China – AI 2026/27

An analysis of AI in Chinese financial services, examining the shift from FinTech to financial foundation models and AI agents, their use across banking, securities, asset management, insurance and payments, and China’s evolving regulatory and compliance framework.

Tags: 2026/27 / AI / China / King & Wood / Yirui Zhang (Alicia)

Yirui Zhang (Alicia)
Partner

Alicia is a partner in the Technology, Media, Entertainment and Cross-border Intellectual Property (TMEC) practice group at King & Wood. She specializes in intellectual property, technology commercialization, and strategic management of data and digital assets, with particular focus on artificial intelligence, fintech, and robotics.

Alicia advises leading enterprises on digital transformation, data transactions, AI compliance, algorithmic governance, and technology ethics frameworks. Her practice spans the application and regulation of emerging technologies at the intersection of finance and innovation. She serves as Director of the Digital Technology and AI Committee of the Shanghai Bar Association and holds advisory roles with the National Facility for Translational Medicine and Shanghai Data Service Provider Association.

Recognized as one of Shanghai’s Outstanding Female Lawyers and ranked in IAM Patent 1000, Alicia holds an LL.B. and B.S. in Computer Science from East China University of Political Science and Law and an LL.M. from the University of Illinois Urbana-Champaign. She is admitted to practice in the PRC and New York, and works in Chinese and English.

AI IN CHINESE FINANCIAL SERVICES: FROM FINTECH TO FOUNDATION MODELS — MARKET PRACTICE, REGULATORY ARCHITECTURE AND THE COMPLIANCE ROADMAP

From FinTech to financial foundation models and AI agents

The application of artificial intelligence (“AI”) in Chinese financial services (“FS”) is best understood as the latest stage of a decades-long digitalisation trajectory. The first stage was financial electronification — the automation of bookkeeping, clearing and branch operations from the 1990s onwards. The second was internet finance, which from around 2013 brought mobile payments, online wealth management and platform-based lending to mass-market scale. The third, financial digitalisation, saw institutions rebuild core systems around data, cloud and APIs. Since around 2023, the sector has entered a fourth stage: the era of financial foundation models and AI agents, in which large language models (“LLMs”) are no longer peripheral productivity tools but are being embedded into credit, trading, advisory, underwriting and claims workflows.

This transition enjoys explicit top-level policy endorsement. The General Office of the State Council’s Guiding Opinions on Advancing the “Five Major Articles” of Finance (Guo Ban Fa [2025] No. 8, March 2025)¹ elevated digital finance — alongside technology finance, green finance, inclusive finance and pension finance — to the status of national strategy, requiring financial institutions to accelerate digital transformation and calling for a sound digital finance governance system in which innovative businesses are brought within the regulatory perimeter. The National Financial Regulatory Administration (“NFRA”) had earlier issued its own implementation opinions on the “Five Major Articles” (Jin Fa [2024] No. 11)2, calling on banking and insurance institutions to advance digital transformation and encouraging technologically leading institutions to export risk-control tools and technology services to smaller institutions. Within this framework, ² AI is positioned not merely as an efficiency tool but as a component of the financial system’s infrastructure.

Market landscape and infrastructure characteristics

Across leading Chinese financial institutions, deployment has converged on a four-layer stack — “foundation model + knowledge base + AI agent + business systems” — that mirrors the global enterprise norm for generative AI. What distinguishes China is not the architecture itself but the choices made within it. Rather than training general-purpose models from scratch, institutions typically deploy domestic open-weight or commercial foundation models (including DeepSeek, Qwen, GLM, Baichuan and Wenxin) on private infrastructure — frequently combining domestic accelerators with still-substantial imported GPU capacity; enhance them with proprietary financial data, knowledge graphs and retrieval-augmented generation (“RAG”); orchestrate task execution through agent frameworks; and wire the outputs into core banking, trading or claims systems. State-owned banks and the larger joint-stock banks lead adoption, while city and rural commercial banks, working with tighter budgets and scarcer talent, have pursued divergent paths. Many begin with narrow, single-scenario pilots, while some have scaled to broader deployment—for instance, Jiangsu Jiangnan Rural Commercial Bank now covers more than 30 scenarios. Regulatory guidance encourages larger institutions to export computing power and models to smaller peers, though concrete cases of such exports remain scarce in public reporting.

“Chinese deployment has converged on the global enterprise stack — foundation model + knowledge base + AI agent + business systems. What sets China apart is the configuration, not the architecture: domestic foundation models, private on-premises deployment, and mandatory human review in consequential decisions.”

Three structural features distinguish the Chinese market. First, private, on-premises deployment is the dominant norm — driven above all by data-localisation and financial data-security compliance requirements rather than by unconstrained technical preference; regulators require critical data to remain onshore and expressly encourage private deployment. Second, the policy orientation towards autonomous and controllable technology stacks is unambiguous: adoption of domestic chips and operating systems is advancing under regulatory pressure, but progress is uneven across institutions, core computing power still relies substantially on imported GPUs, and self-controllability remains a work in progress rather than an accomplished fact. Third, human review is systematically retained in consequential customer-facing decisions — initially a prudent response to model-hallucination risk, as Section 5 explains, has now been codified as a regulatory mandate.

Use Cases & Institutional Practices

Banking: from intelligent customer service to intelligent credit

Banking applications have moved well beyond chatbots. In credit, AI now supports automated underwriting for micro and consumer loans — with human review retained at defined checkpoints — while intelligent customer service handles the bulk of routine interactions.

Industrial and Commercial Bank of China (“ICBC”) offers the flagship example. Its enterprise-level, hundred-billion-parameter financial foundation model system, “ICBC Zhiyong” (工银智涌), had empowered more than 20 major business lines and over 200 scenarios, with cumulative calls exceeding one billion, according to its 2024 annual report; ³ its 2025 interim report disclosed further deployment progress, including AI wealth assistants and investment research agents. ⁴ China CITIC Bank has built a three-in-one AI system combining its decision-making AI “CITIC Brain (中信大脑)” — deployed across a large number of business scenarios including credit decisioning, wealth management, risk compliance and operations ⁵ — with its generative “Cangjie (仓颉)” LLM, supporting an additional package of applications across wealth management, customer service, marketing, advisory and risk compliance. China Merchants Bank deploys the “Zhao Xiao Cai (招小财)” AI assistant in transaction banking — and has reported similar response accuracy on routine transaction-banking queries — and is reported to be applying generative-LLM capability to capital management and product pricing.

Securities and asset management: research, advisory and AI-native apps

In the securities sector, adoption has shifted from embedded tools to AI-native applications. Huatai Securities launched “AI Zhangle” (AI涨乐) — what the company describes as the industry’s first AI-native trading application — in October 2025 ⁶: rather than layering AI functions onto a conventional app, it rebuilds the service logic around AI, embedding large models into stock selection, analysis, trading and alerting through multiple AI investment-assistant personas capable of executing complex operations via dialogue. Guotai Haitong Securities, pursuing an “All in AI” strategy, built the “Junhong Lingxi (君弘灵犀)” hundred-billion-parameter multimodal securities model ⁷ and, in July 2025, launched the securities industry’s first all-AI application, “Lingxi (灵犀)” — with version 2.0 following in December 2025. In March 2026 the Junhong Lingxi model was reported to have received the top tier in the China Academy of Information and Communications Technology’s (“CAICT”) assessment of LLM capability in typical financial scenarios. ⁸

“Chinese brokerages have moved from embedding AI tools into existing apps to launching AI-native applications — a shift from ‘functions looking for users’ to ‘AI delivering services’.”

These deployments operate within the existing perimeter of investor suitability, advisory licensing and distribution rules — the use of an LLM interface does not exempt an institution from suitability management, record-keeping or content-review obligations, which apply on a technology-neutral basis.

Insurance and payments: underwriting, claims and anti-fraud

Insurance shows perhaps the deepest operational integration. Under Ping An Group’s “AI in All” strategy, the group has reported an internal portfolio of 67 vertical foundation models (some at or above the hundred-billion-parameter scale) and over 23,000 deployed AI-agent applications, with company-disclosed adoption metrics across its approximately 230,000-strong workforce and more than 650 operational scenarios in the first half of 2025. ⁹ Its 2025 annual results disclose that AI agents handled 1.702 billion customer service interactions, covering 80% of total service volume ¹⁰; 93% of policies issued through the auto dealer channel are now underwritten intelligently within one minute ¹¹; and AI-powered anti-fraud systems intercepted fraudulent claims, reducing losses by RMB 10.51 billion in 2025 — the third consecutive year above RMB 10 billion ¹². China Pacific Insurance (CPIC) P&C developed “Lingxi (灵析)” — what the company describes as the industry’s first claims AI employee integrating workflow assistance, risk alerting and quality management across the full claims chain ¹³: it extracts photo and document information and generates assessment reports in approximately one minute, with material reductions in manual data-entry volume and gains in pilot-team productivity as reported by the company.

In payments, AI underpins real-time fraud interception, merchant risk scoring and customer service at massive scale, building on the infrastructure of the mobile payment era.

Value Creation vs. Legal & Compliance Risks

Core value

The value case is consistent across segments: efficiency in unstructured information processing (research reports, disclosures, claims documents); stronger risk early-warning and anti-fraud capability; lower operating costs (ICBC’s digital workforce alone is reported to approximate tens of thousands of person-years of capacity); and financial inclusion, as automated credit scoring extends lending to small businesses and underbanked consumers.

Core legal and compliance risks

Data security and personal information protection. Financial data is heavily regulated: sensitive personal information (expressly including financial accounts) requires separate consent; the “minimum necessary” principle, together with the classification and grading requirements for important data under the Data Security Law, constrains training-data assembly; and cross-border data transfer (“CBDT”) restrictions shape — and often rule out — architectures involving offshore models or group-level platforms.

Algorithm governance and model defects. The opacity of complex models (“black box”) collides with explainability expectations; LLM hallucination creates mis-selling and misinformation risks in advisory contexts; and algorithmic discrimination in credit and pricing engages both the algorithm rules ¹⁴ and PIPL Section 2435 on automated decision-making.

Intellectual property and data provenance. The legality of training corpora, licensing risks in third-party API arrangements, and the contested copyright status of AI-generated content17 all remain live issues in Chinese law.

Novel cyber and systemic risks. Prompt injection, model poisoning and supply-chain compromise feature expressly in the latest regulatory guidance34; at the market level, homogeneous models raise concerns about convergent strategies, herding and the amplification of volatility — a systemic-risk framing now explicitly addressed by the financial regulator21 (See Section 4 China’s AI Regulatory & Legal Framework).

China’s AI Regulatory & Legal Framework

China has no single comprehensive AI statute. Instead, obligations arise from four layers of rules, which stack upon one another. For FS institutions the practical question is never whether a single rule applies, but how the layers interact: a customer-facing chatbot at a bank may simultaneously be a “generative AI service” under the Cyberspace Administration of China (“CAC”) rules16, a “high-risk application” under the NFRA’s June 2026 guidance21, an “automated decision” under the Personal Information Protection Law (“PIPL”)35, and a regulated advisory activity under securities law. This section maps each layer and — critically — how each connects to financial regulation and practice.

Layer One: horizontal AI and algorithm governance (led by the CAC)

Administered principally by the CAC, this layer targets AI techniques directly. Four instruments form the core, and each has a distinct financial-services interface:

Algorithm Recommendation Provisions (effective 1 March 2022)14 — China’s first algorithm-specific regulation requires service providers with “public opinion attributes or social mobilisation capacity” to complete an algorithm filing, grants users the right to opt out of personalised recommendation, and prohibits unreasonable differential treatment in trading conditions. For FS institutions, this reaches personalised product recommendation in banking and wealth-management apps, robo-advisory distribution logic, and “big-data-enabled price discrimination” in loan pricing and insurance quotation — practices that overlap with, and are reinforced by, PIPL Section 24 and financial consumer-protection rules.

Deep Synthesis Provisions (effective 10 January 2023) ¹⁵ — impose labelling, real-name verification, biometric consent and security assessment duties on providers of deep synthesis technology (synthetic media, digital humans, voice synthesis). These rules cut both ways for finance: they govern institutions’ own deployments — digital human tellers and livestream anchors, synthetic voice in telemarketing and collections — and they supply part of the legal toolkit against deepfake-enabled fraud targeting facial-recognition onboarding and voice authentication.

Interim Measures for Generative AI Services (effective 15 August 2023) ¹⁶ — the centrepiece. The Measures regulate GenAI services offered to the public; research, development and use that does not offer services to the domestic public are expressly outside their scope. Providers of public-facing services must ensure the legality of training data, protect IP and personal information, prevent discrimination, maintain content integrity, handle complaints, and — where the service has public opinion attributes — undergo security assessment and algorithm filing. The internal-use carve-out is the pivot of every FS deployment analysis: an internal coding assistant or document-summarisation tool generally sits outside the regime, whereas a customer-facing chatbot, AI research generator or marketing-copy engine falls squarely within it — and institutions are well advised to treat borderline “employee-facing but customer-impacting” deployments conservatively.

AI-Generated and Synthesized Content Labelling Measures (effective 1 September 2025) ¹⁷ — require dual explicit and implicit labelling of AI-generated content (visible markers plus embedded metadata), prohibit tampering with labels, and oblige dissemination platforms and app distribution channels to verify labelling. A companion mandatory national standard, GB 45438-2025 (Cybersecurity Technology — Labelling Method for AI-Generated Synthetic Content), supplies the technical specifications. In finance, this bears directly on AI-generated marketing materials, research summaries, digital-human video and customer communications — and creates a review checkpoint for compliance and investor-education content pipelines.

The filing regime has scaled rapidly: across CAC monthly publications the population of filed GenAI services grew from 346 in March 2025 toward 868 in April 2026, with a parallel group of registered AI applications and functions ¹⁸; mid-2026 the same filing regime was extended to on-device large models, with the first cohort of smartphone-side generative AI services — including Apple Intelligence — recorded as filed ¹⁹. The threshold question for an FS institution is therefore factual and structural: is the deployment public-facing, does it rely on an externally sourced model (and is that model filed?), and which entity in the group is the “provider”? As this Section shows, the financial regulator has now imported these thresholds directly into sectoral supervision33.

Layer Two: financial-sector rules (NFRA, PBOC, CSRC)

The policy trajectory. Sectoral AI governance did not appear overnight. The PBOC’s FinTech Development Plan (2022–2025) first called for algorithm management rules, model security assessment and audit systems, disclosure of algorithmic decision logic39, and stress-testing against black-box, herding and discrimination risks. The “Five Major Articles” framework then made digital finance a strategic priority, and the NFRA’s December 2025 implementation plan for high-quality digital finance in banking and insurance committed the regulator to building a classified and graded management framework for secure AI application, with human-intervention mechanisms covering important business processes and key nodes. ²⁰ That commitment crystallised six months later.

The NFRA’s AI Guiding Opinions (June 2026) — the sectoral centrepiece. ²¹ On 18 June 2026, the NFRA issued the Guiding Opinions on the Secure Development and Application of Artificial Intelligence in Banking and Insurance (Jin Fa [2026] No. 8, the “AI Guiding Opinions”), setting out 32 guidance items across governance, development and application, data governance, computing power, risk management, capability building and supervision. The AI Guiding Opinions convert previously high-level principles into concrete, examinable expectations:

Governance: the board must designate a dedicated committee responsible for AI; institutions must establish full-lifecycle management systems covering demand analysis, data preparation, development, deployment, iteration and decommissioning22;

Classified and graded risk management: AI applications must be inventoried and tiered by business criticality, scale, customer impact and model complexity and dependence — operationalising the December 2025 commitment NFRA – AI Guiding Opinions, sec. 15. See: https://www.nfra.gov.cn/cn/view/pages/ItemDetail.html?docId=1261784&itemId=928 ;

High-risk application admission: GenAI applications involving fund transactions, asset valuation, credit approval, underwriting and claims, or risk management — or otherwise directly affecting customers’ interests or the conclusion of financial contracts — are deemed high-risk and may be implemented only with the approval of the institution’s risk management committee24;

Human oversight: human supervision and intervention mechanisms must be established at key nodes of high-risk applications, with defined emergency shutdown and model-exit conditions and backup manual processes ²²; where a model’s explainability is insufficient, it may be used in high-risk scenarios only as an auxiliary tool, with humans making the final decision; decisions affecting customer rights or carrying material financial impact require manual review nodes;

Model admission and external models: GenAI models are subject to admission management assessing efficacy, security and compliance; externally sourced GenAI models must have completed the CAC filing described in Section 4 — a direct interlock between Layer One and Layer Two that makes the CAC filing status of a vendor’s model a gating due-diligence item in every FS procurement ²³;

Data and personal information: names, ID numbers, phone numbers and bank card numbers and other personal and privacy data must not be used for GenAI model training or optimisation — a sectoral rule that goes further than PIPL’s consent architecture ²⁴, ²⁵ and effectively mandates de-identification pipelines before any training or fine-tuning27;

Computing power: institutions must build secure, autonomous and controllable computing bases; large institutions are encouraged to export computing capacity and mature models to small and medium-sized institutions — an explicit policy lever against a two-tier market ²⁶;

Ethics and fairness: institutions must establish AI ethics review and monitoring systems, and be able to justify the legitimacy of using protected characteristics ²⁷;

Agent-specific security: for AI agents, the AI Guiding Opinions name the failure modes to be controlled — memory poisoning, identity privilege escalation, tool abuse and loss of operational control — and require defences against prompt injection, chain-of-thought injection and multimodal attacks ²⁸;

Transparency and auditability: institutions must set explainability standards for high-risk scenarios, prominently label AI-generated content and proactively inform financial consumers17; logs must be retained for no less than the duration of the relevant business; periodic audits of models and algorithms are required ²⁹;

Outsourcing and supply chain: named-list management of outsourcing partners, internal evaluation frameworks for external models, concentration-risk management and open-source component review, with express warnings against supply-chain poisoning ³⁰;

Reporting duty: institutions deploying GenAI in public-facing services or high-risk scenarios must report to the NFRA or its local offices — note how the “public-facing” trigger borrows the CAC’s threshold vocabulary ³¹.

The AI Guiding Opinions also address market-level risk directly: institutions must guard against convergent investment strategies and the amplification of market volatility, and are prohibited from abusing AI to generate false information or manipulate prices ³². The PBOC’s FinTech Development Plan had already flagged herding risk among the dangers to be stress-tested[author‑note: see Section 4 — Layer Two policy-trajectory paragraph]; the AI Guiding Opinions go a step further, converting model-homogeneity risk into concrete supervisory obligations — a signal with obvious relevance for securities and asset-management business supervised by the CSRC as well.

Securities side. The CSRC has not issued an AI-specific rule; its IT management, algorithmic trading and outsourcing rules apply a technology-neutral lens, with suitability, record-keeping and content obligations applied to AI-generated research and advisory outputs. In practice, securities firms face the same Layer One duties as everyone else (GenAI filing for public-facing apps16,17, labelling of AI-generated content), while the AI-native trading and advisory applications described in Section 2 remain subject to the full weight of investor-suitability and licensing rules.

Layer Three: the data law trilogy

The Cybersecurity Law, Data Security Law and PIPL form the foundation on which both layers above rest. Four provisions matter most for AI in FS:

First, PIPL Section 2435: automated decision-making must be transparent and produce fair and impartial results; unreasonable differential treatment in “transaction conditions, including price” is prohibited; where automated decisions materially affect an individual’s rights, the individual has the right to an explanation and the right to refuse a decision made solely by automated means. Functionally, this captures credit scoring, automated underwriting, dynamic pricing — the exact scenarios the NFRA has now designated as high-risk ³³ — and PIPL separately requires a prior personal-information-protection impact assessment for automated decision-making.

Secondly, PIPL’s sensitive-information rules 36: financial accounts, identity documents and personal property status are expressly enumerated as sensitive personal information, processing of which requires a specific purpose, sufficient necessity and separate consent. Combined with the NFRA’s prohibition on using names, ID numbers, phone numbers and bank card numbers in GenAI training ³⁴, this creates a two-level constraint: even with consent, sectoral guidance now bars the most identifying fields from training corpora altogether.

Thirdly, the Network Data Security Management Regulation (State Council Decree No. 790, effective 1 January 2025) ³⁵ consolidates data-handling obligations across the trilogy: it reiterates separate consent for sensitive personal information36, obliges providers of GenAI services to strengthen training-data security management16, and imposes annual risk-assessment and reporting duties on handlers of important data. For financial institutions — archetypal handlers of important data — this adds a recurring, board-visible compliance cycle that intersects with the NFRA’s own periodic audit expectations31.

Fourthly, the CBDT regime (security assessment, standard contract, or certification) shapes AI architecture choices: architectures that send prompts or customer data to offshore model APIs, or that consolidate data on group-level platforms abroad, will frequently be impracticable or require case-by-case structuring — a principal reason domestic models and private deployment dominate Chinese FS practice (Section 1), and a live issue for multinational institutions running global AI stacks.

Layer Four: technical standards and industry norms

The PBOC’s financial industry standards supply the technical grammar for the rules above. JR/T 0221—2021, the Evaluation Specification of AI Algorithms in Financial Applications ³⁶, sets the evaluation dimensions, including security, explainability, accuracy and performance — against which financial AI algorithms are assessed. JR/T 0287—2023, the Guidelines for Information Disclosure of AI Algorithms in Financial Applications ³⁷, prescribes disclosure principles, forms and content across seven aspects — algorithm combination, algorithm logic, algorithm application, algorithm data, the responsible entity, algorithm changes and algorithm audits — to enhance explainability and protect financial consumers’ right to know. Together with the mandatory national standard GB 45438-2025 on content labelling ³⁸, these standards convert statutory abstractions (“transparency”, “explainability”) into testable engineering requirements. The NFRA’s AI Guiding Opinions contemplate a further technical framework and security standards for GenAI in banking and insurance, to be developed with relevant departments — expect sector-specific GenAI security standards to follow ³⁹.

Core Regulatory Principles & Operational Compliance Roadmap

Four core principles of Chinese AI-in-finance regulation

Balancing development and security. Adoption is state policy; the filing, assessment and labelling infrastructure is designed to enable adoption safely, not to impede it. The “AI Plus” Initiative (State Council Document No. 11 of 2025) ⁴⁰ expressly names finance among the sectors for wide deployment of smart terminals and AI agents, targeting penetration of next-generation intelligent terminals and AI agents in over 70% of applications by 2027.

“Who uses, who is responsible” ⁴¹. The financial institution — as service provider and technology user — bears ultimate responsibility, regardless of outsourcing or third-party models32; internal accountability must be traceable at every node.

Full-lifecycle governance ⁴². From development and testing through deployment, monitoring and exit, each stage carries defined controls: admission management, classified grading23, human intervention25, logging and audit31, and decommissioning conditions.

Human-machine collaboration and human final decision ⁴³. Human-in-the-loop is not aspirational but codified: insufficiently explainable AI in high-risk scenarios is auxiliary only, and decisions affecting customer rights or with material financial impact require human review nodes25.

“’Who uses, who is responsible’: in China, the financial institution bears ultimate responsibility for its AI — outsourcing and third-party models dilute nothing32.”

A compliance roadmap for deploying AI in Chinese financial institutions

Board level. Articulate AI principles reflecting the institution’s values and risk appetite; designate the dedicated board committee contemplated by the AI Guiding Opinions ⁴⁴; approve high-risk deployments through the risk management committee24.

Governance framework. Integrate AI strategy into overall risk management; establish an AI risk committee or equivalent; maintain a classified and graded AI application inventory23; define accountability and escalation lines; put ethics review on a standing footing29.

Operational processes. Run use-case and impact assessments before deployment (mapping each use case against the CAC’s “public-facing” threshold16 and the NFRA’s high-risk catalogue24); adopt policies for labelling17, disclosure and consumer notification consistent with the Labelling Measures and JR/T 0287—202339; build logging and information-management systems meeting the retention expectations31 (no less than the business duration); train staff; and maintain continuing communication with regulators — including the NFRA reporting duty for public-facing or high-risk GenAI applications33.

Model deployment. Decide deployment and compute architecture (on-premises vs. cloud) against data localisation and CBDT rules; map data flows, with particular care that names, ID numbers, phone numbers, bank card numbers and other personal and privacy data are excluded from GenAI training27; conduct supply-chain due diligence on procured models — verifying CAC filing status for external GenAI models26 — and allocate responsibilities, audit rights and exit arrangements contractually32.

Future Outlook & Strategic Conclusion

Several trajectories appear durable. First, financial foundation models are becoming general-purpose industry infrastructure, reshaping how products and services are delivered, how institutions operate and how risk is managed — the shift from “+AI” to “AI+”, in the phrasing now common among Chinese bank technology leaders. Second, human-AI collaboration is likely to remain the dominant working model for the long term — a prediction now embedded in regulation44 itself, which reserves final decision-making to humans in high-risk scenarios25. Third, the risk agenda — accuracy and explainability, privacy and IP, ethics and bias — will continue to demand proactive governance: risk assessment, human oversight and transparent communication.

Two further developments warrant close monitoring. On the legislative front, the NPC Standing Committee’s 2026 legislative work plan retains AI legislation as a preparatory review item, while the State Council’s 2026 legislative plan calls for improving AI governance and accelerating comprehensive legislation for the healthy development of AI — suggesting that today’s layered, “small-incision” rules will progressively consolidate. On the international front, China is investing in global AI governance — the Global AI Governance Action Plan issued at the 2025 World Artificial Intelligence Conference, and the proposed World AI Cooperation Organization — which may shape future cross-border regulatory convergence relevant to multinational FS institutions.

“Watch the consolidation: China is moving from fragmented, sector-specific rules towards comprehensive AI legislation — and from technique regulation towards a risk-tiered, full-lifecycle model already previewed by the financial regulator43.”

For financial institutions, the strategic conclusion is straightforward: continuous monitoring of regulatory change is now a compliance function in its own right. Institutions that have already built use-case inventories, classification and grading23, human oversight mechanisms25, model risk management38,39 and data governance36,37 will find themselves substantially compliant with whatever form the eventual comprehensive statute takes — and best placed to capture the productivity gains that Chinese FS institutions are already reporting at scale.

Belize – AI 2026/27

Germany – AI 2026/27

China – AI 2026/27

Yirui Zhang (Alicia) Partner

AI IN CHINESE FINANCIAL SERVICES: FROM FINTECH TO FOUNDATION MODELS — MARKET PRACTICE, REGULATORY ARCHITECTURE AND THE COMPLIANCE ROADMAP

From FinTech to financial foundation models and AI agents

Market landscape and infrastructure characteristics

Use Cases & Institutional Practices

Securities and asset management: research, advisory and AI-native apps

Insurance and payments: underwriting, claims and anti-fraud

Value Creation vs. Legal & Compliance Risks

Core legal and compliance risks

Layer One: horizontal AI and algorithm governance (led by the CAC)

Layer Two: financial-sector rules (NFRA, PBOC, CSRC)

Layer Three: the data law trilogy

Layer Four: technical standards and industry norms

Core Regulatory Principles & Operational Compliance Roadmap

A compliance roadmap for deploying AI in Chinese financial institutions

Future Outlook & Strategic Conclusion

Foreword – The World Bank – Mr Dessislav Dobrev

Cyprus AI – Pelaghias, Christodoulou, Vrachas

Finland AI 2025/26 – Fondia

Hong Kong AI 2025/26 – Bird & Bird

Turkey AI 2025/26 – Yazicioglu

UK AI 2025/26 – Simmons & Simmons

South Korea AI 2025/26 -Kim & Chang

US AI 2025/26 – Eversheds Sutherland

UAE AI 2025/26 – Gibson Dunn & Crutcher

International AI in Financial Services Review 2025/26

Ghana – AI 2026/27

Japan – AI 2026/27

Saudi Arabia – AI 2026/27

Romania – AI 2026/27

US – AI 2026/27

Kazakhstan – AI 2026/27

Hungary – AI 2026/27

South Africa – AI 2026/27

Belize – AI 2026/27

UK – AI 2026/27

Mexico – AI 2026/27

Kazakhstan – AI 2026/27

Germany – AI 2026/27

Foreword – AI 2026/27

International AI in Financial Services Review 2026/27

Navigation

Yirui Zhang (Alicia)
Partner