AGI Kingdom Insights
Cross-boundary AI data between Hong Kong and the Greater Bay Area.
A partner diligence guide to what moves, who can access it, why it is needed, where it is stored, and which legal or voluntary route may apply.
Scope
“AI data” is not one legal category.
AI collaboration across Hong Kong and the Mainland cities of the Greater Bay Area can involve more data flows than a project description initially suggests. Training or evaluation datasets are only part of the picture. Prompts, user feedback, system logs, model telemetry, embeddings, support tickets, account records, backups and remote technical access may also cross a boundary.
The first question is whether any part of the proposed flow relates to an identifiable living individual. If it does, it may be personal data under Hong Kong’s Personal Data (Privacy) Ordinance (PDPO), regardless of whether the project calls it a dataset, log, prompt, feature vector or model input. Information that is genuinely outside that definition may be governed by other contractual, confidentiality, cybersecurity, sector or Mainland requirements. The label “AI data” does not answer those questions.
This guide separates three issues that should not be merged: Hong Kong’s existing personal-data requirements; the voluntary Greater Bay Area Standard Contract route; and practical partner diligence that reduces operational risk but is not itself a statement of law. The correct route depends on the actual parties, data, direction, locations, purpose and system design.
Hong Kong legal baseline
Moving data across a boundary does not remove the PDPO duties that already apply.
The Office of the Privacy Commissioner for Personal Data describes the PDPO as technology-neutral and principle-based. Its Data Protection Principles govern how a Hong Kong data user collects, holds, processes and uses personal data. An AI project does not acquire a separate exemption because it uses a large model, an autonomous agent or a novel technical architecture.
| Area | Official Hong Kong baseline | AI-project question |
|---|---|---|
| Collection | Personal data should be collected for a lawful purpose directly related to the data user’s function or activity, and be necessary, adequate but not excessive. | Does every field, prompt attachment and log element serve the defined use case? |
| Notice | When collecting directly, the data subject should receive the information required under the PDPO, including the purpose and classes of potential transferees. | Would a reasonable person understand the AI use and the receiving parties? |
| Accuracy and retention | All practicable steps should be taken to keep personal data accurate and not longer than necessary. | Who corrects source errors, and when are prompts, logs, embeddings and backups erased? |
| Use | A new purpose unrelated to the original collection purpose generally requires the data subject’s express and voluntary consent. | Is vendor model training or product improvement a different purpose from providing the contracted service? |
| Security | All practicable steps should protect personal data from unauthorised or accidental access, processing, erasure, loss or use. | Can overseas support, subprocessors, plugins or connected agents reach the data? |
| Processors | Where another party processes data on behalf of the data user, contractual or other means should address applicable retention and security requirements. | Do the contract, architecture and actual vendor settings say the same thing? |
| Transparency and rights | Data users should maintain open personal-data practices and workable access and correction procedures. | Can the team locate a person’s data across the AI workflow and respond appropriately? |
The PCPD also publishes Recommended Model Contractual Clauses for cross-border transfers. They provide two reference structures—data user to data user and data user to data processor—and address matters including purpose limitation, security, retention and erasure, onward transfer, and data-subject requests. They can be useful drafting tools, but they do not turn an unsuitable architecture into a compliant one or replace project-specific professional advice.
Optional facilitation route
The GBA Standard Contract is voluntary, bounded and operationally specific.
The Digital Policy Office states that the Standard Contract for Cross-boundary Flow of Personal Information within the Guangdong-Hong Kong-Macao Greater Bay Area—the GBA Standard Contract—is adopted on a voluntary basis. Since 1 November 2024, the facilitation measure has been open to all industry sectors where the relevant parties are in Hong Kong and one of nine Mainland GBA cities: Guangzhou, Shenzhen, Zhuhai, Foshan, Huizhou, Dongguan, Zhongshan, Jiangmen and Zhaoqing. Despite the measure’s formal name, the current route does not include Macao.
The measure is designed for eligible cross-boundary flows of personal information between Hong Kong and those nine cities. It is not a general approval for all data, all locations, every overseas affiliate or every AI service. Personal information classified or promulgated by relevant authorities or regions as critical data is excluded. The HKSAR Government expressly states that the measure does not displace the Privacy Commissioner’s supervisory and regulatory role and that personal-data processing and cross-boundary transfer in Hong Kong continue to be regulated under the PDPO.
| Feature | Official position | Operational implication |
|---|---|---|
| Adoption | The Hong Kong facilitation measure is voluntary. | Eligibility does not remove the need to assess all applicable Hong Kong, Mainland and sector requirements. |
| Geography | It covers Hong Kong and the nine listed Mainland GBA cities; Macao is currently outside the route. | Confirm the registration or location of the parties and every access, storage and onward-transfer path. |
| Critical data | Personal information classified or promulgated by relevant authorities or regions as critical data is excluded. | Do not rely on this route until the data classification has been checked against current official requirements. |
| Templates | The parties use the prescribed contract and undertaking templates. | Do not silently rewrite a required template into a generic cloud or data-processing agreement. |
| Impact assessment | The DPO states that the personal information protection impact assessment must be conducted within three months before the filing date. | Complete a current, fact-specific assessment; do not reuse an old report after a material architecture change. |
| Filing | Hong Kong’s personal information processor or recipient must file with the DPO within ten working days from the contract’s effective date, which the current guide identifies as its signing date. | Use the latest DPO filing guide and verify which party must submit which materials in each jurisdiction. |
| Direction | One contract covers one direction of flow. | A two-way flow requires two contract sets with sender and recipient roles reversed. |
| Outside-GBA access | Information received through this route must not be provided outside the GBA, including through transfer, storage or access. | Global support, overseas administrators, external model APIs and multi-region backups can be decisive design issues. |
The restriction on provision outside the GBA deserves attention before the team selects cloud infrastructure or a model provider. A service may be hosted in Hong Kong while still permitting administrators, subprocessors, monitoring services or automated systems outside the GBA to access information. A region label on a sales page is not a complete data-flow map.
Mainland requirements, exemptions and alternative mechanisms can change and depend on facts that this article cannot determine. Organisations should use the current official materials and qualified advisers in the relevant jurisdiction rather than treating the voluntary Hong Kong facilitation route as an automatic answer.
Partner diligence
Map the real system before selecting the legal or contractual route.
The following sequence is a practical risk-control framework. It is not a substitute for legal, privacy, cybersecurity or sector advice.
1. Map the complete data flow.
Record each data category, originating system and jurisdiction, sending and receiving party, processing purpose, storage location, user and administrator access location, subprocessor, onward recipient, retention period, and deletion or return process. Include system logs, backups, monitoring tools, model-evaluation records, prompts, embeddings, vector stores and remote support—not only the primary dataset.
2. Separate necessary data from convenient data.
For each field or dataset, ask whether it is genuinely required for the defined outcome. Consider whether aggregated, synthetic, masked or otherwise de-identified information can serve the purpose; whether processing can occur locally with only a result crossing the boundary; and whether the proposed retention period corresponds to a real operational need. A file labelled “anonymous” should not be assumed to fall outside privacy controls if individuals remain identifiable in practice.
3. Verify purpose and authority.
Check whether the proposed use matches what was communicated when the data was collected. A model that learns from the data, a vendor that reuses it for evaluation, or a provider that improves a shared service may create a materially different purpose from processing the information solely to deliver a contracted service. Document the applicable notice, consent or other authority; a general reference to “AI improvement” is not a precise operating instruction.
4. Diligence the entire technical chain.
Ask each partner and provider to identify hosting and backup regions; remote access locations; subcontractors and model providers; prompt, output and log retention; use for training or service improvement; encryption and key control; role-based permissions and audit logging; incident notification; deletion from active systems and backups; and procedures for applicable access and correction requests. Contract terms should match actual settings and network paths.
5. Allocate responsibility and change control.
Name the owner who can approve, pause or terminate the flow. Define who handles security incidents, data-subject enquiries, deletion evidence and regulator contact. Reopen the assessment if the dataset, purpose, model provider, cloud region, recipients, access locations, retention period or onward-transfer arrangement changes. A short pilot using non-personal, synthetic or tightly minimised data can be a lower-risk way to validate the workflow before a broader transfer.
Stop conditions
An incomplete data map is a reason to pause—not a reason to write a broader contract.
Pause the transfer and obtain appropriately qualified review if any of the following remains unresolved:
- the team cannot determine whether the information contains personal data or sensitive, confidential or regulated information;
- the team cannot determine whether relevant authorities or regions have classified or promulgated the information as critical data;
- the purpose, legal counterparty or recipient is described only in broad terms;
- a provider may train a shared model on transferred prompts, outputs, logs or datasets;
- hosting, backups, telemetry, remote support, plugins or subprocessors may operate outside the approved geography;
- an onward recipient or administrator cannot be identified;
- the project cannot implement required retention, deletion, access or correction processes;
- security and incident-response responsibilities are not allocated between the parties;
- the proposed use differs materially from the purpose communicated to individuals;
- high-impact or sensitive uses have not received an appropriate impact assessment and specialist review; or
- the project intends to rely on the GBA Standard Contract while permitting transfer, storage or access outside the GBA.
The decision can be expressed simply. Continue when the data, purpose, parties, route and controls are verified. Pause when a bounded check or lower-risk pilot can resolve a material unknown. Terminate the proposed flow when essential rights, security or geographic restrictions cannot be met. Concentrate resources only after a controlled pilot shows repeatable value and the governance process works under realistic conditions.
Frequently asked questions
Five questions about cross-boundary AI data.
Does all AI-related data count as personal data in Hong Kong?
No. AI data is not a single legal category. Information relating to an identifiable living individual may be personal data under the PDPO, including information embedded in prompts, logs, support records or datasets. The assessment depends on the actual information and context, not the project label.
Is the GBA Standard Contract mandatory for every Hong Kong–Mainland AI project?
No. The DPO describes it as a voluntary facilitation measure for eligible cross-boundary flows of personal information between Hong Kong and the nine Mainland GBA cities. The project must still assess the PDPO, applicable Mainland requirements, sector rules and its actual architecture.
Can an overseas cloud administrator access information transferred under the GBA Standard Contract?
The DPO states that personal information received through the GBA Standard Contract must not be provided to organisations or individuals outside the GBA, including through transfer, storage or access. Teams should verify the real locations of administrators, subprocessors, backups and support before relying on this route.
Does a two-way flow require one contract or two?
Official DPO guidance states that each contract covers one direction. A two-way flow therefore requires two contract sets, with the sender and recipient roles reversed.
When should a cross-boundary AI pilot be paused?
Pause when the data, purpose, recipient, access geography, subprocessors, model-training use, retention, deletion, security responsibilities or applicable route cannot be clearly verified. Sensitive or high-impact uses should receive appropriately qualified legal, privacy, security and sector review.
Official sources
Primary Hong Kong materials used for this guide.
The diligence framework and stop conditions are AGI Kingdom Group’s independent editorial synthesis. The official materials below provide legal and administrative context; none certifies, approves, funds or endorses this guide or AGI Kingdom Group.
- Digital Policy Office — GBA Standard Contract facilitation measure, including current eligibility, workflow, FAQs and documents.
- Digital Policy Office — Filing Guidelines on the GBA Standard Contract, March 2026.
- Digital Policy Office — Implementation Guidelines on the GBA Standard Contract.
- Privacy Commissioner for Personal Data — Guidance on Cross-boundary Data Transfer: GBA Standard Contract, December 2023.
- Privacy Commissioner for Personal Data — The Personal Data (Privacy) Ordinance at a Glance.
- Privacy Commissioner for Personal Data — Recommended Model Contractual Clauses for Cross-border Transfer of Personal Data, May 2022.
- Privacy Commissioner for Personal Data — Artificial Intelligence: Model Personal Data Protection Framework, June 2024.
- HKSAR Government — GBA Standard Contract facilitation measure extended to all sectors, 1 November 2024.
Method and limitations
Use the latest official materials and the facts of the proposed flow.
This page was last reviewed on 1 September 2026. Administrative procedures, filing documents, Mainland requirements, sector rules, technical services and data architectures can change. Current official documents and decisions of the relevant authorities take precedence.
This article provides general information only. It is not legal, regulatory, privacy, cybersecurity, engineering, investment or other professional advice, and it does not determine any organisation’s eligibility for the GBA Standard Contract or another transfer mechanism. AGI Kingdom Group does not claim government endorsement, regulatory approval, client experience or implementation capability through the publication of this guide. Each project should obtain qualified advice appropriate to its parties, data, transfer direction, technical architecture and sector before acting.