Governing Generative AI Across Financial Institutions: A Framework for Generative AI Risk Control

Alessandra Lin; Dennis Mao; Yiqing Wang; Yixin Kang

REVIEW 3 major objections 4 minor 22 references

Governing Generative AI Across Financial Institutions: A Framework for Generative AI Risk Control

T0 review · 3 major / 4 minor · reviewed 2026-08-02 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Generative AI uses in finance reduce to five capability patterns — knowledge synthesis, content generation, analytical assistance, interaction, and workflow orchestration — that map onto major financial functions, the paper argues.

desk verdict A competent survey of GenAI use cases in finance, but the five-capability taxonomy is not actually used consistently in the paper's own mapping table, and the title overpromises. read the letter →

arxiv 2607.04103 v3 pith:WY5ECZVI submitted 2026-07-05 q-fin.RM cs.LG

Dennis Mao , Alessandra Lin , Yixin Kang , Yiqing Wang This is my paper

classification q-fin.RMcs.LG

keywords generativeAIfinancialservicesbankingcapitalmarketsinvestmentresearchtechnologylargelanguagemodelsretrieval-augmentedgeneration

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper sets out to give financial institutions a reusable way to see where generative AI fits across their operations. It argues that the many possible uses reduce to five capability patterns — knowledge synthesis, content generation, analytical assistance, interaction, and workflow orchestration — and that these patterns recur in banking, capital markets, insurance, payments, and wealth management. It maps each pattern to representative functions such as investment research, lending, fraud investigation, operations, reporting, and insurance claims, and describes the technical architectures that support them. A sympathetic reader is left with a shared vocabulary for scoping generative AI initiatives and a design principle: connect generative systems to deterministic, specialized engines rather than letting them reason or calculate alone.

What carries the argument

The load-bearing object is the five-pattern capability taxonomy (knowledge synthesis, content generation, analytical assistance, interaction, and workflow orchestration) together with the architecture patterns in Table 1 — retrieval-augmented generation, tool-using copilots, multimodal document intelligence, and agentic workflows. The taxonomy gives practitioners a single lens for grouping use cases; the architecture patterns tell them how to build each group. Table 2 then consolidates the capability-to-function mapping, making the framework directly inspectable and reusable.

What would settle it

Collect a sample of generative AI applications actually deployed in financial institutions — from public disclosures, vendor case studies, or regulator reports — and classify each into exactly one of the five patterns. If a large share fits none of them, or fits several without a dominant pattern, the taxonomy's organizing power collapses.

Watch

Extended reading notes

Core claim

The paper's central claim is that generative AI's contribution to finance is broader than conversational interfaces: the same core capabilities — synthesizing large document collections, generating content and code, assisting analysis, interacting conversationally, and orchestrating multi-step workflows — underpin applications across the financial value chain. The paper deliberately presents an application-oriented landscape rather than a control or regulatory framework. It identifies hybrid architecture as the practical key: retrieval-augmented generation supplies current context, tool-using systems call deterministic calculators and databases, multimodal models parse charts and forms, and

Load-bearing premise

The framework stands or falls on the claim that the five capability patterns are a coherent and complete way to partition generative AI uses in finance, and that the cited examples represent actual practice.

Editorial extensions

If this is right

A financial institution can inventory its generative AI initiatives by asking which of the five capability patterns each one belongs to, then choose the supporting architecture accordingly: RAG for synthesis, tool use for analysis, agents for workflow orchestration.
Applications that combine unstructured information with deterministic calculations — research, lending review, fraud investigation, reporting — are where the paper expects the largest near-term value.
The paper's design principle implies that models should not be asked to perform arithmetic or retrieval in isolation; systems should call specialized engines for numerical and data work.
The same capability vocabulary applies across nearly every financial function, so experience gained in one domain can transfer to another.
Future research priorities highlighted by the paper include finance-specific benchmarks, grounded generation, numerical and temporal reasoning, and evaluation of agentic workflows.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Because the paper stops at the application landscape, an obvious next step is to turn the five patterns into an oversight tool: supervisors and internal auditors could use the same map to ask where generative AI risk concentrations sit. That is an extension, not something the paper claims.
The taxonomy's categories are not mutually exclusive in practice; one application can be a knowledge-synthesis, analytical-assistance, and workflow-orchestration product at once, so the map may function more as a diagnostic lens than a strict filing system.
A direct test of the framework: build a corpus of publicly described financial generative AI deployments and classify each into the five patterns; the rate of 'no fit' or 'multiple fits' would show how much revision the taxonomy needs.
The paper's hybrid-design principle implies that model risk in finance will increasingly live in the seams between language models and deterministic engines — a place where existing risk frameworks have no settled playbook.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

This is a readable, well-organized survey of generative AI applications in finance, and worth knowing about if you work in the area. But the stress-test is right: the central claim—a reusable vocabulary of five capability patterns—doesn't hold up in the paper's own Table 2.

What it does well: it compiles a broad set of applications across banking, wealth, lending, fraud, operations, and insurance, and links them to technical patterns like RAG, tool-using copilots, multimodal document intelligence, and agentic workflows. The best parts are the discussion of hybrid designs (language model as interface, deterministic engines for calculation) and the candid limitations section (§3.11) on hallucination, numerical reasoning, and evaluation gaps. As an orientation for practitioners, it serves a purpose.

The soft spots are real. Section 2 defines five capability patterns, but Table 2's 'Primary capability' column uses entries like 'Personalization and generation,' 'Document intelligence and generation,' 'Multimodal extraction and tools,' 'Coding assistance,' and 'Multimodal synthesis.' These come from the architecture table (Table 1) rather than the capability list. That is not cosmetic: the mapping cannot be expressed in the claimed vocabulary, so the paper's contribution as a 'reusable vocabulary and mapping' is not actually delivered. There is also no selection criteria, completeness argument, or inter-rater reliability check for the taxonomy, so a reader cannot tell whether the five patterns are coherent or exhaustive. Minor but notable: the title promises a governance/risk-control framework while Section 1 explicitly says the paper is not proposing one, and there is a placeholder 'Figure??' in §3—a quality-control issue.

Who is this for? Practitioners wanting a quick landscape, or researchers looking for an example of a taxonomy that needs validation before adoption. It is a survey, not a research result with new empirical facts.

I would send it to peer review—editors should let referees push for internal consistency and a title that matches the content—but I would not cite it as a foundation for my own work until the taxonomy is fixed and validated.

Referee Report

3 major / 4 minor

Summary. The paper proposes an application-oriented taxonomy of generative AI in finance, centered on five capability patterns (knowledge synthesis, content/communication generation, analytical/coding assistance, interactive assistance, and workflow orchestration) and maps these to major financial functions including investment research, wealth management, customer service, lending, risk/fraud, operations, reporting, software/data, and insurance. It also surveys common technical architectures (retrieval-augmented generation, tool-using copilots, multimodal document intelligence, agentic workflows) and identifies sources of business value and open technical challenges. The paper is a conceptual synthesis with no new empirical data; its contribution is a proposed vocabulary and mapping that practitioners could use to scope GenAI initiatives.

Significance. If the taxonomy and mapping were internally consistent and adequately grounded, the paper would provide a useful shared vocabulary for describing and comparing GenAI applications in finance, particularly the emphasis on separating generative language layers from deterministic calculation (§2.3) and the candid treatment of limitations (§3.11). The paper draws on a broad literature and acknowledges open problems such as hallucination, numerical accuracy, and evaluation gaps. However, the framework's value is undermined by the inconsistent use of the claimed five-pattern vocabulary in Table 2 and the mismatch between the title's 'risk control' promise and the stated non-framework scope. These issues are central and fixable, but they require substantive revision rather than minor copyediting.

major comments (3)

[§3.9, Table 2] The 'Primary capability' column does not consistently use the five capability patterns defined in §2. It contains entries such as 'Personalization and generation', 'Document intelligence and generation', 'Multimodal extraction and tools', 'Coding assistance', and 'Multimodal synthesis' — none of which are among the five patterns (knowledge synthesis, content/communication generation, analytical/coding assistance, interactive assistance, workflow orchestration). Some entries appear to be technical patterns from Table 1 rather than capability patterns. This is not cosmetic: the abstract and §1 claim that the paper organizes uses around five capability patterns and maps them to financial functions, but the mapping in Table 2 cannot be expressed in that vocabulary. The central 'reusable vocabulary and mapping' claim is therefore not delivered as stated. The authors should either revise Table
[Title and §1] The title promises 'A Framework for Generative AI Risk Control,' but §1 states: 'Rather than proposing a control or regulatory framework, it focuses on what generative systems may do, where they may be deployed, and how their technical capabilities translate into business use.' This is a direct contradiction between the advertised contribution and the actual content. A reader picking up the paper on the basis of the title would expect at least a discussion of risk-control design, governance, or regulatory alignment, none of which is present. The title should be revised to reflect the paper's real scope, or the paper should add a section that actually addresses risk control. Without this, the manuscript's framing is misleading.
[§2] The five capability patterns are introduced 'through some usage case examples' with no selection criteria, no formal definitions, and no argument for completeness or disjointness. The categories appear to overlap in practice: for example, knowledge synthesis (§2.1) and interactive assistance (§2.4) both subsume question-answering over documents, and workflow orchestration (§2.5) may comprise the other patterns. Because the paper's central claim is that these patterns form a partition of GenAI applications in finance, the taxonomy needs at least an explicit statement that it is a heuristic grouping, with criteria for assigning a use case to a pattern. Without this, the mapping in Table 2 is not principled and a reader cannot judge whether major use classes are missed or double-counted.

minor comments (4)

[§3, line with 'Figure??'] There is an unresolved placeholder 'Figure??' in the sentence introducing the capability-to-application map. The figure is missing, which disrupts the reader's ability to follow the mapping described in the text.
[§2.1] 'extract key words' is likely meant to be 'extract key facts' or 'extract keywords'; as written it is ambiguous and could be mistaken for a lexical extraction task rather than semantic summarization.
[Abstract vs §2] The abstract lists the five patterns as 'knowledge synthesis, content generation, analytical assistance, interaction, and workflow orchestration,' but the section headings are 'Content and Communication Generation,' 'Analytical and Coding Assistance,' and 'Interactive Assistance.' Harmonize the names to avoid apparent inconsistency.
[References] Several references have inconsistent formatting (e.g., [17] uses 'et al.' after some author lists, while [11] has a partly garbled publisher name). A final proofreading pass is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; survey-style taxonomy has internal consistency issues but no logical circularity.

full rationale

The paper makes no predictive or first-principles derivation that could reduce to its own inputs. It is a survey and taxonomy: Section 2 proposes five capability patterns, and Section 3 maps them to financial functions using external literature and illustrative use cases. There are no fitted parameters, no equations, no benchmark claims, and no load-bearing chain of self-citations; the references are to independent external work. The closest issues are non-circular consistency problems: Table 2's 'Primary capability' column contains entries such as 'Personalization and generation,' 'Document intelligence and generation,' 'Multimodal extraction and tools,' 'Coding assistance,' and 'Multimodal synthesis' that are not among the five patterns defined in Section 2, and Section 1 states the paper is 'Rather than proposing a control or regulatory framework' despite a title promising a risk-control framework. These are internal-consistency and framing concerns, not cases in which a result is assumed by construction or a fitted input is renamed as a prediction. No circular step can be quoted, so the circularity score is 0.

Assumptions & free parameters 0 free parameters · 3 assumptions · 0 invented entities

The paper's central contribution is a taxonomy; it introduces one unvalidated framework (the five capability patterns) and depends on the fidelity of the cited literature. There are no fitted parameters or invented physical entities.

assumptions (3)

ad hoc to paper The five capability patterns (knowledge synthesis, content generation, analytical assistance, interaction, workflow orchestration) constitute a valid and useful partition of GenAI applications in finance.
Introduced by the authors without empirical or literature-derived justification; Section 2 defines the patterns but gives no selection criteria.
domain assumption The cited sources accurately describe feasible or deployed financial use cases.
The paper does not independently verify any application; it relies on references including preprints and SSRN papers.
domain assumption Generative models can perform the described tasks well enough to generate business value despite known limitations.
Section 3.10 asserts value channels, while Section 3.11 concedes hallucination, unreliable numerical reasoning, and incomplete retrieval; the optimistic framing assumes these limitations are surmountable.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Governing Generative AI Across Financial Institutions: A Framework for Generative AI Risk Control." pith.science (2026). https://pith.science/paper/WY5ECZVI

@misc{pith2026260704103,
  author       = {Pith},
  title        = {Pith review of: Governing Generative AI Across Financial Institutions: A Framework for Generative AI Risk Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WY5ECZVI}},
  note         = {Machine review of arXiv:2607.04103}
}

read the original abstract

Generative artificial intelligence is moving from general-purpose experimentation toward specialized applications across banking, capital markets, insurance, payments, and wealth management. Its main contribution is not limited to conversational interfaces. Modern generative systems can synthesize large document collections, extract information from unstructured data, generate software and analytical code, create scenario narratives, support research workflows, and coordinate multi-step tasks. These capabilities make generative AI especially relevant to finance, where decisions often depend on combining quantitative data with contracts, policies,filings, news, customer communications, and expert judgment. This paper presents an application-oriented view of generative AI in finance. It organizes potential uses around five capability patterns, including knowledge synthesis, content generation, analytical assistance, interaction, and workflow orchestration, and maps them to major financia functions. Representative applications include investment research, customer service, lending support, fraud investigation, financial reporting, operations automation, software development, and personalized financial guidance. The paper also discusses common technical architectures, such as retrieval-augmented generation, tool-using assistants, multimodal models, and agentic workflows, and identifies practical factors that shape business value. The resulting landscape provides a foundation for researchers and practitioners seeking to understand where generative AI may produce the greatest operational and analytical impact in financial services

Figures

Figures reproduced from arXiv: 2607.04103 by the authors.

Review history (2 revisions) →

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

22 extracted references · 2 canonical work pages

[1]

Rishi Bommasani et al.On the Opportunities and Risks of Foundation Models. 2021. arXiv: 2108.07258 [cs.LG].url:https://arxiv.org/abs/2108.07258

arXiv 2021
[2]

A survey of large language models for financial applications: Progress, prospects and challenges

Yuqi Nie et al. “A survey of large language models for financial applications: Progress, prospects and challenges”. In:arXiv preprint arXiv:2406.11903(2024)

arXiv 2024
[3]

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

Patrick Lewis et al. “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”. In:Advances in Neural Information Processing Systems. Vol. 33. 2020, pp. 9459–9474.url: https://papers.nips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abst ract.html

2020
[4]

Agentic ai systems applied to tasks in financial services: Modeling and model risk management crews

Izunna Okpala, Ashkan Golgoon, and Arjun Ravi Kannan. “Agentic ai systems applied to tasks in financial services: Modeling and model risk management crews”. In:arXiv preprint arXiv:2502.05439(2025)

arXiv 2025
[5]

Unleashing the power of text for credit default prediction: Compar- ing human-generated and AI-generated texts

Zongxiao Wu et al. “Unleashing the power of text for credit default prediction: Compar- ing human-generated and AI-generated texts”. In:SSRN Electronic Journal. https://doi. org/10.2139/ssrn4601317 (2023)

work page doi:10.2139/ssrn4601317 2023
[6]

YihangChenetal.Does RAG Know When Retrieval Is Wrong? Diagnosing Context Compliance under Knowledge Conflict. 2026. arXiv:2605.14473 [cs.CL].url: https://arxiv.org/abs /2605.14473

arXiv 2026
[7]

Eleanor W Dillon et al.Shifting work patterns with generative ai. Tech. rep. National Bureau of Economic Research, 2025

2025
[8]

Generative Artificial Intelligence in Finance: A Systematic Literature Review and a Research Agenda

Yunfei Chen et al. “Generative Artificial Intelligence in Finance: A Systematic Literature Review and a Research Agenda”. In:Digital Technologies Research and Applications4.2 (2025), pp. 14–32

2025

Show all 22 references

[9]

Generative AI for finance: applications, case studies and challenges

Siva Sai et al. “Generative AI for finance: applications, case studies and challenges”. In:Expert Systems42.3 (2025), e70018

2025
[10]

Su Wang et al.When Safe Skills Collide: Measuring Compositional Risk in Agent Skill Ecosystems. 2026. arXiv:2606.00448 [cs.SE].url:https://arxiv.org/abs/2606.00448

2026 arXiv
[11]

Report LLMs in Finance

Giulio Bagattini et al.Leveraging Large Language Models in Finance: Pathways to Responsible Adoption. Report LLMs in Finance. Published 4 June 2025. European Securities et al., June 2025.url: https://www.esma.europa.eu/sites/default/files/2025-06/LLMs_in_fina nce_-_ILB_ESMA_Tu...

2025
[12]

AI in wealth management-transforming personal finance for the better

Constantinos Challoumis. “AI in wealth management-transforming personal finance for the better”. In:XVI International Scientific Conference. 2024, pp. 30–61

2024
[13]

Assessing the impact of generative artificial intelligence on customer engagement in business-to-customer scenarios

Jooyoung Kim and Gyunghyun Choi. “Assessing the impact of generative artificial intelligence on customer engagement in business-to-customer scenarios”. In:Asia-pacific Journal of Convergent Research Interchange10.2 (2024), pp. 89–104. 9

2024
[14]

Intellichain stars at the regulations challenge task: A large language model for financial regulation

Shijia Jiang et al. “Intellichain stars at the regulations challenge task: A large language model for financial regulation”. In:Proceedings of the Joint Workshop of the 9th Financial Technology and Natural Language Processing (FinNLP), the 6th Financial Narrative Processing (F...

2025
[15]

Interpretable LLMs for credit risk: A systematic review and taxonomy

Muhammed Golec and Maha AlabdulJalil. “Interpretable LLMs for credit risk: A systematic review and taxonomy”. In:Expert Systems with Applications(2025), p. 130941

2025
[16]

Credit risk meets large language models: Building a risk indicator from loan descriptions in p2p lending

Mario Sanz-Guerrero and Javier Arroyo. “Credit risk meets large language models: Building a risk indicator from loan descriptions in p2p lending”. In:arXiv preprint arXiv:2401.16458 (2024)

2024 arXiv
[17]

Generative AI and financial crimes: a quantitative systematic literature review

Milind Tiwari et al. “Generative AI and financial crimes: a quantitative systematic literature review”. In:Crime Science15.1 (2026), p. 5

2026
[18]

AI-driven intelligent document processing for banking and finance

Ramesh Pingili. “AI-driven intelligent document processing for banking and finance”. In: International Journal of Management & Entrepreneurship Research7.2 (2025), pp. 98–109

2025
[19]

A comprehensive review of generative AI in finance

David Kuo Chuen Lee et al. “A comprehensive review of generative AI in finance”. In:FinTech 3.3 (2024), pp. 460–478

2024
[20]

Generative AI for Claims Exceptions and Investigations: Enhancing Resolution Efficiency in Complex Insurance Processes

Keerthi Amistapuram. “Generative AI for Claims Exceptions and Investigations: Enhancing Resolution Efficiency in Complex Insurance Processes”. In:Available at SSRN 5785482(2025)

2025
[21]

Survey of Hallucination in Natural Language Generation

Ziwei Ji et al. “Survey of Hallucination in Natural Language Generation”. In:ACM Computing Surveys55.12 (2023), pp. 1–38.doi: 10.1145/3571730.url: https://doi.org/10.1145/35 71730

2023 doi
[22]

On faithfulness and factuality in abstractive summarization

Joshua Maynez et al. “On faithfulness and factuality in abstractive summarization”. In: Proceedings of the 58th annual meeting of the association for computational linguistics. 2020, pp. 1906–1919. 10

2020

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.

[1] [1]

Rishi Bommasani et al.On the Opportunities and Risks of Foundation Models. 2021. arXiv: 2108.07258 [cs.LG].url:https://arxiv.org/abs/2108.07258

arXiv 2021

[2] [2]

A survey of large language models for financial applications: Progress, prospects and challenges

Yuqi Nie et al. “A survey of large language models for financial applications: Progress, prospects and challenges”. In:arXiv preprint arXiv:2406.11903(2024)

arXiv 2024

[3] [3]

Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks

Patrick Lewis et al. “Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks”. In:Advances in Neural Information Processing Systems. Vol. 33. 2020, pp. 9459–9474.url: https://papers.nips.cc/paper/2020/hash/6b493230205f780e1bc26945df7481e5-Abst ract.html

2020

[4] [4]

Agentic ai systems applied to tasks in financial services: Modeling and model risk management crews

Izunna Okpala, Ashkan Golgoon, and Arjun Ravi Kannan. “Agentic ai systems applied to tasks in financial services: Modeling and model risk management crews”. In:arXiv preprint arXiv:2502.05439(2025)

arXiv 2025

[5] [5]

Unleashing the power of text for credit default prediction: Compar- ing human-generated and AI-generated texts

Zongxiao Wu et al. “Unleashing the power of text for credit default prediction: Compar- ing human-generated and AI-generated texts”. In:SSRN Electronic Journal. https://doi. org/10.2139/ssrn4601317 (2023)

work page doi:10.2139/ssrn4601317 2023

[6] [6]

YihangChenetal.Does RAG Know When Retrieval Is Wrong? Diagnosing Context Compliance under Knowledge Conflict. 2026. arXiv:2605.14473 [cs.CL].url: https://arxiv.org/abs /2605.14473

arXiv 2026

[7] [7]

Eleanor W Dillon et al.Shifting work patterns with generative ai. Tech. rep. National Bureau of Economic Research, 2025

2025

[8] [8]

Generative Artificial Intelligence in Finance: A Systematic Literature Review and a Research Agenda

Yunfei Chen et al. “Generative Artificial Intelligence in Finance: A Systematic Literature Review and a Research Agenda”. In:Digital Technologies Research and Applications4.2 (2025), pp. 14–32

2025