, credit losses and other changes in financial performance or risk exposure. ML models can update projections as new data arrives, estimate ranges of possible outcomes and identify variables associated with a change in expectations.

The same techniques support financial planning and analysis (FP&A), although corporate forecasting is only one part of the opportunity. Treasury teams may need to model funding and liquidity under different conditions, while risk teams assess how portfolios could respond to changes in interest rates, defaults or market volatility. Generative AI can make those analyses easier to explore by summarizing variances, answering questions about assumptions and preparing scenario narratives.

Apparent precision can conceal inconsistent inputs, however. Different businesses may calculate revenue, exposure or customer value differently, and source systems may update on incompatible schedules. Historical relationships may also weaken when market conditions or customer behavior change.

For this reason, effective forecasting requires more than a model trained on a large data set. Institutions need agreed-upon definitions, explicit assumptions and a process for comparing forecasts with actual results. Scenario outputs should also communicate uncertainty rather than collapsing a range of possible outcomes into one authoritative-looking number.

Regulatory compliance and reporting

AI can help compliance teams search regulatory material, classify documents, compare policy versions and monitor activity that may require investigation. Generative models can summarize lengthy guidance or prepare an initial draft of a report, while ML supports activities such as AML alert prioritization and transaction monitoring.

Human oversight is especially important when a workflow combines retrieval, generation and action. Institutions need to know which source supported a statement, which model produced the analysis and who approved the result. Any corrections must be reflected in the final record rather than remaining inside an isolated conversation with an AI assistant.

Operational resilience regulations add another dimension. DORA, which has applied in the EU since January 2025, establishes requirements around ICT risk management, incident reporting, resilience testing and third-party technology risk for covered financial entities. AI may help organizations analyze risks and documentation, but its own infrastructure, dependencies and failure scenarios also form part of the environment those institutions must manage.

Access to advanced models is widening. The harder-to-reproduce asset is the institution’s own data, together with the definitions, relationships and controls that make it usable.

The relationship between data governance and AI competitive advantage is especially direct in financial services, where proprietary information often includes sensitive customer data, transactions, positions, internal policies and risk assessments. In the Snowflake-sponsored Omdia research, 57% of financial services respondents said fragmented systems and data silos were slowing AI initiatives, while 89% identified data engineering as important to AI project success.

Shared financial context

A unified platform can bring together data from core banking, payments, markets, customer systems and external providers. Physical consolidation alone doesn’t create a reliable AI foundation, however. Data also needs consistent identity, meaning and relationships.

For example, a customer might appear under several identifiers across deposit, lending and wealth-management systems. The terms “account,” “exposure” or “revenue” may refer to different calculations depending on the business unit. An AI system needs enough semantic context to distinguish those meanings and select the correct source for the question or workflow.

Shared definitions also reduce the risk that different models act on incompatible versions of the business. When an agent retrieves a policy, queries a transaction and invokes a risk model, each part of the workflow should interpret customers, products and financial measures consistently.

Freshness matched to the decision

“Real-time data” is often treated as a universal requirement, but appropriate freshness depends on the decision. For example, a transaction authorization will require current information and a subsecond response, while an intraday trading model needs synchronized market data and positions. Monthly management reporting can tolerate a longer delay, provided users know when the data was last updated and which accounting period it represents.

Designing every pipeline for the lowest possible latency adds cost and complexity without necessarily improving the outcome. The better approach is to define freshness requirements at the use-case level, monitor whether they’re being met and prevent a model or agent from presenting stale data as current.

Lineage and proportionate model risk controls

For consequential applications, financial institutions may need to reconstruct how a result was produced. That record can include source data, transformations, model and prompt versions, retrieved documents, policies, tool calls and human approvals.

The level of control should reflect the system’s purpose and materiality. A tool summarizing internal research, for example, doesn’t require the same validation and oversight as a credit model or trading system. Applying identical controls to every use case would create unnecessary friction, and weak controls around a consequential system create risk.

Model risk management frameworks provide a structure for making those distinctions. The UK PRA’s SS1/23 addresses model identification, governance, development, independent validation and risk mitigation for banks within its scope, including AI and ML where they’re used as modeling techniques. Canada’s OSFI has taken a broader approach in its revised Guideline E-23, which explicitly incorporates AI and ML into enterprise-wide model lifecycle governance and takes effect in May 2027.

Controlled collaboration with external data

Financial AI frequently relies on data that originates outside the institution. Market feeds, economic indicators, consortium data and information from business partners may improve fraud analysis, risk models or investment research.

Traditional data exchange creates additional copies that must be secured, updated and governed. Secure sharing and data clean rooms offer an alternative, allowing participants to collaborate while controlling which data and queries each party can access. Where an analysis requires aggregate patterns rather than individual records, privacy-enhancing techniques such as differential privacy may provide another layer of protection.

Financial institutions still need clear purposes, data rights, quality standards and controls over what can be derived or exported. The architecture should enforce those decisions rather than relying on every downstream application to implement them independently.

Snowflake’s AI Data Cloud for Financial Services gives institutions an environment for working with internal and external financial data, developing predictive models and building generative and agentic applications.

For predictive applications such as fraud detection, credit risk and financial forecasting, Snowflake ML supports the model lifecycle alongside governed enterprise data. Snowflake Feature Store helps teams define, manage and reuse the features supplied to models, while Snowflake Model Registry records model versions, metrics and metadata and controls who can manage or run them. Those capabilities help institutions maintain consistency between model development and production rather than recreating features and controls in separate systems.

Snowflake Cortex AI extends the same environment to generative and agentic AI. Cortex Search retrieves relevant information from unstructured content, while Cortex Analyst lets applications answer natural-language questions over structured data. Cortex Agents can coordinate retrieval, models and tools across multistep workflows, and Cortex AI Functions can parse, classify and extract information from documents such as filings, reports and forms. Snowflake CoWork gives business users a governed interface for asking questions of enterprise data, receiving cited answers and completing work across connected tools.

Snowflake Horizon Catalog applies context and governance across those data and AI assets. Semantic views establish business-aligned definitions, while data quality monitoring and column-level lineage help teams trace an AI-generated answer to its underlying sources. Sensitive-data classification, masking policies, row-level access controls and AI guardrails help enforce boundaries around what people, applications and agents can retrieve or disclose. Because those controls operate at the data and query layers, financial institutions don’t necessarily need to recreate a separate policy framework for each AI interface.

Organizations can also extend their AI systems with external information. Snowflake Marketplace and Secure Data Sharing provide access to third-party data that organizations can combine with their own governed information, while Snowflake Data Clean Rooms allow multiple parties to analyze combined data under approved controls without giving participants direct access to the underlying raw records. That approach can support research, risk and fraud workflows that depend on partner, consortium or commercial data while limiting the additional copies created through conventional exchanges.

Together, these capabilities connect traditional ML, generative AI and agentic workflows to a shared data and governance foundation. Teams can select the models and operating controls appropriate to each financial use case while keeping business definitions, access policies and lineage connected as the applications evolve.

See how AI is being used in other industries >

Explore the AI in Industries Hub:

• AI in Advertising
• AI in Cybersecurity
• AI in Healthcare and Life Sciences
• AI in Manufacturing
• AI in Retail