From Conceptual Frenzy to Pragmatic Evolution

/

Produced by | Business Show

If last year's stories about AI agents were filtered through a sexy lens of 'disrupting imagination,' this year's narrative has shifted to a pragmatic undertone of 'solving practical problems.'

From 2023 to the present, the development of AI agents has transitioned from concept to practice. At this year's World Artificial Intelligence Conference, we witnessed major vendors continue to unveil their new generation of AI agent products, increasingly implemented in vertically segmented fields such as finance, education, and entertainment.

Many people refer to 2025 as the 'first year of AI agents.' However, the concept of AI agents dates back to before 2023.

Wang Wei, CTO of AntChain, noted in an interview with 'Business Show' and other media that November 2024 was already dubbed the 'first year of AI agents,' and by July of this year, the peak of this inaugural year was nearing its conclusion.

Wang Wei's implication is that the speed of technological iteration is incredibly fast, especially for AI agents based on large models. Take ChatGPT as an example; it reached 1 billion users and achieved 36.5 billion queries, 14 times faster than Google.

In more specific industry sectors, such as the transformation of AI agents in the financial industry, is also accelerating. However, for AI agents to truly take off, they still face numerous challenges. At the very least, they need to surmount the triple thresholds of 'technical reliability, data controllability, and ecosystem coordination.'

01 Three Years of Wild Running for AI Agents: From Concept to Industrial Penetration

From 2023 to 2025, the evolution trajectory of AI agents has traced a steep curve. Before 2023, the development of AI agents was still in the conceptual budding stage, evolving from 'tool-based' to preliminary intelligence.

Wang Aihua, Deputy Chief Engineer of the China Academy of Information and Communications Technology, believes that before 2023, AI agents were still in the stage of 'agent-type tools,' with even customer service robots only capable of providing mechanical, standardized responses. At that time, AI agents struggled to complete even the simplest cross-scenario conversations, let alone make autonomous decisions.

2024 became a pivotal turning point in the development of AI agents. The infusion of large model technology suddenly equipped AI agents with basic capabilities of 'reasoning-memory-action.' For instance, some first-generation financial AI agents launched in 2024 were already capable of completing semi-automated processes of 'customer consultation-demand matching-business handling' in financial scenarios.

However, core risk control processes still required manual 'checkpoints,' and their reliability and scenario adaptability needed further enhancement.

Entering 2025, AI agents have exhibited a markedly different demeanor, beginning to enter the stage of industrial penetration, with vertical deepening and large-scale implementation. AI agents at this stage primarily present three core characteristics. First, there is a shift from 'generalization' to 'industry specialization.' For example, many deeply customized AI agents have emerged in fields such as finance, energy, and industry. AntChain has launched the Agentar full-stack enterprise-level AI agent platform, which has jointly developed over 100 financial AI agent application solutions across various industries through capabilities such as knowledge engineering, evaluation, security risk control, MCP, and large financial models, implementing them in core scenarios like intelligent risk control, marketing, and wealth management.

The industry believes that industry scenarios have far higher requirements for the 'professionalism' and 'reliability' of AI agents than for general capabilities. For example, the financial sector needs to strictly avoid 'model hallucinations,' while the industrial sector needs to adapt to extreme environments such as high temperatures and high pressures.

Second, multi-agent coordination has become the mainstream model. By coordinating multiple AI agents, the 'capability boundary' issue of single AI agents is addressed, adapting to the diverse needs of complex industrial scenarios, such as the coordination of 'marketing agents + risk control agents + compliance agents' in the financial sector, thereby covering the entire business chain. Third, there is an upgrade from 'auxiliary tools' to 'productivity engines.' It is reported that some banks have deployed over 1,000 AI agents, among which credit risk identification AI agents can increase the credit efficiency of small and medium-sized customers by 10 times, and data dynamic AI agents enable 'one-sentence invocation of bank-wide data,' promoting a shift in business models from 'people seeking services' to 'services seeking people.' In other words, AI agents at this stage are no longer limited to 'cost reduction and efficiency enhancement' but are creating more new value through data insight and process reconstruction.

02 Finance and Energy: The Battleground for AI Agent Implementation

In this year's WAIC industry map, AI agent applications in the financial and energy sectors stand out the most.

What is the core logic behind this?

Wang Wei believes that these two industries have the highest degree of digitization, the highest data density, and the most urgent need for efficiency improvements.

Taking finance as an example, with daily transaction data in the tens of millions and multi-level risk control rules, the error rate of traditional manual processing exceeds 3%, while AI agents can reduce this figure to less than 0.5%.

However, the implementation process is not without challenges. For example, 80% of financial institutions only test AI agents in non-core scenarios such as customer service, with core processes like clearing and risk control still relying on manual labor.

This caution stems from dual anxieties – they are both worried about missing out on the AI dividend and afraid of security risks arising from immature technology.

In other words, the anxiety of many institutions does not stem from rejecting AI but from worrying about how to safely and effectively apply this new technology to their business scenarios to truly solve problems. At the same time, they see peers achieving results with AI in their business and look forward to achieving 'overtaking in curves' through AI agents.

Regardless of the choices made by financial institutions and banks, the complexity of financial scenarios cannot be avoided.

How does AntChain respond to this? Its strategy revolves around three 'E's.

First is Expertise. Instead of following the path of general large models, it formulates a financial task system covering six major categories and 66 sub-categories across full scenarios such as banks and securities based on long-term financial experience. Using this framework, a professional training dataset is constructed from hundreds of billions of data points, with the addition of principle-based synthetic data to ensure compliance, making the model an 'expert right out of the factory.'

Second is Efficiency. During training, resources are dynamically allocated to enhance the performance and learning efficiency of complex financial tasks, achieving 'shallow tuning for high performance,' ensuring that general capabilities do not degrade, reducing subsequent data and computing power consumption for fine-tuning in business applications, and lowering the threshold for enterprise implementation.

Third is Evolution. A high-frequency agile iteration mechanism is established to continuously absorb information on financial policies, market dynamics, etc., quickly fixing model issues, ensuring that knowledge, capabilities, and compliance keep up with industry changes, and evolving continuously in real business scenarios.

Breakthroughs in the energy sector also rely on scenario adaptation. According to Wang Kuanxin, General Manager of Industrial AI Technology Management at Supcon Technology, its industrial AI agents have achieved unmanned operation of refining units through a combination of 'time-series large models + edge control.' Nine AI agents work together and can support the autonomous operation of units for over a week, a significant breakthrough in high-temperature and high-pressure industrial environments.

It is reported that AntChain has jointly launched over 100 financial AI agent solutions with partners in the financial industry, covering four major areas of banking, securities, insurance, and general finance. Financial institutions can 'plug and play' these solutions, improving the work efficiency of frontline employees by over 80%.

03 Triple Thresholds: A Tough Battle for Technology, Data, and Ecosystem

Despite their rapid development momentum, the large-scale implementation of AI agents still needs to surmount three formidable thresholds.

Computing power remains a 'sword of Damocles' hanging over the industry. Zheng Weimin, Professor of Computer Science and Technology at Tsinghua University, directly points out the 'pain point' – large model inference relies on GPU clusters, and the current cost of inference computing power for large models is still high.

He explains, 'Whether it's inference or training, the cost of computing power is still very expensive. Within the inference cost, manpower accounts for 3%, data for 2%, and computing power for 95%. Most of the money is spent on computing power. The daily inference cost of ChatGPT is $700,000. The daily inference cost of DeepSeek V3 is approximately $87,000.'

The industry believes that large models themselves have high costs. Although related hardware and technology are developing, compared to past software sales, cost remains an important factor for financial institutions to consider.

The inference efficiency of a single card may be slow, but multiple invocations place extremely high demands on underlying computing power. How to balance the cost of technological investment with commercial value has become an issue that financial institutions and technology providers need to address.

The issue of model 'hallucinations' is even more troubling for financial institutions. Especially in scenarios such as credit approval, AI agents occasionally provide erroneous information, and customers generally require 'you must explain the reason for every decision.' This demand for explainability is difficult to meet solely through Prompt technology.

Zhang Peng, General Manager of the AI Algorithm Technology Department at AntChain, explains that customers need to understand the thought process behind the large model's responses and require explainability, necessitating a reasoning model to address this. Before this, only Prompt could be used to force the model to think, but the effect was not ideal.

This is also the reason why AntChain, in collaboration with institutions such as the Industrial and Commercial Bank of China, Bank of Ningbo, Beijing Frontier Institute of Financial Regulatory Technology, and Shanghai Artificial Intelligence Industry Association, jointly launched the Finova large model financial application evaluation benchmark.

In other words, using stricter testing to force the improvement of model reliability. 'Customer demands are driving technological evolution. From initially asking 'why don't you have a reasoning model?' to now asking 'why should I believe your answer?' market requirements are becoming more specific and stringent,' said Zhang Peng.

Challenges at the data level are equally thorny. Li Nan, Vice President of Greatech Intelligent, complained, 'The data standards in the manufacturing industry are chaotic. For the calculation of the same product's qualification rate, Factory A and Factory B can differ by three versions.' This chaos results in 70% of effort being spent on data governance, severely slowing down the implementation pace of AI agents.

Finally, at the level of industrial implementation, there are also challenges in building trust and ecosystem coordination. First, there is a strong wait-and-see attitude in the industry. Financial institutions only test AI agents in non-core scenarios (such as customer service), with core businesses (such as clearing and risk control) still relying on manual labor. Some institutions are concerned about 'insufficient technological maturity' and 'ambiguous responsibility definitions.' For example, when an AI agent's decision-making mistake leads to losses, it is difficult to clarify responsibility attribution.

The fragmentation of ecosystem coordination is another obstacle. Because the AI agent industry chain involves multiple links such as hardware, models, and applications, it is difficult for a single enterprise to cover full-chain capabilities. For example, banks need to coordinate with model providers, system integrators, regulatory agencies, and other parties when deploying AI agents, with high coordination costs.

To reduce costs, AntChain has launched a full-size model family and provided customers with models of different specifications, such as the 32B and 8B versions of the financial inference large model Agentar-Fin-R1 and the MOE architecture model based on the Bailing large model, allowing customers to choose based on their own computing power and scenario needs, balancing cost and effectiveness.

Conclusion

Currently, AI agents are still in a rapid development stage, but to achieve a true breakthrough, they need to make significant strides in technological improvement, cost control, and meeting differentiated needs.

A research report released by the CITIC think tank points out that in 2025, AI large models will develop towards stronger, more efficient, and more reliable directions, presenting a pattern of deepening reasoning models and the proliferation of AI agent models.

This year is also a year of accelerated AI application implementation, but the story of AI agents this year has been less 'sexy' and more 'tough battles.' This may well be the inevitable path for technology to change the world.

As Tu Guangshao, Chair of the Shanghai Finance Institute and former Executive Vice Mayor of Shanghai, noted at the 2025 World Artificial Intelligence Conference forum, "The genuine worth of AI agents is not measured by the allure of the concept but by their ability to transform into a 'viable productivity force' that drives industrial evolution."

When the industry shifts its focus from discussing "what AI agents can do" to concentrating on "how AI agents can address scenario-specific problems," the true industrial potential of AI agents will begin to be fully realized. [End]