Summary of Key Points
In 2026, the implementation of large-scale financial models has entered a "race period," but most projects are still at the "proof of concept" (POC) or internal demonstration stage, with very few generating actual business value. The problem does not lie with the models themselves, but rather with three major challenges on the data side (disorganized data, difficulty for AI to understand the data, and challenges in application) and a critical gap on the organizational side (lack of teams that can integrate business, data, and engineering capabilities). The solution lies in rebuilding data sovereignty through a three-tier approach: "data governance → AI readiness → application implementation." This approach replaces comprehensive data management with a "scenario-driven, incremental" approach, and breaks down organizational barriers with the FDE (Frontier Deployment Engineers) model and four key types of professionals.
I. Implementation of Financial Large Models: Much Hype, Little Actual Output
Leading financial institutions are all working on large-model projects, covering areas such as investment research, risk management, and customer service, but most are still at the conceptual verification stage (for example, demonstrating to management that AI can answer simple questions). Why isn't there more practical use? It's not that the models are ineffective—general large models are improving rapidly and becoming more cost-effective. The real issue is the data, as well as the lack of professionals who can connect business, data, and technology.
II. Three Major Data Challenges Hindering Implementation
Even with advanced models, they struggle to function effectively due to three key data problems:
Challenge 1: Disorganized Basic Data
Financial institutions have numerous systems (e.g., different vendors for banking credit and wealth management, and incompatible systems for securities brokerage and asset management), leading to:
- The same customer having different identifiers across systems (e.g., "Zhang San 123" in the credit system and "Mr. Zhang 456" in the wealth management system, making it impossible for AI to recognize them as the same person).
- Inconsistent data formats (some in tables, some in text, with frequent changes).
- Poor data quality (it's unclear how certain indicators are calculated, and any changes require manual email notifications, increasing the risk of errors).
In short, the data is not organized in a way that AI can effectively use.
Challenge 2: Data Not in a Format AI Can Understand
Even if the data is available, it's not in a format that AI can process directly. For example, while a financial report shows "net profit of 500 million," AI doesn't understand what "net profit" means, where the figure comes from, or how it relates to other indicators. This creates three problems for AI: difficulty in finding relevant data, inconsistent interfaces, and an inability to understand industry-specific terms (e.g., the meaning of "leverage ratio" varies between banks and securities firms).
Challenge 3: AI Generates Inaccurate Results
General large models lack specialized financial data and update slowly, often leading to incorrect information (e.g., miscalculated price changes). Additionally, strict regulatory requirements mean that only specific data can be used. For instance, when using AI for bond rating, it might generate arbitrary numbers without verifiable sources, which is unacceptable in the financial industry.
III. Rebuilding Data Sovereignty: Three Steps to Make Data Work for AI
To solve these data issues, a three-tier approach is needed:
1. Data Governance: Organize the Data
This involves more than just assigning consistent identifiers; it requires addressing three issues:
- Unified Master Data: The same customer/business entity should have a consistent identifier across all systems (e.g., "Zhang San 001").
- Clear Data lineage: Track the origin of each data point (e.g., whether it comes from financial reports or third-party reports) and maintain version control.
- Standardized Indicators: Define indicators (e.g., EBITDA) clearly and specify their use cases to prevent misuse by AI.
This cannot be done by technical personnel alone; business and data experts must collaborate to ensure that indicators are used correctly in practical contexts.
2. AI Readiness: Convert Data into a Format AI Can Understand
Even well-organized data needs to be translated into a format that AI can process. This involves creating a "business knowledge framework" that defines terms clearly (e.g., clarifying the difference between "leverage ratio" in banking and securities contexts) and labeling data (e.g., classifying "net profit" as a "financial indicator" to indicate profitability).
3. Application Implementation: Restrict AI's Actions
To prevent AI from producing incorrect results, measures such as decoupling complex processes into smaller steps (e.g., breaking bond investment into steps like bond selection, rating, and trading) and enforcing constraints (e.g., using only verified data and automating high-risk decisions) are necessary. This ensures that AI uses accurate data and can be easily audited.
IV. Practical Implementation Strategy
Waiting to complete all data governance before deploying large models is unfeasible, as data continues to evolve. The better approach is to focus on high-value use cases (e.g., bond rating adjustments) and implement a "data governance → AI readiness → application implementation" cycle. This not only solves practical problems but also builds experience for future use in other scenarios. For example, a bank's data platform has integrated more than 300 systems and established 2,900 data standards, using a combination of technology, standards, and governance processes.
V. The Key to Success: Collaborative Teams of Four Types of Professionals
Having the right data and technology in place is not enough; a collaborative team of four types of professionals is essential:
1. Data Governance Experts: Organize data into usable assets (e.g., by establishing standards and tracking its origin).
2. Industry Business Experts: Understand financial processes and convert business challenges into data requirements.
3. AI Full-Stack Engineers: Develop AI systems that can handle these requirements.
4. High-Level Consulting Experts: Create reusable templates based on project experiences.
While these professionals are not scarce individually, coordinating their efforts is challenging. Leading institutions are already training FDE (Frontier Deployment Engineers) to bridge the gap between technology and business.
VI. Conclusion
The success of financial large models depends on a combination of data, use cases, and skilled teams. The focus should not be on the models themselves, but on creating a seamless data flow and leveraging experience to build practical capabilities. Only by doing so can large models truly generate business value.