Bank AI is not what you think it is

Introduction: Since launching the "All in AI" strategy, AI capabilities have reached the core of Bank of Beijing's operations.

On one side is fire, on the other is ice.

How big is the temperature difference between the two? Open a stock account and take a look to find out.

Chips and computing power are being fiercely pursued by capital, while traditional economic sectors are neglected. The valuation ceiling for businesses reliant on manual labor is right there.

However, the divide is not insurmountable.

The commercialization of AI ultimately relies on the real economy to absorb it. Rationality and judgment, the two most core assets of the financial industry, are migrating from human experience systems to data and algorithm systems.

Ray Dalio, founder of Bridgewater Associates, once wrote in Fortune magazine that AI will affect almost everything, and humanity stands on the edge of a new era; "the days when humans make decisions solely based on personal judgment are coming to an end."

Inside Bridgewater, Dalio has translated decision-making principles into algorithms, with computer systems assisting investment judgments. Bridgewater's AIA Macro Fund exceeded $5 billion in scale by the end of 2025, performing on par with strategies led by human fund managers, while providing independent alpha returns.

A survey released by Barclays in June this year covered 410 global investment institutions, showing that 72% of hedge funds use AI tools daily. The most widely applied area is research and analysis, followed by security screening.

In the banking sector, which also belongs to the financial industry, AI applications are accelerating as well.

Goldman Sachs' "One Goldman Sachs 3.0" embeds AI into six core workflows, with over half of employees using the GS AI platform daily. JPMorgan Chase invests over $17 billion annually in technology, supporting the implementation of about 450 AI applications. ING Bank uses a centralized AI platform to push the success rate of pilot projects to 90%.

Bringing the focus back to China, the issues become more concrete: For the vast majority of Chinese banks, tech investments amounting to tens of billions of dollars represent an unreachable threshold.

High thresholds are just one aspect. More tricky is that for the domestic strict regulatory environment, introducing AI into finance inherently brings new risk exposures. Deep neural network decisions are like black boxes, and large models may generate hallucinations, producing unreliable results. Quality defects in training data will be systematically amplified, and algorithms may produce biased or discriminatory outcomes. Moreover, once AI is deeply embedded in core links such as credit approval and risk assessment, technical failures could propagate along the business chain, amplifying systemic risks.

Is there a path that is both feasible and safe?

On June 18, the National Financial Regulatory Administration issued the "Guiding Opinions on the Safe Development and Application of Artificial Intelligence in the Banking and Insurance Industries" (hereinafter referred to as the Guiding Opinions). This is the first special regulatory document for the full process of artificial intelligence in China's banking and insurance industries.

The "Guiding Opinions" clarify that financial institutions developing and applying AI should adhere to four major principles: user responsibility, autonomous controllability, pragmatism and efficiency, and safe development. It proposes 32 guiding opinions from seven aspects: governance architecture, development and application, data governance, computing power construction, risk management, capacity enhancement, and guarantees and supervision. The core orientation is to guide the industry to start from actual business needs, solve real problems, and create real value.

Bank of Beijing's (601169.SH) financial practice path aligns highly with the regulatory direction. Compared to the grand narratives of Wall Street, it may be closer to the true operational reality of China's banking industry.

As a city commercial bank with an asset scale of 5 trillion yuan, Bank of Beijing has run through a complete system—from computing power base to agent matrix, from organizational reconstruction to business restructuring—using less than 5% of its revenue for investment.

01 "All in AI"

Traditional banking business models are built on information asymmetry.

Relationship managers hold enterprise information, risk control experts possess approval authority, and tellers execute standard operations. This human-centric system naturally suffers from manpower bottlenecks, experience gaps, and efficiency ceilings. The larger the scale, the higher the management costs. Marginal benefits diminish, and growth curves inevitably flatten.

Bank of Beijing's "All in AI" strategy attempts to break this ceiling. By using computing power to share the burden of labor, algorithms to assist experience, and agents to optimize processes, it transforms the bank from a labor-intensive organization toward a technology-driven direction.

This transformation is driven by top-level design and implemented from top to bottom.

In February 2025, Bank of Beijing fully launched its "All in AI" strategy, led by the top executive, establishing a special task force mechanism, significantly compressing the decision-making chain.

This is consistent with the "Guiding Opinions" requirement for banks' governance architecture to "strengthen top-level design and coordinated management."

Digital transformation has its stages. Early digitalization focused on "online-ization," solving the problem of "whether systems exist." At that time, AI was merely an efficiency tool, assisting humans in completing work.

After three years of transformation, Bank of Beijing entered the "Comprehensive Digital Operation 2.0" stage, where the core orientation shifted from "system launch" to "value creation."

"All in AI" is a key lever in this stage. The role of AI has gradually upgraded from business assistance to becoming an engine driving growth.

As Ming Lisong, Chief Information Officer of Bank of Beijing, said at the earnings briefing: "Investing in technology is investing in the future."

Over the past few years, Bank of Beijing's technology investment accounted for a stable proportion of over 4.5% of revenue, further increasing to 4.8% in 2025, placing it at a relatively high level among peers.

These investments have yielded a series of infrastructure upgrades: technology stack transformation, breaking down system silos, building an enterprise-grade data foundation, upgrading security defenses, implementing forward-looking AI layouts, and forming a digital talent team. Essentially, this is a bank upgrading its underlying architecture.

02 Computing Power and Agents

It took five years for Bank of Beijing to complete the leap from "buying computing power" to "building an ecosystem."

During these five years, Bank of Beijing's insistence on self-research and domestic adaptation directions remained consistent with the "autonomous controllability" principle in the subsequently issued "Guiding Opinions."

In June 2021, the "Jingzhi Brain" project was launched, aiming to build an enterprise-wide artificial intelligence platform. Two years later, the third phase of "Jingzhi Brain" was completed and rolled out to various branch institutions. That same year, the Financial Intelligent Innovation Laboratory was created, beginning the construction of a domestic computing power base.

2025 was a key turning point.

In the first half of the year, Bank of Beijing built the "1213" AI system, including an integrated computing power base, two major model development and operation platforms, over 100 AI capabilities, and over 300 AI application scenarios. Computing power, demand, and capabilities were managed in coordination for the first time.

In the second half of the year, a fully self-developed integrated heterogeneous computing power scheduling platform was deployed. From computing resources to front-end applications, a complete link was connected.

In March 2026, Bank of Beijing released the "Intelligent Computing Platform," compatible with heterogeneous domestic computing power, coordinating the scheduling of large and small models.

At the large model level, Bank of Beijing is one of the first institutions in the financial industry to achieve an autonomous and controllable closed loop.

Starting in 2023, mainstream large models were introduced, simultaneously building an AI innovation platform, and subsequently completing the joint research and development platform for AI application scenarios. In February 2025, Bank of Beijing partnered with Huawei to achieve full-stack domestic adaptation for DeepSeek financial applications, covering everything from underlying computing power to large models to business applications.

The underlying technology stack is also upgrading synchronously. The self-developed "Financial Operating System" connects all data and devices downward and supports agile development upward. On the "Jingzhi Brain" platform, production time was reduced by an average of 60%, and R&D delivery capability improved by 20%. The core system completed the "mainframe cloudification" migration, and the database underwent distributed transformation.

In the "Guiding Opinions," the prudent exploration of AI technology research and development and the construction of financial agents were placed in a position of considerable importance.

In this direction, from introducing an agent orchestration and creation platform in 2024 to embedding over 300 agents into Bank of Beijing's business processes by the end of 2025. Completing this scale in just over a year is quite rare among city commercial banks.

These 300+ agents are divided into three categories:

About 180 "Fast-Handle Agents" are responsible for standardized scenarios with clear processes and simple operations. Pre-loan screening, quota inquiries, and routine operations, which previously required manual effort for over ten minutes, now respond in seconds.

About 90 "Deep-Research Agents" handle complex scenarios requiring the integration of multi-source information, adopting a collaborative mode of "machines do the work, humans make the final review." They capture data in real-time and automatically identify deviations，弥补 ing omissions in human judgment.

About 30 "Cruise Agents" link market databases, competitor reports, and internal core systems, monitoring business indicators and dynamics 7x24 hours a day, achieving early perception and early warning of risks.

These three types of agents form a complete AI agent system, with intuitive and quantifiable efficiency improvements.

In the field of intelligent trading, in businesses such as pledged repurchase, AI empowers the entire chain, improving transaction processing efficiency by nearly 10 times. Supply chain ecological graphs empower relationship managers to precisely acquire customers, improving efficiency by 5 times. The mobile banking intelligent customer service completes an average of 6.26 million user consultations per month, with a success rate of 92%.

Additionally, Bank of Beijing added over 30 automated processes based on the RPA platform, deployed 20 robots, and cumulatively saved 300 person-months of manual labor.

And the indicator that truly proves AI is deeply embedded in business is large model activity. Compared to two years ago, the Token activity of large models in 2026 grew by 27 times.

03 From "+AI" to "AI+"

Stacking an AI module onto old processes is tool thinking, still stuck in the 初级 stage.

What Bank of Beijing is doing now is another thing: redesigning business processes with AI as the origin.

The difference between the two mindsets lies in that the former solves efficiency problems, while the latter cuts into the structural bottom layer.

Starting in 2024, all new projects across the bank are required to follow "AI First," meaning AI is treated as the default option from the requirement definition stage, rather than waiting until the system is built to think about "where can we add an AI."

For any new business initiation, the primary question is: Can AI be embedded, replaced, or improve the original process?

Following this logic, AI begins to permeate every link of credit approval, risk control, product design, customer service, and technology R&D.

Changes in the credit field are the most typical. AI-driven risk assessment is changing the traditional reliance on collateral.

Bank of Beijing's self-developed "Sci-Tech Radar" evaluation system does not just look at financial statements. It also profiles tech enterprises from dimensions such as technical strength and innovation potential, activating the core value of light assets of tech enterprises.

The cumulative disbursement of "Pilot AI Loan" has broken through 160 billion yuan, and "Sci-Tech e-Loan" and "R&D Loan" have allowed a large number of light-asset, high-risk tech enterprises to obtain low-cost financing for the first time.

By the end of the first quarter of 2026, Bank of Beijing's sci-tech finance loan balance broke through 480 billion yuan, serving over 30,000 specialized and sophisticated enterprises. 82% of ChiNext-listed companies, 76% of STAR Market-listed companies, and 75% of BSE-listed companies in the Beijing region are within its service scope.

It is worth noting that credit approval is explicitly listed in the "Guiding Opinions" as a category of "high-risk AI applications." For such applications, key links must establish human supervision and intervention mechanisms, clearly defining emergency stop conditions. Bank of Beijing's deep-research agents adopt a collaborative mode of "machine processing, human final review," ensuring that the final decision-making power remains in human hands.

Risk control is also changing. The new generation credit risk management system integrates the entire credit approval process, allowing AI to transform risk identification from "post-event remediation" to "pre-event warning." In 2025, Bank of Beijing's non-performing loan ratio dropped to 1.29%, declining for the fifth consecutive year. Registered clients of the open bank increased by 35% year-on-year, with 19.85 million individual customers on the mobile banking app and 7.899 million monthly active users.

Technological change ultimately falls upon people and organizations. Bank of Beijing's technology personnel have nearly tripled compared to the beginning of the "14th Five-Year Plan," and the business-technology integration training camp has cumulatively cultivated thousands of composite talents. In 2025, over 300 high-value business scenarios at the head office and branches have been implemented.

More representative is the change in management mechanisms.

Bank of Beijing has established a control mechanism similar to HR for agents, standardizing production, issuance, authorization, supervision, and human-machine collaboration. Agents are incorporated into the labor production system just like human employees. This means profound changes have occurred in the internal production relations of the bank.

Efficiency improvements driven by AI will eventually be reflected in financial data.

In the first quarter of 2026, Bank of Beijing's revenue was 19.599 billion yuan, up 14.43% year-on-year; net profit attributable to parent company shareholders was 8.098 billion yuan, up 5.55% year-on-year.

AI transformation is increasingly shaping the ship of Theseus that is the bank.

From credit approval to risk control, from customer touchpoints to product design, each plank is equipped with AI-driven capabilities one by one.

Importantly, it is sailing into a completely new body of water. In this body of water, the rules of the game are being rewritten.

Computing power iteration speed, agent collaboration density, and large model knowledge depth are replacing branch numbers and staff size, becoming the new yardstick for measuring a bank's competitiveness. Their half-lives are shorter, and update frequencies are higher. This shifts the advantage from a bank's original static stock advantages to dynamic evolutionary capabilities.

Industrial changes previously unimaginable are gradually becoming clear.

Size no longer matters; speed determines life and death. Medium-sized banks, after deep AI transformation, may surpass opponents several times their size in operational efficiency.

Under this new logic, banks that complete underlying architecture upgrades first will be closer to the definition of future banks than any existing peers.