Chinese Banks Deploy AI Agents at Scale: 2026 Half-Year Report Analysis

Overall Landscape: From LLM Pilots to Batch Agent Deployment

By late August 2026, all 42 listed Chinese banks had disclosed their half-year reports. Unlike previous years where AI was mentioned only under generic "technology investment", this cycle AI became a systematic disclosure focus. Quantitative metrics — agent counts, token consumption, and equivalent man-hours replaced — appeared densely. Token consumption was written into financial report narratives for the first time, becoming a new industry yardstick for AI application scale. The competition has moved from "whether they have AI" to a phase emphasizing both scale and control.

If 2025 was the year of large-model deployment, the hallmark of H1 2026 is that intelligent agents have replaced the models themselves as the core carrier for production use. Banks universally upgraded from "can chat" Q&A assistants to "can execute" digital employees. AI has risen from a technology-department tool to a whole-bank strategic priority, with several banks establishing AI task forces led by the chairman or president.

The industry shows a clear tiered structure: the six large state-owned banks lead in agent quantity and capability building; joint-stock banks pursue systematic layout around key scenarios; city and rural commercial banks focus on characteristic scenarios and digital employees via lightweight paths.

State-Owned Big Six: Platformization, Data Governance, Organizational Restructuring

Common traits across the six banks:

Platformization: unified model and compute infrastructure.

Data governance intensified: examples include Postal Savings Bank's "Data for AI", ICBC's enterprise-level trusted data service platform, and ABC's data architecture adapted to the full agent lifecycle. The competitive focus is shifting from "models" to "data + governance".

Organizational restructuring: operating model shifting to "agents handle repetitive work, humans handle exceptions and customer communication".

Joint-Stock Banks: Systematic Key Scenarios, Dense Quantitative Disclosure

China Merchants Bank: Continued strengthening of "cloud + AI + middle platform" intelligent foundation, building a full-stack large-model technical system covering compute, models, and applications. Four-pronged synergy: technical system, knowledge management, scenario application, ecosystem building.

Technical system: daily token throughput up >78% YoY; 256 domain models deployed (+40% vs end-2025).

Knowledge management: deepened enterprise-level knowledge management and application system.

Scenario application: 1,386 intelligent scenarios landed (+62% vs end-2025) across retail, wholesale, risk control, operations, office. AI contributed 13.88 million equivalent man-hours in efficiency gains.

Ecosystem building: acted as founder, maintainer, core contributor to mainstream open-source projects; contributed 200+ core technical achievements to global AI open-source ecosystem.

Ping An Bank: H1 2026 launched 140 operations-audit agents covering 55 business scenarios, boosting efficiency in anti-money-laundering due diligence, account review, fund clearing, cross-border remittance. By June-end, 35 consumer-protection review point models deployed, intelligent review coverage >90%.

Accelerated large-model capability system: 450+ large-model application scenarios landed by June 2026, covering precision marketing, intelligent advisory, compliance risk control, office assistance.

Perfected compute platform: unified scheduling and efficient management of diverse compute resources; H1 daily token consumption >5.3 billion, +130% YoY.

Advanced high-performance inference and multimodal model iteration: mainstream large-model ecosystem products deployed at scale, deployment cycle shortened to one week.

CITIC Bank: Core strategy "scenario-driven + AI-enabled", fully implementing "AI+" action. Pushed AI construction from "point exploration" to "systematic building, scaled application"; cumulative AI service scenarios >1,900. Built group-level agent management platform, focusing on investment research, intelligent trading, customer service.

Industrial Bank: Elevated digital operations level. With strategy system as core and intelligence as engine, systematically strengthened digital operations. Based on centralized head-office digital operations capability and digital marketing platform, built strategy system and branded rights-activity system, deepened channel synergy across mobile banking, enterprise WeChat, remote operations. H1 deployed 21,800 strategies, touched 597 million customer-times, drove comprehensive financial assets of 178.459 billion yuan. Strategy intelligent audit covers 52 sub-agent modules, manual verification consistency 99.96%; strategy intelligent evaluation accuracy 100%.

China Zheshang Bank: Adjusted "technology-driven bank" positioning from "maintain peer-relative leadership" to "local relative leadership". AI narrative shifted from unified assistant/capability hub to enterprise-level agent platform and first batch of key business scenarios. Actively implementing national "AI+" strategy, improving compute clusters, enterprise-level agent platform, confirming first batch of key AI construction scenarios, piloting retail AI assistant, corporate AI assistant, intelligent credit risk control, advancing intelligent R&D. Regular "AI+Finance" special training to accelerate full-staff AI literacy.

Regional and Private Banks: Lightweight, Characteristic Scenarios, Calculating "Small Accounts"

Bank of Jiangsu: Launched "AI+Finance" strategic project, fully implementing financial process intelligent engineering, reshaping traditional financial processes with AI tools. H1 2026 daily token usage up 18x vs 2025; AI applications cumulatively replaced 1.2 million man-hours; efficiently replicated private-banking marketing service capability; comprehensively improved risk control precision and timeliness.

Bank of Nanjing: Deepened AI "Double Hundred" plan; H1 added 30 high-value business agents, cumulative 156 agents built, covering marketing, operations, risk control, office.

Bank of Chongqing: Based on "456" digital transformation framework, focused on "AI+data" empowerment and risk prevention. Built generative AI application management platform, deployed Qwen large model, added 7 new scenarios including "intelligent compliance assistant". In May and July 2026 issued procurement for 34.79 million yuan "high-performance compute expansion" and 4 million yuan "AI large-model application scenario R&D", aiming to move AI from single-point trials to systematic landing.

WeBank: As of July 2026, daily token consumption >20 billion (up from 5 billion at end-2025, multi-fold growth); formed 80+ digital employees "onboarded", 7×24 human-machine collaboration mode. In R&D, digital employees handle security ticket reception, requirement breakdown, code editing, automated testing to release end-to-end, compressing original ~2-week cycle to under 30 minutes, efficiency improvement >10x.

Five Trend Judgments

Agents upgrade from "Q&A assist" to "digital employees": Universal leap from "can chat" to "can execute". ICBC digital employees undertook 55,000 man-years workload in 2025; CMB "equip every line with an assistant"; PSBC transaction robot "Youxiaozhu" processed >12 trillion yuan transactions.

AI-native and Agentic AI become new narrative: All six big banks' half-year reports mention "AI-native". Industry defines Agentic AI as tool evolution stage 3.0, driving shift from "human executes, AI assists" to "human supervises, AI executes".

Governance and compliance define the track: Regulatory documents such as "Implementation Opinions on Standardized Application and Innovative Development of Intelligent Agents" and "Guidance on Secure Development and Application of AI in Banking and Insurance" have landed, adding a "competing on governance" dimension.

Cost anxiety behind the token race: Management rhetoric shifted from "competing on token scale" to "competing on token efficiency". CMB disclosed 55% token consumption reduction via planning optimization; refined operations replacing extensive expansion.

Organizational restructuring begins: AI elevated from tech line to whole-bank strategy; human-machine collaboration operating model entering institutionalization at multiple banks.

Sobering Realities: Three Issues Not to Ignore

Still in transition period: Industry at "scenario verification toward scaled replication" stage. Agents concentrated in low-to-medium risk links: internal assistants, customer service assistance, credit material processing. Autonomous execution of core financial transactions rare; high-risk links like credit approval still human-final.

Effectiveness evaluation remains a bottleneck: Marketing gains hard to separate from organic growth; risk control gains appear as "losses not occurred" not hitting P&L man-hours easy to count but decision quality, customer retention hard to monetize; disclosure calibers differ across banks, cross-bank comparability questionable.

Short-term ROI pressure is real: A joint-stock bank president stated AI "may not directly reflect in financial statement gains, even phase-wise showing as expenditure and cost" — industry needs longer horizon to view AI input-output.

2026 half-year reports give a clear answer: bank AI competition has moved from "have or not" to "scale + control"; agent count, token consumption, equivalent man-hours are the new hard currency. But the true watershed is not how many agents are launched, but whether end-to-end operational efficiency can be reconstructed under safety and compliance. From this cycle, no bank has crossed that line — H1 compared speed, H2 will compare governance, efficiency, and patience.

Note: Data sourced from commercial banks' 2026 half-year reports and public disclosures, industry regulatory documents.