Technical Summary: Governing Generative AI Across Financial Institutions

Problem Statement

Financial institutions manage vast quantities of both structured data (prices, transaction histories, risk measures) and unstructured information (contracts, regulatory filings, research notes, customer communications). Traditional analytical systems excel at well-specified tasks with structured data but lack the flexibility to search, interpret, summarize, or synthesize information expressed in natural language. While generative artificial intelligence (GenAI) offers the potential to bridge this gap by creating, transforming, and reasoning over text, code, and images, there is a need to map these technical capabilities to specific financial functions and understand how they translate into business value without proposing a new regulatory framework.

Methodology

This paper adopts an application-oriented perspective rather than a control or regulatory framework. The methodology involves:

1. Categorization of Capabilities: Identifying five core capability patterns of generative systems: Knowledge Synthesis, Content Generation, Analytical Assistance, Interactive Assistance, and Workflow Orchestration.
2. Architectural Analysis: Examining common technical patterns used to build finance-oriented GenAI, including Retrieval-Augmented Generation (RAG), tool-using assistants, multimodal models, and agentic workflows.
3. Mapping to Financial Functions: Systematically mapping these capabilities and architectures to major financial domains, including investment research, wealth management, lending, risk analysis, operations, and insurance.
4. Value and Challenge Assessment: Synthesizing the sources of business value and identifying persistent technical limitations and research gaps.

Key Contributions

1. A Capability-to-Application Landscape

The paper organizes potential GenAI uses in finance around five distinct patterns:

• Knowledge Synthesis: Utilizing RAG to search, compare, and summarize fragmented evidence across filings, reports, and emails. Outputs include earnings-call summaries, covenant extraction, and due-diligence briefs.
• Content and Communication Generation: Transforming technical inputs into coherent communications (e.g., client emails, research notes) and standardizing documents (e.g., converting analyst notes to templates).
• Analytical and Coding Assistance: Acting as an interface to databases and statistical libraries. Models generate SQL/Python code, interpret statistical outputs, and explain quantitative results, while separating language generation from deterministic calculation.
• Interactive Assistance: Enabling iterative exploration of financial information via conversational interfaces in contact centers and analytical workbenches, supporting multilingual interaction.
• Workflow Orchestration: Extending generation into multi-step execution where "agents" perform sequences of actions (retrieval, calculation, drafting) across systems, distinct from assistants that primarily draft or recommend.

2. Technical Architecture Patterns

The paper identifies five technical patterns essential for finance-oriented applications (Table 1):

• Prompt-based Assistants: General or domain-adapted models for drafting and brainstorming without direct private system access.
• Retrieval-Augmented Generation (RAG): Models accessing selected internal/external knowledge bases for research and contract review.
• Tool-Using Copilots: Models calling external calculators, pricing engines, and databases for portfolio analysis and financial modeling.
• Multimodal Document Intelligence: Systems processing text alongside tables, charts, and scanned images for financial statement spreading and KYC reviews.
• Agentic Workflows: Multi-step planning and execution across tools for due diligence and exception handling.

3. Domain-Specific Applications

The paper maps these capabilities to specific financial functions:

• Investment Research: Summarizing earnings calls, extracting management guidance, and synthesizing macroeconomic data.
• Wealth Management: Personalizing client briefs, explaining portfolio performance, and simulating scenarios (e.g., retirement timing) while relying on deterministic engines for numerical calculations.
• Lending: Extracting data from tax forms and bank statements, summarizing borrower narratives, and drafting credit memoranda.
• Risk & Fraud: Organizing heterogeneous evidence (transactions, alerts) into investigation timelines and case narratives.
• Operations & Reporting: Automating exception handling, trade confirmations, and generating narrative commentary for financial statements.
• Software Development: Generating data pipelines, unit tests, and translating legacy code.
• Insurance: Summarizing policies, extracting claim information from multimodal inputs (photos, forms), and drafting correspondence.

4. Sources of Business Value

The paper identifies five primary channels for value creation:

1. Reducing search and reading time via task-specific summaries.
2. Accelerating document production through high-quality first drafts.
3. Broadening access to analytical tools via natural language interfaces.
4. Improving workflow continuity by carrying context from research to documentation.
5. Enabling new products based on personalized, conversational, and multimodal interaction.

Results and Limitations

The paper does not present empirical experimental results but rather a synthesized landscape of current and near-future applications. It concludes that the most promising applications are hybrid systems that combine generative models with domain-specific data and specialized tools. In this architecture, RAG provides access to current documents, analytical engines perform reliable calculations, and multimodal models interpret complex forms, allowing GenAI to complement rather than replace established financial analytics.

The paper explicitly notes several technical limitations that remain central to development:

• Hallucinations: Models may produce unsupported statements or fabricated citations.
• Contextual Degradation: Performance may decline with very long documents, incomplete retrieval, or highly specialized terminology.
• Numerical Reasoning: Models are unreliable at calculation without external tools.
• Evaluation Gaps: There is a need for finance-specific benchmarks, grounded generation techniques, and better evaluation of agentic workflows.

Significance

The paper positions itself as a foundational resource for researchers and practitioners seeking to understand where generative AI may produce the greatest operational and analytical impact in financial services. It argues that GenAI introduces a new interaction layer across financial data, documents, software, and human expertise. Its significance lies in shifting the focus from conversational interfaces to a broader view where GenAI supports investment research, lending, fraud investigation, and software development by transforming unstructured information into searchable knowledge and multi-step digital workflows. The paper advocates for a future where hybrid systems—connecting generative interfaces to deterministic analytics—become the standard for financial institutions.