The core argument of the article is that the competition in financial AI is not about who can build a more conversational "financial version of ChatGPT," but rather about who can deeply integrate into financial professionals' daily tools—such as Excel, PowerPoint, and Word—and core business processes—such as due diligence and approval—and directly deliver formal, reviewable, and archivable outputs.

Author: Resonant Ones

Source: Suichu.AI

The competition in financial AI isn't about "who can chat," but about "who can integrate into Excel, PowerPoint, and approval workflows."

Many people believe that the competition in financial AI is about training a larger model that better understands finance.

But Claude for Financial Services revealed the real answer: the core of financial AI is not the model, but the workflow.

It’s not about having AI chat with users about stocks; it’s about having AI integrate into Excel, PowerPoint, Word, investment research, investment banking, due diligence, compliance, reconciliation, and approval workflows.

This is crucial for domestic entrepreneurs. If you're still building a "financial version of ChatGPT," you're likely to be absorbed by big tech companies, data terminals, and office suites; but if you can take over the daily repetitive Excel, PowerPoint, Word, and approval workflows of financial institutions, your real opportunity is just beginning.

A real-world scenario

Last month, I spoke with a friend who works in private equity. Their team conducted due diligence on a consumer company and received a data room with 17 folders and over 400 documents—contracts, audit reports, bank statements, order details, interview transcripts, and management materials.

Previously, it took a VP with two analysts two weeks to produce a decent initial draft of an IC memo.

What about now? If a person (or an agent) could complete data organization, risk tagging, identification of missing items, and draft generation within 24 hours—would you think customers would pay for that?

This is not science fiction. Claude for Financial Services is already doing this—and it’s open-sourcing not an app, but a product paradigm of “Agent + Skill + Connector + Deliverable + Human Approval.”

Let’s start with the first insight. The product structure of Claude for Financial Services is straightforward: Agents handle end-to-end tasks, Skills encapsulate financial industry workflows, Connectors integrate with financial data and enterprise systems, and Excel, PowerPoint, and Word deliver the final outputs—backed by permissions, citations, audit trails, and human review to ensure suitability for financial institutions.

In the past, financial AI took the form of you asking a question and the AI providing an answer. But what financial institutions truly need is: given a set of documents, deliver a tangible output that can be reviewed, cited, archived, and integrated into business systems. The difference between these two approaches is significant. The value of financial AI lies in the deliverable, not in the chat interface.

Another notable change is that domestic financial institutions are no longer standing by.

From 2025 to 2026, I see the adoption scenarios falling into three tiers. Banks are leading the way: China Construction Bank has successfully deployed DeepSeek privately, covering hundreds of use cases. CITIC Construction Investment Fund used DeepSeek for REIT due diligence, reducing a workload previously requiring five employees over 70 days to just one employee in 10 days—achieving a 30-fold increase in efficiency.

Securities firms and insurance companies are also catching up: CITIC Construction Securities has launched AI advisory services based on multi-agent systems; PICC Property & Casualty has integrated DeepSeek to build a professional knowledge base; and Ping An’s large model has been invoked 818 million times in six months.

But what’s truly interesting is the third tier—private equity, asset management, and wealth management. They have abundant data, substantial budgets, and high delivery pressures, yet most are still in the POC stage. This isn’t lagging—it’s the startup window of opportunity.

When it comes to startups entering this space, many immediately think of building a financial version of ChatGPT. But this approach carries significant risk, as it simultaneously pits you against three types of formidable competitors.

Model providers will continue to reduce the cost of general-purpose capabilities. Financial data terminals like Wind, Choice, iFinD, and Tonghuashun already possess data and user access; once AI is integrated, it becomes difficult to charge separately for broad financial Q&A. Large financial institutions are more inclined to build their own internal AI middleware platforms, incorporating general capabilities into their existing permission systems.

A startup is fighting head-on, facing attacks on three fronts.

But if you shift your perspective—from the entry point to the operational layer—the situation changes. What is meant by a vertical operational layer? It means deeply integrating AI around a specific role, a specific process, or a specific deliverable. For example: structuring due diligence documents for private equity or investment banking, auditing financial models in Excel, preliminary review of loan application materials, automatically generating compliance checklists, assisting in the review of insurance claims and underwriting documents, and automatically summarizing client meeting minutes.

These directions may not seem as grand as "financial large models," but they are closer to customers' budgets.

What products are worth developing?

In summary, all four conditions must be met simultaneously.

Can handle the data

High-value use cases typically require integration with internal company documents, CRM systems, cloud storage, email, contracts, and approval workflows. Processing only public web pages offers limited value.

The process works.

Financial users won't change their work habits for AI. The product must integrate into the tools they already use: Excel, PowerPoint, Feishu, WeCom, DingTalk, WPS, and CRM.

Submit the documents

Financial institutions don't pay for answers—they pay for materials. Only when you can deliver checklists, memos, decks, and Excel files do they have the willingness to pay.

Retain clear responsibility boundaries

AI must support citation, traceability, permissions, auditing, and human review. It must not provide investment advice, execute automated trades, or replace final approval.

If any one of these four is missing, the product will struggle to enter a real production environment.

If we widen our perspective to look ahead at the next 24 months, I believe there are seven subsectors worth the most attention.

Research and due diligence come first. With abundant materials, tight deadlines, and clear deliverables, this is the closest to Hebbia and Rogo’s approach.

Second is Excel model auditing—investment banks, private equity, credit, and asset management all rely heavily on Excel, where formula errors, hard-coded values, and inconsistent assumptions are common, offering tremendous potential for AI assistance.

Credit approval assistance ranks third, requiring document preliminary review, cash flow analysis, risk extraction, and credit report generation for both banks and non-bank institutions. Compliance review ranks fourth, with tasks such as system comparison, marketing material review, and KYC checks all well-suited for AI assistants that provide citable and traceable support.

Fund administration and financial operations, including reconciliation, valuation, fee verification, and audit documentation, are highly standardized processes with high costs of error.

Insurance claims and underwriting documents are numerous, rules are complex, and review pressures are high, yet manual verification is still required.

Finally, there are the Client Manager and Investment Advisor Copilots—not AI providing investment advice directly, but assisting advisors with pre-meeting preparation, product explanations, meeting minutes, and CRM updates.

These seven directions share a common prerequisite: the product must be auditable, citable, and privatizable.

Financial institutions will not accept "AI probably said this." Where do the numbers come from? Where are the citations? Who has reviewed them? Has the data left the domain? These are prerequisites for procurement decisions. Therefore, from the outset, you must design citation tracing, human approval workflows, data isolation, and audit trails. This is not a compliance cost—it’s a product moat.

There’s an even larger trend: once model capabilities become commoditized, opportunities shift to workflows, connectors, and governance layers. Just as cloud computing turned IT infrastructure into APIs, enabling a new generation of entrepreneurs to build SaaS on top, the same is true today with large models—those who can encapsulate industry-specific workflows on top will gain a competitive edge.

The financial industry features high information density in knowledge work, strict formatting requirements, and strong accountability constraints—characteristics that make it unsuitable for rapid coverage by general-purpose AI. This is precisely the safe zone for startups.

How can a startup enter the market?

Don't start with the platform.

Identify a narrow use case: with real data, a fixed template, a clear deliverable, human review, a departmental budget, and the ability to validate ROI within 60–90 days.

Don't say that:

I am building an AI platform for financial institutions.

Say it this way:

I will first help the PE/FA team automatically structure the Data Room materials to generate a due diligence Q&A, a risk checklist, and a draft IC memo.

The more specific, the easier it is to close a deal.

Is the maximum risk being replaced by big companies?

The general entry point will be replaced. General financial Q&A, basic research report summaries, and simple data queries can easily be covered by large models and data terminals.

But vertical deep processes will not.

Large companies are unwilling to do the messy work for every niche role. The real challenge lies in integrating with the client’s internal systems, understanding their workflow, adapting to their templates, and accompanying them from POC through to production.

These cannot be automatically resolved by a model API.