The 8,000 Ghost Employees: How China's Banks Quietly Built an AI Workforce

*Trading screens in Shanghai's Lujiazui financial district. Behind the glow, a quiet revolution: China's banking system has absorbed AI more deeply than any other industry. (Image: Unsplash)*

At 7:00 AM on a Tuesday in Chongqing, before a single customer walks through the doors, a bank branch manager opens her phone. Overnight, an AI platform called "Chongyin Shubao" has analyzed her branch's deposits, loan pipeline, and operational metrics. By 7:15, a personalized morning report arrives — one of 26,000 such reports the system generates every month across Chongqing Bank's network. She reads it over tea. No analyst stayed late to produce it. No data team compiled the figures. The report simply exists, delivered before she knew she needed it.

This is not a pilot project. It is not a proof of concept. This is standard operating procedure at a mid-tier Chinese city commercial bank in 2026 — and it represents the tip of an iceberg that most of the world has not yet noticed.

While global attention fixates on China's AI model labs — DeepSeek, Moonshot, Zhipu — the country's banking system has been quietly executing the largest enterprise AI deployment on the planet. The scale is difficult to overstate. The Industrial and Commercial Bank of China (ICBC), the world's largest bank by assets, now consumes more than 10 billion AI tokens every single day — a nearly hundredfold increase from just two years ago. Its flagship large language model, "Gongyin Zhiyong" (ICBC Wisdom Surge), runs in over 600 business scenarios. And an AI agent dedicated to personal relationship managers handled more than 22 million customer service interactions in the first half of 2026 alone.

These numbers come not from press releases or investor decks, but from ICBC's own interim earnings call in August 2026. They are audited, operational, and real. And they tell a story that transcends any single bank: China's financial sector has crossed the threshold from AI experimentation to AI dependence.

The Trillion-Token Bank

To understand how ICBC got here, you need to understand what changed in 2026. For four years, the bank had pursued a strategy it called "Digital ICBC" — a broad digitization push that moved services online, built mobile apps, and automated back-office processes. It was impressive by traditional banking standards, but it was fundamentally about digitizing existing workflows, not rethinking them.

In early 2026, the bank quietly renamed the initiative. "Digital ICBC" became "AI-ICBC." The change was more than cosmetic. Vice President Zhao Guide put it bluntly on the August earnings call: digital intelligence is not a choice, he said. It is a mandatory question.

The new architecture, which ICBC calls its "1+1+3" system, reveals the ambition. The first "1" is the bank's upgraded technology foundation. The second "1" is a unified enterprise data vault, bringing together customer records, transaction histories, and market data into a single accessible ecosystem. And the "3" represents three dedicated AI agent platforms: one for corporate clients, one for retail customers, and one for the bank's 400,000+ employees.

That last number deserves attention. ICBC is not deploying AI for a tech-savvy elite. It is giving AI tools to every relationship manager, every branch officer, every risk analyst in an organization that employs more people than most mid-sized cities have residents. The bank has also trained over 12,000 data analysts and continues to expand its technology team, which numbered 36,000 as early as 2023.

The operational metrics are staggering. The "Gongyin Zhiyong" large model — which ICBC built in-house — has been deployed across more than 600 scenarios spanning every major business line. In investment and trading, an intelligent pricing assistant has pushed the automation rate of transactions to 96%, while transaction volume has surged 50% year-over-year. In credit risk, an intelligent lending assistant now supports loan officers through the entire origination process, from initial assessment to documentation. The bank's information systems maintain an availability rate above 99.99% even under this AI load.

*Sources: ICBC H1 2026 Earnings Call (August 28, 2026), China Daily, Baidu Baike*

ICBC's H1 2026 financial results provide context for why this matters. The bank extended ¥3 trillion in loans to technology companies — a figure that reflects its role as the financial engine of China's AI buildout. Its digital economy lending reached ¥1.26 trillion, up nearly 20%. Cross-border RMB business volume hit ¥5.5 trillion in a single half. The bank's non-performing loan ratio stood at a manageable 1.29%, with provision coverage of 217.58%. In other words, the AI transformation is happening inside a bank that is simultaneously growing, lending more, and maintaining credit discipline.

But ICBC is only the most visible example of a system-wide shift.

8,000 Ghost Employees

At China Merchants Bank (CMB), headquartered across the country in Shenzhen, the transformation has taken a different form — one that CIO Zhou Tianhong described with a number that stopped industry observers cold.

By the end of 2025, CMB had deployed 856 large-model application scenarios across its operations. Those AI systems replaced more than 15.56 million hours of manual work in a single year. Zhou did the math for his audience: that is equivalent to the output of over 8,000 full-time employees.

Eight thousand. At a bank that employs roughly 120,000 people, AI systems now perform work equivalent to roughly 7% of the entire human workforce. And unlike human workers, these systems operate around the clock, do not take vacations, do not file grievances, and improve continuously as the underlying models advance.

The comparison is not entirely fair — AI systems do not exercise judgment the way experienced bankers do, and CMB has been careful to frame this as augmentation rather than replacement. But the efficiency gains are concrete and measurable. An AI assistant deployed for account managers has driven a 20% increase in the average transaction volume per customer — a metric that directly translates to revenue. The bank's digital employees handle routine operations across retail banking, wealth management, and customer service.

*AI-powered dashboards now deliver operational intelligence to Chinese bank managers before they arrive at work. CMB's AI systems replaced the equivalent of 8,000+ full-time employees' annual output. (Image: Unsplash)*

What makes CMB's deployment particularly instructive is its approach to risk. Zhou has been publicly candid about the dangers of large language model hallucinations in financial contexts. A bank cannot afford an AI that invents a credit rating, misstates a regulatory requirement, or advises a customer incorrectly on an investment product. CMB has implemented multiple layers of verification and human oversight, and the bank continues to increase its AI safety investment in 2026.

This tension — between the enormous efficiency gains AI offers and the existential risks of deploying probabilistic systems in a domain that demands precision — runs through every major Chinese bank's AI strategy. It is not a solved problem. But the direction of travel is clear: more deployment, more automation, more trust in the systems.

The City Banks Join the Race

The transformation extends far beyond the two giants. China's banking system comprises over 4,500 institutions, and the AI wave is reaching every tier.

Agricultural Bank of China (ABC), the country's third-largest lender, is deploying AI-powered recognition systems for risk management and using big data modeling to sharpen early-warning systems for loan defaults. The bank is advancing "AI plus wealth management" applications that build integrated service models spanning customer insights, intelligent asset allocation, and post-investment support. ABC President Wang Zhiheng has framed this explicitly within the bank's strategic mandate: turning technological innovation into the greatest driver of high-quality development.

At Chongqing Bank, the "Chongyin Shubao" platform exemplifies what AI looks like at the city commercial bank level. In the first half of 2026, it pushed 26,000 operational morning reports to branch managers and provided over 2,300 intelligent data analysis services. These are not exotic applications — they are the bread-and-butter of daily bank management, now automated.

Nanjing Bank has gone further in operational automation, deploying digital employees that handle subscription and redemption processing, dividend distribution, and other standardized operations for more than 150 fund companies. A task that once required armies of back-office staff is now performed by AI agents that work continuously, make no transcription errors, and flag anomalies for human review.

The pattern is consistent: Chinese banks are not using AI to build flashy consumer-facing features. They are deploying it in the operational core — risk management, transaction processing, reporting — where the volume is highest and the cost savings are most dramatic.

*Sources: Company disclosures, The Asian Banker, Caixin, China Daily*

The WeBank Playbook

If the state-owned giants represent AI deployment at maximum scale, WeBank represents something perhaps more radical: a bank built AI-native from the ground up.

WeBank — founded in 2014 as China's first digital-only bank, backed by Tencent — has never operated physical branches. Its entire customer interface has always been digital. But in 2026, the distinction between "digital bank" and "AI bank" has become meaningful in a new way. The Asian Banker named WeBank the Best AI-Driven Bank of the Year in Asia Pacific, citing its development of an "AI-native operating model" that embeds artificial intelligence across product design, operational management, and customer interaction.

The difference matters. Traditional banks, even those as advanced as ICBC and CMB, are layering AI onto existing organizational structures, existing workflows, and existing business logic. WeBank is doing something different: redesigning its operating model around what AI makes possible, rather than bolting AI onto what already exists. Its loan underwriting, fraud detection, customer service, and product recommendation systems were designed with machine intelligence as the primary actor and human judgment as the exception handler.

This distinction — retrofit versus native — may prove to be the most important strategic divide in the industry over the next five years. Retrofitted AI delivers incremental efficiency. Native AI enables fundamentally different economics.

The Numbers Behind the Transformation

Step back, and the aggregate picture is remarkable. China's fintech market — the broad ecosystem of technology-enabled financial services — was valued at $59.39 billion in 2026 and is projected to reach $123.78 billion by 2031, representing a compound annual growth rate of 15.82%. That growth is being powered by AI adoption that, according to The Asian Banker's 2026 research, now features in more than 60% of AI-related submissions from Chinese financial institutions — a figure that reflects the industry's shift from experimentation to operational deployment.

The academic evidence supports the business case. A 2026 study using a multi-period difference-in-differences model across 17 Chinese A-share listed commercial banks found that AI adoption significantly improves bank profitability through enhanced operational efficiency, reduced information asymmetry, and improved risk assessment.

*Sources: Mordor Intelligence, The Asian Banker 2026, MDPI Journal of Risk and Financial Management, QuestMobile*

The consumer dimension reinforces the picture. QuestMobile reported that as of July 2026, China's AI-powered assistant products collectively reached 1.13 billion users, up 47.3% year-over-year. Total usage frequency surged 162.8%. Chinese consumers are not merely aware of AI — they are integrating it into daily financial decisions at a pace that dwarfs most Western markets.

Why Global Banking Should Pay Attention

For international observers, the Chinese banking AI story carries implications that extend well beyond China's borders.

First, there is the question of competitive template. Western banks have largely approached AI through the lens of cost reduction — automating call centers, streamlining compliance, reducing headcount in back-office functions. Chinese banks are doing those things too, but they are simultaneously using AI to expand their addressable market. ICBC's ¥3 trillion in technology company loans, its ¥1.26 trillion digital economy lending portfolio, and its AI-powered service delivery to hundreds of millions of customers represent growth, not just efficiency. The Chinese model suggests that AI in banking is not primarily a cost story — it is a growth story that happens to also reduce costs.

Second, there is the question of regulatory posture. China's banking regulator has taken a comparatively permissive approach to AI deployment in financial services, emphasizing innovation and sandbox-style experimentation while maintaining baseline requirements around data security, model transparency, and consumer protection. The revised Cybersecurity Law, effective January 2026, added AI-specific provisions around risk assessment and human controllability. But the overall trajectory is toward enabling deployment, not constraining it. This stands in contrast to the EU's AI Act, which imposes more stringent requirements on "high-risk" AI applications — a category that includes most banking use cases.

Third, and perhaps most consequentially, there is the question of what happens when AI-native banking meets AI-native everything else. If Chinese banks are building AI agents that can autonomously manage corporate treasuries, optimize supply chain financing, and execute cross-border settlements, and Chinese e-commerce platforms are building AI agents that can autonomously source products, negotiate prices, and manage inventory, the intersection of these capabilities points toward a financial system where an enormous share of routine transactions are executed machine-to-machine, with humans setting strategy and parameters rather than approving individual transactions.

The Uncomfortable Questions

None of this is without risk. The CMB CIO's public acknowledgment of hallucination risk is not corporate humility — it is a genuine operational concern. Large language models are probabilistic systems. Banking is a domain of legal precision. The gap between those two realities is not fully bridgeable with current technology, and every Chinese bank deploying AI at scale is grappling with the same fundamental question: how much autonomy is too much?

The employment question is equally thorny. CMB's 8,000 FTE-equivalent in replaced work does not mean 8,000 people lost their jobs — the bank has been careful to frame AI as augmenting human workers, not replacing them. But the trajectory is clear. If AI systems continue improving at their current pace, and if Chinese banks continue expanding their AI deployments at the current rate, the total volume of human work displaced by machines in China's banking sector will grow from millions of hours per year to tens of millions.

There is also a competitive concentration risk. The banks that can afford to build trillion-token AI infrastructures will pull further ahead of smaller institutions. The gap between AI-haves and AI-have-nots in Chinese banking could become a self-reinforcing cycle: better AI leads to better customer experiences, which attracts more deposits, which funds more AI investment.

And finally, there is the systemic risk dimension. China's banking system manages the world's largest pool of household savings. If AI systems managing trillions of yuan in assets develop correlated failure modes — the same hallucination, the same data quality issue, the same model vulnerability propagated across hundreds of institutions using similar underlying technology — the potential for cascading disruption is real. Chinese regulators are aware of this risk, and the July 2026 "Finance China" conference in Beijing explicitly elevated governance alongside innovation as dual priorities. But awareness and mitigation are not the same thing.

What Comes Next

Looking ahead, the trajectory is defined by three converging forces.

The first is policy. The 15th Five-Year Plan (2026-2030) explicitly positions AI as a driver of financial sector modernization. ICBC's own plan, embedded within this national framework, calls for accelerating intelligent transformation, developing enterprise-level data infrastructure, and innovating financial AI agents to "explore new paradigms for AI-empowered finance." This is not a technology strategy — it is industrial policy applied to banking.

The second is technology. As Chinese AI models continue improving — and as the cost of inference continues falling — the capabilities available to banks will expand dramatically. Today's AI agents handle customer service interactions and generate morning reports. Tomorrow's will underwrite complex commercial loans, negotiate derivative contracts, and manage real-time treasury operations across multiple currencies and jurisdictions.

The third is competition. When one major bank achieves a 20% increase in per-customer transaction volume through AI, every competitor must respond. When one bank deploys 856 LLM scenarios, every peer institution faces pressure to match or exceed that number. The dynamic is self-accelerating: each deployment raises the competitive bar, forcing the next round of investment.

By 2030, the Chinese banking system will likely look fundamentally different from today's. Human bankers will focus on relationship management, complex judgment calls, and strategic decisions. AI agents will handle the vast majority of operational tasks — processing, reporting, routine underwriting, standard customer interactions. The banks that thrive will be those that manage this transition most effectively: deploying AI aggressively enough to capture the efficiency gains, but carefully enough to maintain the trust and precision that banking demands.

The 8,000 ghost employees at CMB are not a curiosity. They are a preview.

Social Media Voices

From Zhihu:

"工行一天消耗的token比很多小国家全年用量都大，这才是真正的'新质生产力'。我们总在讨论AI实验室，但真正的落地是在银行。"

>

*"ICBC consumes more tokens in a day than many small countries use in a year. This is what 'new quality productive forces' actually look like. We keep talking about AI labs, but the real deployment is happening in banks."*

>

— 知乎用户 "金融科技观察者", 2.4K upvotes

From Twitter/X:

"China Merchants Bank replaced 8,000 FTE-equivalent work with LLMs in one year. Western banks are still running chatbot pilots. The gap in enterprise AI deployment between China and everyone else is not closing — it's widening."

>

— @fintech_watcher, 1.8K likes

From Xiaohongshu:

"在银行工作的表姐说，她们现在早上不用做报表了，AI自动生成。她说既开心又害怕——开心的是不用加班，害怕的是不知道自己还能做什么。"

>

*"My cousin who works at a bank says they don't have to prepare morning reports anymore — the AI generates them automatically. She says she feels both relieved and scared — relieved she doesn't have to work overtime, scared about what she'll do next."*

>

— 小红书用户 "城市观察日记", 3.1K likes

From Weibo:

"银行都在用AI了，但幻觉问题怎么解决？万一AI给错了一个数字，那就是真金白银的损失。招商银行CIO能公开承认这个问题，至少态度是诚实的。"

>

*"Banks are all using AI now, but how do you solve the hallucination problem? If an AI gets one number wrong, that's real money lost. CMB's CIO publicly admitting this issue — at least the attitude is honest."*

>

— 微博用户 "风控老兵", 1.2K reposts

From GitHub:

"Interesting that ICBC built their own LLM instead of using DeepSeek or Qwen. A bank with 400K employees rolling their own foundation model shows how far enterprise AI has come — and how much internal capability they're building."

>

— @quant_dev_cn, discussion thread

From Douban:

"每次看这种新闻都很矛盾。一方面觉得中国AI应用确实厉害，另一方面担心基层员工怎么办。八万人的工作量被替代，背后是多少家庭？"

>

*"Every time I read news like this I feel conflicted. On one hand, China's AI applications are genuinely impressive. On the other hand, what happens to frontline employees? Eight thousand people's worth of work replaced — how many families are behind that number?"*

>

— 豆瓣用户 "阿尔法城的难民", 876 responses

*Related reading: The Agentic Cloud War: Huawei and Alibaba's Battle for China's AI Agent Future | DeepSeek's Silence and the V4 Pro Question | Moonshot AI's $50 Billion Sovereign AI Capital Blueprint | China's AI Agent Army: Inside the OpenClaw Workforce Revolution*