Asia-Pacific: China Agent Governance, CSRC AI Plans, Job Risk Reviews; Japan Gennai Standards; Korea Data Zone Opens

On 19 October 2026, China’s Ministry of Commerce (MOFCOM), the Office of the Central Cyberspace Affairs Commission and seven other departments made public the “Opinions on Accelerating Technology-Driven Development in Commerce” (Shang Jian Han [2026] No. 551). Alongside support for commercial AI agents in shopping assistance, business operations and logistics, the document proposes tiered governance by application scenario and explores publishing a negative list for data use.

The document is dated 13 October 2026 and was published on the official website on 19 October. It is a published set of policy opinions, with implementation coordinated by MOFCOM and the relevant departments. Item 15 incorporates use-case tiering, data-use boundaries and risk monitoring into commercial AI agent governance, calling for clear boundaries and rules for the collection, use and sharing of commercial data, and supporting stress testing and sandbox pilots within controlled limits. The negative list remains exploratory. The document does not yet specify prohibited uses, agent risk tiers, filing procedures or uniform log-retention periods; it therefore does not establish a complete new deployment-approval regime for commercial AI agents.

Retailers, e-commerce businesses, logistics companies and agent providers should consider extending governance beyond the model to actual business actions and their data flows. For shopping-assistance agents, check whether user profiles and order records are sent to external models and whether providers may use those data for their own training. For agents that can change prices, issue coupons, process refunds or adjust deliveries, restrict tool permissions and operating scope, require human confirmation for high-impact actions, and test unauthorised calls, repeated execution and stopping under abnormal conditions. Procurement contracts should define data purposes, subcontracted processing, version-change notifications and responsibilities for assisting incident investigations. Operational records should link user authorisation, key actions and response outcomes. These are deployment recommendations based on the policy direction; specific configurations still depend on business risks and applicable rules.

Tiered governance for commercial AI agents has a more direct policy precedent. On 8 May 2026, the Cyberspace Administration of China, the National Development and Reform Commission, and the Ministry of Industry and Information Technology published the “Implementation Opinions on the Compliant Application and Innovative Development of AI Agents”. That document already proposed classification and tiering by use case and potential impact, with measures such as filing, testing and recalls of problematic products in sensitive fields and key industries under the relevant rules. For some lower-risk areas, including leisure, entertainment and routine office work, it envisaged compliance self-assessments, information reporting and platform management. The new opinions apply this direction more specifically to commerce, adding commercial data-use boundaries and exploration of a negative list. Enterprises still need to determine the applicable arrangements from their functions, sector and potential impact, and monitor whether commercial use-case tiering methods, negative-list entries and sandbox participation conditions are further specified.

On 19 October 2026, China Securities Regulatory Commission (CSRC) Chairman Wu Qing told the 2026 Financial Street Forum annual conference that the regulator would soon issue and implement guidance on safe and compliant AI use in the securities and futures industry. The guidance would improve regulatory rules on business development and data security and promote safe, efficient use in risk management, investment research and other priority scenarios.

This is an advance regulatory policy announcement; specific requirements await the formal document. The speech identified the guidance as one of four measures to be introduced soon, and highlighted algorithmic black boxes and AI-enabled cyberattacks as challenges to the operational resilience of financial infrastructure. The confirmed development is therefore an intention to improve rules for industry AI use; the announcement does not establish new model-filing, mandatory certification or uniform deployment-approval requirements.

Securities firms, futures companies and their technology providers should consider inventorying use cases across internal research assistance, customer service, risk-management decisions and trade execution, identifying accountable owners, models called and accessible data. Investment research systems should check cited sources, data currency and financial figures, preventing unreviewed generated content from flowing directly into research conclusions or client materials. Before client holdings, transaction records or unpublished research enter an external model, review data-processing permissions, confidentiality boundaries and the provider’s training uses. Risk-management applications can be tested for false positives, false negatives and unusual market conditions, with human review arrangements. Connections to trading tools warrant further restrictions on execution permissions and preparation of stopping and rollback measures. Procurement contracts should address model-update notifications, service interruptions and responsibilities for assisting incident investigations. These are implementation-readiness recommendations; any specific new obligations depend on the formal document.

The announcement builds on the industry’s existing information-security regime. The “Administrative Measures for Cybersecurity and Information Security in the Securities and Futures Industry”, effective from 1 May 2023, already address pre-deployment testing, supplier management, protection of investors’ personal information and emergency response. They also specify that institutions’ cybersecurity and information-security responsibilities are not transferred or reduced through the use of external products or services. For important information systems within scope, Article 18 already requires business logs to be retained for at least five years and system logs for at least six months. While awaiting the AI-specific guidance, firms can check whether relevant applications are covered by existing system-management and recordkeeping arrangements. Subsequent review should focus on the formal document’s scope, use-case classification, model evaluation, external services and implementation timetable.

On 23 October 2026, the “Opinions on Adapting to Advances in Artificial Intelligence to Promote Employment-Friendly Development” (Guo Ban Fa [2026] No. 29), issued by the General Office of the State Council, were made public. They set out the establishment of a mechanism to assess employment risks from AI adoption, mitigation of job displacement, and governance of AI use in human resources management.

Dated 16 October 2026, these are published policy opinions to be implemented by regional authorities and relevant departments, with coordination by human resources and social security departments and development and reform departments. The core arrangements call for employment-risk assessments alongside project reviews for AI projects in market-access areas that may cause substantial employment disruption, with assessments in other areas reflecting their specific circumstances. Enterprises are encouraged to explore unemployment buffer periods, alongside arrangements for advance notice and job-transition support during large-scale adoption. The opinions also call for addressing, under the law, employment discrimination, fraudulent recruitment, and infringements of privacy and rest rights arising from algorithmic misuse. They do not yet specify a quantitative threshold for substantial employment disruption, a uniform assessment procedure or a fixed buffer period.

Enterprises automating customer service, content processing or back-office operations should consider incorporating job impacts into project initiation and reviews of deployment expansion. Record which tasks AI assists and which job responsibilities or staffing requirements change, and budget for employee communications, training, redeployment and human takeover. For AI recruitment, interviewing, performance evaluation and scheduling systems, examine the relevance of input data to job requirements, test screening bias and erroneous rejections, and retain decision rationales, human-review records and corrections. Procurement reviews should check whether providers support version tracking, review of results and deletion of personal information. Business, HR, technology and compliance teams should jointly confirm deployment conditions. These are implementation-readiness recommendations; specific assessment, reporting and employment requirements still need to be checked against applicable rules and subsequent arrangements.

The State Council’s “Opinions on Deepening the Implementation of the AI Plus Initiative”, published in August 2025, already called for stronger employment-risk assessment. The new opinions further specify links to project review, job-transition support, and monitoring and response mechanisms. Subsequent developments include how local and sector authorities flesh out these arrangements. For regional comparison, Annex III of the EU Artificial Intelligence Act includes recruitment screening, certain employment-related decisions, and worker monitoring and evaluation within its high-risk classification framework, subject to the relevant conditions and exceptions. China’s new opinions also address the broader job-displacement effects of business automation. Enterprises operating across regions can maintain a shared project register while separately assessing organisational employment changes and the effects of systems on individual rights.

On 20 October 2026, Japan’s Digital Agency published governance, cybersecurity and safety standards for the Government AI platform Gennai. The standards define the responsibilities of users, administrators and developers, alongside the start of a phased rollout of AI agent environments to ministries and agencies.

The Gennai Standards were developed with support from Japan’s AI Safety Institute (AISI). They directly govern the relevant roles within Gennai and can serve as a reference for local governments and enterprises. For agents that autonomously execute sequences of tasks, the standards require limits on purposes and execution permissions, staff confirmation or approval before important operations, and controls for log retention, suspension of use and periodic evaluation. For Gennai developers, they also require production changes to undergo review and approval; technical measures must prevent any party from tampering with or deleting security-critical logs. Least privilege, separation of permissions and incident-recovery procedures are also required.

Suppliers involved in government projects, and enterprises drawing on this approach, should consider bringing authorisation, approvals, tool execution, logs and suspension capabilities together in deployment acceptance checks. Configure separate permissions for information retrieval, code execution, file modification and external-system calls. Define human-confirmation and blocking conditions in advance for high-impact actions such as bulk updates and file deletion. Acceptance testing should cover prompt injection, unauthorised reads and anomalous continued execution, verify that suspension blocks subsequent tool calls, and allocate responsibility for credential revocation, data recovery and reactivation. Procurement reviews should check whether the organisation can manage permissions, export audit evidence, and migrate data and tool configurations when changing models or cloud services. These are deployment recommendations; specific requirements depend on project roles, contracts and applicable rules.

Gennai began a large-scale trial for approximately 180,000 government staff on 28 May 2026. The new announcement advances AI agent environments and role-based governance rules together. The published plans include introducing Gennai Kobo, an environment for building and running more advanced agents, around December 2026, and releasing the relevant environments as Gennai OSS Ver.2.0 around February 2027. The standards also call for decoupling functions such as cloud services, models, data, permissions and logs, and validating alternative configurations and migration paths during normal operations to prepare for service discontinuation, price changes or performance degradation. Enterprises can incorporate supplier substitutability into architecture and procurement assessments; subsequent rollout and open-source release dates remain subject to official announcements.

On 23 October 2026, Korea’s Personal Information Protection Commission (PIPC) announced that the Personal Information Innovation Zone operated by the Daegu Digital Innovation Promotion Agency (DIP) had begun full operation that day, with AI robotics and healthcare data research and use identified as priority areas for support.

This is the opening of an operating institution under the existing innovation-zone scheme. The mechanism uses environmental controls, including multi-factor authentication and real-time screen recording, to support more flexible use of pseudonymised data. This includes relaxing the degree of pseudonymisation, linking data with different linkage keys, retaining data over longer periods for ongoing research, and enabling reuse by third parties. It also supports sample-based checks of large video and image datasets and validation of privacy-enhancing technologies. These facilities depend on the security conditions of the designated environment; the specific data, research purposes and scope of external sharing still need to be verified for each project.

Robotics and healthcare AI companies and research teams should consider the zone in decisions on data research environments and processing workflows. Before applying, confirm available data, access eligibility and whether the project permits model training or fine-tuning, or is limited to analysis and validation. Where sensory data contain faces, speech or trajectories, assess residual identifying features and the risks of linkage with other data. Project design should also clarify the conditions for taking model weights, vectors, statistical results and samples out of the zone. Where external model APIs are required, first verify connectivity and transmission arrangements before deciding on execution within the zone or testing in an isolated environment. Contracts should allocate responsibility for access auditing, data retention, deletion and handling abnormal events. These are deployment recommendations; specific processing arrangements depend on operating rules and the project’s permitted scope.

Daegu received preliminary designation in April 2026, then developed analysis facilities and improved its systems and networks. Following PIPC on-site verification and remediation, it received final designation and began operation on 23 October. DIP’s existing Pseudonymised Information Utilisation Support Centre can work with the zone to provide basic support and in-depth analysis services. PIPC also announced plans to trial an inter-zone linkage hub in November 2026 to support data linkage and use across operating institutions, and to invite applications for further operators in the first quarter of 2027. Enterprises can track the hub’s actual connectivity, data-linkage conditions and available resources, aligning research schedules with the data-processing capabilities that are permitted.