Jian Chen: Core Demands and Cost Constraints to Find Out the Optimal AI Solution Suitable for Enterprises | FEMBA

Release time：2025-08-29    

AI big models are like a “double-edged sword,” restructuring social productivity and technological ethical boundaries while simultaneously arousing a series of disputes like data security and hallucination traps. So how can we consolidate the security line of defense amid algorithmic black box and data torrent? How to effectively crack seemingly perfect AI hallucinations? And what exactly is the optimal balance point of localized deployment between cost control and efficiency increase?  

Recently, Dr. Jian Chen, Professor of Practice in Finance at FISF and Researcher of Chengdu Fudan Western International Financial Research Institute, accepted a special interview of FEMBA, expounding the challenges for AI large language models in enterprises applications and corresponding solutions.

Jian Chen

Professor of Practice in Finance at FISF

Teaching Professor of EMBA

Researcher of Chengdu Fudan Western International Financial Research Institute

With the widespread application of AI big models, how to ensure data security?

AI itself is merely a tool beyond good and evil. The essence of data security is about balance: overprotected data won’t lead to good models; while to generate a good model, data can’t be protected too well.

Jian Chen proposed to “identify problems in the process of development and settle the problems in the process of development.” It’s not appropriate to make extremely strict standards upon data protection in early stages so as to avoid stifling possibilities of big language models too early.

So far, Europe has the most rigorous policies to protect data, having made a series of regulating rules beforehand, which enterprises must not exceed by even one step; the US on the other hand, though having also launched certain acts upon data security and privacy protection, leaves relatively loose space overall, tending to make amendments and revisions after specific problems emerge.

What is the “key” to a successful localized deployment?

As big language models become larger in scale through increasingly complex training, it is inevitable that a series of problems will arise, such as overfitting, AI hallucination and poor interpretability.

Chen pointed out that for a model, the simpler the stabler; and vice versa. Currently the Deepseek model has approximately 670 billion parameters, and for Grok-3 and ChatGPT-4, the numbers go beyond a trillion. This is the fundamental reason why AI can “fabricate information.”

As for this problem, Jian Chen further shared two major solutions:

01 Building multiple AI Agents. Break down a complex mission into multiple specified tasks, and respectively distribute the detailed work to multiple AI Agents in different fields, intriguing the powerful efficacy exceeding a single big language model through the mutual collaboration and complementary advantages of multiple AI Agents.

02 Making use of RAG (Retrieval-augmented Generation) and Fine-tuning Technique. Retrieve relevant information in restricted external corpus/repository to help the big model give more precise and comprehensive answers; or conduct relevant knowledge training, optimization and adjustments on the big model beforehand so that it can better fit in the whole-new task fields.

What is the “key” to a successful localized deployment?

According to Jian Chen, whatever method is employed for AI localized deployment, the very first step enterprises need to take is to understand and clarify their own core demands. Only on the basis of their own core demands and relevant cost constraints can they truly find the optimal solutions.

For example, if a financial institution wants to build a research report big model with extremely high requirements for the timeliness and accuracy of data information, the best solution is to add the RAG technology and a professional corpus or news database updated in real time to the general AI big model, so that AI can call the latest data at any time for investment research analysis and report generation.

For another example, for a banking institution planning to design a credit card anti-fraud model, the timeliness of data information is not so important. Instead, it is more crucial to precisely identify various fraudulent information. Therefore, a professional corpus of fraudulent information should be set up in advance and updated irregularly, which, together with pre-training on the AI big model, can ensure AI can identify and capture in a fine-tuned manner all the suspicious fraudulent activities and information as much as possible.

Are AI big models “influencers” or just “passersby” in the financial circle at present?

Financial statement analysis, credit risk analysis, valuation and pricing, investment, mergers and acquisitions, compliance supervision... AI big language models have a wide range of application prospects in various sub-fields of the financial industry.

Meanwhile, Jian Chen also mentioned that due to problems of low accuracy and poor interpretability still existing in AI big models right now, most people prefer to use traditional machine-learning algorithms for core risk management and asset pricing; while in cases of unstructured data and speech emotion analysis, AI big models have already been able to satisfy market demands, assisting institutions to further enhance quality and improve efficiency in internal text processing and compliance management.

Reconstructing Intellect, Navigating the Future. [Series Interviews of FEMBA Professors] is in full swing. Next session we will focus on the heavyweight topic of “management education in the era of AI,” and continue under the guidance of Professor Jian Chen to have a profound reflection on the original value of humans and the ultimate significance of EMBA education, helping more enterprise leaders and elite entrepreneurs to actively embrace the new commercial age and to set foot on new journeys in a deliberate manner. Stay tuned for more!

FISF FEMBA

The EMBA Program of Fudan International School of Finance, established on Fudan’s profound disciplinary foundation, led by “internationalization” and “actual combat” to construct a new educational paradigm of China’s EMBA 2.0, aims to cultivate internationalized business leaders who have a sharp insight into Chinese financial power, take the lead in China’s economic future, and act as the source of energy for individual growth and enterprise development. Right now, 2026 enrollment is in progress. We anticipate to welcome entrepreneurs embracing ambitions and patterns in FISF EMBA, and create infinite possibilities together in the era of uncertainty!

Finance is a pillar of the nation and a compulsory course for modern entrepreneurs.