Difficulties in applying large models in China

When the halo of technology illuminates the entire industry, the dust that falls to the ground first sticks to the dirtiest corners.

ChatGPT emerged, DeepSeek ignited China's AI enthusiasm, and then Agent technology made significant progress in 2025. Companies such as OpenAI, Cursor, and Manus achieved technological breakthroughs through reinforcement learning fine-tuning (RFT) and environmental understanding. Programming Agents evolved into general-purpose agents, and vertical products such as Vantel and Gamma showed great potential.

No one is afraid of missing out on this wave of technology, and industries such as medical care, finance, manufacturing, and retail are also flocking to it, scrambling to "buy computing power, install architecture, and put labels on" in the hope of achieving leapfrog efficiency improvements and cost reductions and increased efficiency.

However, in the process of applying big models, various industries have exposed the common dilemmas of "serious mismatch between cost input and actual value", "stumbling in the technology deployment link", "data governance and system docking like climbing a cliff" and "misalignment between specialized needs and general big models".

This time, we will try to stand at the current time node and analyze the application pains and future thinking of big models in different scenarios through real cases.

1.  Medical industry: All-in-one devices are selling well, but clinical use is difficult to achieve

A “massive” deployment boom

At the beginning of this year, the DeepSeek large model came out. In just three months, thousands of hospitals officially announced local deployment. Manufacturers concentrated on promoting slogans such as "the computing power cost is only one-tenth of OpenAI" and "the all-in-one machine kills all competing products on the market". Many hospital information department heads immediately made decisions: the budget increased from hundreds of thousands to millions, and there was a tide of "buying computing power and getting on board".

However, after millions of dollars of hardware costs were spent in the computer room, very few doctors actually "enter" the model interface. According to the directors of the information departments of several tertiary hospitals, the large model all-in-one machine is ultimately used for drafting administrative documents and assisting in scientific research, which is far from the real clinical scenarios of "assisted diagnosis" and "image analysis".

What is the reason? On the one hand, doctors are afraid to trust the model because of its "illusion" and "high error rate". On the other hand, the data formats and standards of more than 100 HIS, PACS, LIS and other systems in the hospital are different. It is like asking the big model to "move mountains with bare hands" to feed massive medical records, images and test reports. The information department often spends several months and millions of yuan, and the results are still unsatisfactory.

The huge contrast between cost and value

Take a top tertiary hospital as an example. In order to deploy the full version of DeepSeek, they invested about 5 million yuan in NVIDIA H100 graphics card hardware alone. Ideally, this computing power can provide doctors with "second-level response" and "multi-modal fusion" instant help in pre-consultation, electronic medical record generation, and personalized treatment plan recommendation.

However, when it was actually applied in clinical practice, doctors found that: for the same imaging report, manual reading takes several minutes, while the conclusion given by the large model often takes another five minutes to verify, and the error rate is as high as more than 20%; the generated diagnostic suggestions contain outdated information that does not meet the latest guidelines. The result is "longer time and higher risk", and medical staff would rather choose "traditional processes" than take the risk of being misled by the model.

At the same time, the hidden costs of hospital information system integration are an invisible iceberg: the medical records, tests, and imaging data formats of frontline departments in hospitals vary, and even in the same system, the same field may be labeled differently in different departments. "Hemoglobin" may be written as "Hb", "HGB", "hematocrit", etc. Some handwritten test sheets are simply full of abbreviations and typos.

To integrate such unstructured data into the big model pipeline, the Information Department needs to mobilize an entire data governance team to proofread, label, and format each item, and then spend several months to complete fine-tuning training; and these investments are often exchanged for only a stopgap measure of "drafting administrative documents", which is a far cry from real high-frequency use in clinical practice.

An attempt to break through the disease-specific vertical model

Faced with the "high opening and low closing" of general large models in clinical applications, many leading tertiary hospitals have begun to explore the feasibility of "specialized vertical models".

Shanghai Renji Hospital and Ant Group spent more than a year polishing a vertical large model for urology. After feeding the model with 2,132 expert-confirmed question-answer pairs and more than 25,000 diagnosis and treatment evidences, the model achieved a diagnostic accuracy rate of nearly 69.8%;

The RuiPath pathology model jointly created by Shanghai Ruijin Hospital and Huawei relies on a library of millions of digital pathology slices, and has an accuracy rate of over 90% in common pathology problems. This type of in-depth attack on specialized diseases allows the big model to be "big and comprehensive" while eliminating the risk of errors and focusing on "deep application in a certain department", becoming the most recognized and preferred path within the hospital.

However, even the specialized model is not a “once and for all” solution: it still faces the pain period of high data collection threshold, long labeling process, and the need for two-way collaboration between clinical and technical personnel.

Only when grassroots hospitals and township health centers can effortlessly connect to these small-scenario models, will it be possible for them to be further connected from the top tertiary hospitals downwards, thus realizing a closed loop of "tiered diagnosis and treatment + AI empowerment".

2. Financial Industry: Risk Control, Smart Investment Advisors and the “Ghost of Compliance” 

AI risk control is popular, but it is difficult to move forward due to the "compliance red line"

The earliest attempts at large-scale models by financial institutions often focused on the areas of "intelligent risk control", "anti-fraud" and "intelligent customer service".

A major bank once brought in an AI company to deploy a language model with a large number of parameters in the computer room of its headquarters, claiming that "only one query is needed to make real-time predictions of abnormal customer transactions." However, once the model output was connected to financial supervision, it was found that "the decision logic of the large model was too black box," making it difficult to provide regulators with a "traceable and explainable" result.

The regulatory authorities require that every risk control decision should be able to provide a closed-loop audit report "from data to conclusion". However, when generating recommendations, large models often rely on probability distribution as output, lacking a "causal chain explanation".

Therefore, the bank removed the big model from the core risk control link and switched to an "auxiliary analysis" mode, allowing the model to generate a preliminary report, and then the manual risk control team conducted a secondary verification of suspicious points. The result was that the model's screening efficiency was only about 10% higher than that of the traditional rule engine, but the compliance cost increased by about 30%, requiring an additional audit team to be assigned to the model, and more personnel to be invested in proofreading and interpretation.

In the end, the bank returned the application scenarios of the big model to "automatic answering of customer hotlines" and "intelligent document retrieval", which was far below the initial expectation of "full-process risk control renovation".

Smart investment advisors are highly praised for their “precision”, but data isolation becomes a natural barrier

Another joint-stock bank once tried to introduce a large model into its personal wealth management subsidiary to create a "smart investment advisor + customized portfolio" service. The project team obtained tens of thousands of historical asset allocation and transaction behavior data of customers, and tried to let the model learn and give personalized asset allocation suggestions.

However, in the face of financial data security management regulations, this batch of data can only circulate in a strict "desensitized sandbox environment", and core transaction details and personal sensitive information cannot be shared with large models at all. Even if the external API call is fully encrypted, the regulator still requires that if the model is to participate in actual investment advisory services, all details such as prediction logic, principal risk warnings, and historical performance differentiation must be transparently disclosed. The inherent "black box" and "illusion" characteristics of large models cannot meet such audit and explainability requirements at all.

In the end, this investment advisory project degenerated into "the suggestions given by the model are for reference only, and the final decision is still checked by the manual investment advisory team." Under the gimmick of endorsing "smart investment advisors", the bank needs to invest a lot of compliance meetings, legal reviews and secondary verifications, but still cannot promise that "the model output is safe and reliable." Therefore, the project team refocused its attention on low-risk links such as "intelligent document writing" and "automatic generation of product descriptions." The distance between courage and ideals was gradually pulled further and further away by the regulatory red line.

The emergence of the "specialized model" in the financial industry: vertical segmentation vs. universal problems

Similar to the medical industry, some financial institutions have begun to explore the idea of ​​"vertical large models". A state-owned fund company has joined hands with a well-known securities firm to incubate an NLP model specifically for listed companies' financial reports. It is only "fed" with the company's annual reports, quarterly reports, research reports and securities firm seminar records for the past five years. The model can make a preliminary prediction of the company's future revenue growth and cash flow risks within 10 seconds.

However, this model can only solve the specific scenario of "financial report text analysis". Once it is expanded to "macroeconomic forecasting", "industry trend analysis" or "compliance auditing", it is necessary to prepare a large amount of labeled data again, which has a huge development cost and is difficult to reuse.

Compared with general large models, financial vertical models can indeed run through a specific scenario faster, but if they are to be fully rolled out in multiple business lines such as "bank bill risk analysis", "credit approval automation", and "investment portfolio intelligent optimization", they still face the same "data island + compliance audit" and "multi-system docking + task switching" problems.

Before they move from the "sandbox" to the production environment, financial institutions often suspend the projects due to changes in business needs, which causes the "pioneering vertical model solutions" to frequently fail to meet expectations during implementation.

3. Manufacturing industry: Digital transformation in the first half, but confusion in the second half of large-scale model implementation 

Ideal smart manufacturing: predictive maintenance, quality control, digital twins

The manufacturing industry's expectations for AI are mostly concentrated in areas such as "equipment predictive maintenance", "production line quality monitoring", "digital twin simulation", and "supply chain optimization".

A large household appliance company introduced a general large model, hoping to build an "intelligent production assistant" through massive equipment operation logs, a large number of process parameters, and historical order delivery data.

In theory, the model should be able to immediately provide "possible causes + corresponding maintenance plans" when abnormal vibrations occur in the equipment, or provide "optimal production scheduling suggestions + cost comparison" when production plans fluctuate.

The “invisible wall” between data boundaries and industrial control systems

But when they tried to connect to multiple industrial control systems such as SCADA, MES, and ERP, they found that all the data was wrapped in layers of "factory intranet + proprietary protocol".

To decrypt, format, and filter these data and then import them into the large model for training, it is necessary to repeatedly confirm the interface definition with the automation factory and OT (Operational Technology) team. Even the field of a temperature sensor must be disassembled from the TCP/IP layer downwards.

For traditional manufacturing companies, the IT department is often only responsible for ERP system maintenance. When it comes to "letting the big model take a look at the equipment operation log", it is like asking an old chef to learn programming. There is no common language between the two.

In the end, this home appliance company spent half a year just to complete the "historical collection of key equipment logs", but before it had time to fine-tune the model, it was interrupted by the emergency transformation plan of the production line, and the project fell into a cooling-off period of "no one cared".

Intelligent factories vs. talent and thinking gap

Even if the data is connected to the big model, there is still a gap in thinking between process engineers and the algorithm team: the probability output of "the equipment may fail next" lacks sufficient operability for the front-line workshop; on the contrary, the conclusion that "when a certain process parameter increases by 0.5%, the yield may decrease by 2%" is more convincing.

This "top-level planning dream of intelligent manufacturing" is often out of touch with the "practical needs of the production site", resulting in the large model results becoming "a beautiful vision on high-level slides" but no practical implementation methods in the workshop.

In order to solve this situation, some companies have begun to try to combine the "vertical marketing model" with the "process data special model". For example, a well-known domestic auto parts factory has jointly developed a small vertical model for "stamping equipment fault diagnosis" with universities, focusing on monitoring the vibration signals, mold wear rate and temperature parameters of stamping machine tools. The model has achieved a fault warning accuracy of more than 80% in a laboratory environment, but to promote it to more than a dozen production bases across the country, it is necessary to transform sensors and upgrade networks for each stamping line, and invest a lot of training costs to let on-site technicians learn "how to read model warnings and how to judge the authenticity of alarms." In the short term, these factories are not enough to see a significant ROI (return on investment), resulting in the factory management's lack of confidence in the large model, and the project often stops at the "internal laboratory" stage.

4.  Retail industry: personalized recommendations and inventory optimization, with mixed results

The tug-of-war between “one thousand faces for one person” and cold start

The retail industry was the first to apply big models to "personalized recommendations", "intelligent customer service" and "supply chain forecasting".

An e-commerce giant once built a universal large model internally, hoping to achieve "full-link connection of user browsing, adding to cart, evaluation and predicted demand", so that the model can push personalized homepages to new users based on massive user behavior data.

However, there is an inherent "cold start" problem: when a new user first registers, there is not enough historical behavior data to learn from, and the model is simply unable to make high-quality recommendations. It can only make general recommendations based on the overall clustering logic, resulting in the actual effect being not much better than traditional collaborative filtering.

On the other hand, when the model's recommendation logic is to be connected to offline stores, there are data silos in the POS system, CRM system, and supply chain WMS system. A supermarket chain tried to connect the online big model with offline stores, hoping to provide customers with a shopping experience of "seeing real-time inventory + personalized discounts when you arrive at the store" on the store APP.

However, since the card swiping system and the QR code scanning system in stores are independent of each other and there is no unified middle platform, the model often experiences data delays or conflicts when reading "real-time inventory" and "membership tags". Once the recommended coupons do not match the inventory, customers will have a bad experience of "the coupons are available on the APP but sold out in the store".

The tug-of-war between price war and model ROI

Profits in the retail industry are generally thin, and many leading companies blindly pursue the gimmick of "technological support". As a result, large model projects "either spend money to buy computing power and refresh data logs, or get stuck in the quagmire of data cleaning", but do not bring a significant increase in final sales.

In order to promote the "big model smart matching" function, a certain chain clothing brand has spent hundreds of thousands of dollars on portrait recognition and product attribute labeling; after it went online, less than 10% of users used the "smart matching" function, and most people were still accustomed to manually browsing store styles.

In the short term, this intelligent matching model increased revenue by less than 1%, while its maintenance cost accounted for 20% of the annual IT budget. At a time when price wars are intensifying, such attempts that "cannot see returns in the short term" are easily labeled as "useless projects" by store operators.

Social e-commerce and the vitality of vertical models

However, there are also some retail companies that have achieved breakthroughs through vertical large models.

A new beauty e-commerce company has partnered with an AI startup to launch a "small language + deep learning" "intelligent makeup trial + skin care Q&A" system: the front end of the system recognizes skin quality and skin color through facial selfies, while the mid- and back-end models combine user skin quality input, product ingredient database, and seasonal climate data to give "personalized makeup trial" and "skin care suggestions."

Since this system only focuses on the "beauty" field and is fully linked with the store shopping guide system, it not only greatly improves user stickiness, but also brings about a 15% increase in average order value.

Compared with the general large model that costs a lot but only provides mediocre experience in the "home page recommendation" section, this vertical and segmented approach is more likely to achieve a quick return on ROI in a "precise scenario".

5.  Common Dilemmas in Large Model Applications

The common dilemma, in a word, is that it is difficult to match cost investment with the actual value.

High hardware and operation costs

Medical: Deployment of large-scale all-in-one machines, the computing power cost of a single machine can easily range from hundreds of thousands to millions. Once the utilization rate is insufficient, it becomes a "high-priced ornament."

Finance: In order to meet compliance audit requirements, dedicated GPU clusters, log audit systems, and compliance teams are needed to barely figure out "why the model gave this conclusion", which accounts for a huge proportion of the cost.

Manufacturing: To connect to the industrial control system, it is necessary to modify the PLC interface, deploy new sensors, and upgrade the OT network, and the cost investment exceeds the model development itself.

Retail: Infrastructure investment such as offline store data integration, POS upgrades, and CRM integration, once the scenario does not work, it will lead to the embarrassment of "spending money to buy a big data platform but not getting a single drop of actual sales."

The difficulty of implementing technology is extremely high

Medical care: Data is scattered and non-standardized, and model output needs to be verified by clinicians, which delays diagnosis and treatment efficiency.

Finance: The regulatory red lines are high and the interpretability requirements are strict, and the “black box” characteristics of the model are difficult to implement in the core risk control links.

Manufacturing: The industrial control system is fragmented, and IT and OT are split. It is like "shooting a cow from a distance" for the model to read production data.

Retail: Online and offline data are disconnected. The forecasting model needs to be connected with the inventory system, ERP system, and membership system, which requires cross-system architecture reform.

Industry specialization vs. general model misalignment

The general large model solves "encyclopedia-like problems" and "multi-field coverage", but it is difficult to go deep into the "segmented pain points" of each industry.

Specialized/vertical models have higher data quality and scenario fit, but their versatility is limited, promotion costs are high, and they require continuous participation from industry experts.

6.  Breakthrough ideas for large model applications

The solution to the problem, in one sentence, is to change from “buying computing power” to “buying value”.

Focus on scenarios and implement them step by step

Medical treatment: Start with "special disease model" and "auxiliary diagnosis", and then copy to other departments after successful verification in a certain department.

Finance: First, run low-risk scenarios such as "intelligent document retrieval" and "automatic generation of compliance reports", and then gradually explore high-risk scenarios such as "risk control warning" and "intelligent investment advisory".

Manufacturing: First, run the vertical fault prediction model on a single production line, and then expand it to the entire factory and across factories after the ROI is clear.

Retail: Start with "online intelligent customer service" and "member portrait refinement", and then integrate them into "inventory and supply chain optimization" after the forecasting effect is stable.

Data center and standardization construction are promoted simultaneously

Build a more mature "data middle platform", unify data standards, and open up interfaces between systems.

Promote industry associations or regulatory authorities to formulate industry data standards to reduce the cost of homogeneous transformation.

Medical engineering, metalworking, industry-university, production and marketing collaboration

The medical field requires clinicians, information departments and algorithm teams to form a "co-creation" mechanism. Doctors tell the algorithm what indicators are needed, the algorithm is continuously refined, and doctors constantly provide feedback and corrections.

The financial field encourages risk control teams, compliance teams and technical teams to jointly establish projects to ensure that model outputs are compliant and auditable.

In the manufacturing field, the IT/OT forces need to be connected, so that field engineers and algorithm engineers can "eat at the same table" to discuss requirements and iterate models in a timely manner.

In the retail field, the marketing team and the data team work closely together to formulate "recommendation rules" and "marketing strategies" so that the model can evolve in sync with business needs.

Building an "explainable + traceable" audit system

The model output must be accompanied by an "explanation layer" to let users know "why this suggestion is given" and "what is the basis" to avoid the "black box decision" from encountering a crisis of trust.

At the same time, "full-link auditing" is implemented in regulatory compliance scenarios to record input, output, and logical paths to ensure that once deviations occur, they can be traced and rectified.

The potential of edge computing and lightweight models

For small and medium-sized institutions with limited hardware resources, they can seek a "lightweight pre-training + online fine-tuning" solution to reduce their reliance on computing power.

Placing the model on edge servers, mobile terminals, or directly adopting cloud-edge collaboration can make training and reasoning more flexible and reduce hardware idle costs.

7.  Conclusion: Big model, running on the road is more promising than standing still

If the wave of deep learning is a "tsunami", then the implementation of large model applications is "surfing".

Healthcare, finance, manufacturing, and retail are several major industries that have their own eager expectations, but they are also faced with real challenges such as "data islands", "deployment costs", "misalignment of specialization", and "compliance audits".

The big model is not a panacea that is applicable everywhere, but more like a "semi-finished puzzle" that requires various industries to continuously polish and iterate between scenarios, data and supervision before it can be truly integrated into the business process.

At present, the most important thing is not "who eats the computing power cake first", but "who can put the small piece of the puzzle in hand together first". From the "hot selling" of medical all-in-one machines to the "wait and see" of financial risk control, from the "sandbox" of manufacturing predictive maintenance to the "traffic trap" of retail personalized recommendation, the application dilemma of large models tells us: only with more solid infrastructure, closer to the scene needs and more sustainable operational thinking, can technology and business, data and compliance, algorithms and professional knowledge form a synergy, so that this "controversial puzzle" can be truly integrated into the industry picture, rather than becoming a gimmick of "high-priced decoration".

In the next few years, the "second wave" of big models in various industries may come: the financial risk control system will further polish the explainable model; the digitization of the entire manufacturing process may give rise to a lighter "edge reasoning" solution; the retail industry will continue to explore incremental growth in the "social e-commerce + vertical model"...

When these fragmented attempts are numerous and in-depth enough, the big model may truly make "AI empowerment" move from paper to reality, allowing those computing machines that were once called "high-priced ornaments" to truly become boosters of "igniting business value."

Waiting for this wave requires more patience, deeper collaboration, and a precise grasp of "cost and value."

When the halo of technology illuminates the entire industry, the dust that falls to the ground first sticks to the dirtiest corners.

Only by running through every scenario steadily can the afterglow of the halo illuminate every inch of land for practical application.