- About Us
- Investor Relations
- ESG
- Product & Service
- Customer Service
Customer Service
- Visiting Taiwan
Internal Company Policy: Guidelines for the Use and Management of Internal Enterprise AI Tools
Section 2: Data Confidentiality Classification and Corresponding AI Tool Requirements
In alignment with the Company's Information Security Management System (ISMS), all data classifications strictly adhere to the Company's existing classification framework: Top Secret, Confidential, and Restricted. Employees must select appropriate AI tools based on the corresponding data classification level. Entering any sensitive data into unauthorized tools is strictly prohibited.
Top Secret: Data whose leakage would cause company-wide operational disruption, catastrophic business impact, or unpredictable financial losses. This includes Personally Identifiable Information (PII), encryption keys / API keys, system credentials/passwords, comprehensive core technology and operational decision-making documents, and core trade secrets.
Confidential: Data whose leakage would impair the Company's competitive edge, advantage competitors, or compromise client rights and interests. This includes restricted project documents, non-public internal operational data, financial forecasts, and unannounced partnership initiatives.
Restricted: Data whose leakage would impact localized business operations or cause limited losses without infringing upon individual client rights. This includes system logs, system configuration files, and network/circuit topology data.
The Company shall, based on the purpose, risk, technical characteristics, and level of impact of an AI system, disclose its use, primary purpose, capabilities, limitations, potential risks, and applicable conditions in an appropriate and easily understandable manner. When users interact with an AI system, they shall, where appropriate, be clearly informed that they are interacting with AI rather than a human. When the Company provides externally AI-generated text, images, video, audio, or other content, appropriate labeling, disclosure, or identification measures shall be adopted based on the level of risk and context of use. When AI participates in or influences decisions that have a significant impact on individual rights and interests, appropriate explanations shall be provided, including the key influencing factors, basis for judgment, applicable limitations, and methods of human review.
The Company's internal MyBro Assistant clearly states:
"MyBro Assistant is an AI and may make mistakes. | v4.0.0 (BETA)"
For applications that utilize large AI models provided by overseas vendors, such as GPT and other large language models (LLMs), the Company continuously monitors updates to model service versions and functionalities. Model service providers also continuously optimize and update their models to mitigate the risk of performance degradation over time. When implementing or updating AI models, the Company conducts functional, accuracy, stability, and output quality testing based on the actual application scenarios to verify the suitability of the models for specific business contexts.
The Company assesses the fairness and potential bias risks of AI models based on the intended use of the AI application and the potential impact it may have. For AI applications that may involve individual rights, personalized decision-making, or differential impacts on different users, the Company conducts testing and validation based on the actual application scenarios during initial implementation and model version updates. The testing evaluates whether the model produces results that are significantly unreasonable, discriminatory, or inconsistent with business requirements.
Currently, the Company primarily manages these risks through application-specific testing and validation, rather than conducting abstract fairness assessments solely at the model level. Test cases are designed according to the purpose of each application, and the model outputs are evaluated across different scenarios. If abnormal or unreasonable results are identified, the Company may take corrective measures such as adjusting prompts, application logic, or model versions, or introducing human review. The application is then retested and validated after the corrective measures have been implemented.
Taiwan Mobile demonstrates concrete execution and quantitative evidence in reducing its ecological footprint. Its green initiatives leverage its telecommunications infrastructure strengths across three key dimensions: physical data center architecture, renewable energy matching, and computational workload optimization.
Taiwan Mobile's IDC cloud data center is the first facility in East Asia to achieve Uptime Institute Tier III certifications across all three disciplines: Design, Constructed Facility, and Operational Sustainability. In 2024, the facility achieved its target of 100% renewable energy use six years ahead of schedule. Its Power Usage Effectiveness (PUE) can reach as low as 1.5, delivering a 25% improvement in energy efficiency compared with conventional data centers with a PUE of 2.0.
Taiwan Mobile has joined the global RE100 renewable energy initiative, pledging 100% renewable electricity enterprise-wide by 2040. In 2025, to counter grid volatility caused by the intermittency of solar and wind generation, the company deployed an AI-driven Green Power Matching Algorithm. This model accurately forecasts and aligns corporate power consumption curves with peak renewable generation windows, maximizing renewable power utilization and decoupling computing workloads from fossil fuels.
Addressing model deployment and computational overhead, Taiwan Mobile partnered with Phison Electronics to introduce the aiDAPTIV+ Turnkey On-Premises AI Solution. Conventional large language model (LLM) training relies heavily on power-dense, costly GPU clusters. The aiDAPTIVCache technology bypasses conventional VRAM bottlenecks by utilizing energy-efficient, cost-effective NAND flash storage as extended memory. This innovation enables enterprises to train and run inference on massive 100B to 405B parameter models using standard workstation hardware, drastically reducing high-end GPU resource dependency and energy consumption.
Taiwan Mobile has established a comprehensive, transparent, and Human-in-the-Loop (HITL) appeal and dispute-resolution mechanism for its flagship AI-driven fraud prediction solution, OP Scam Buster. Leveraging telecom big data and AI models, the system automatically detects and intercepts high-risk phone numbers and phishing URLs to safeguard users. Recognizing that machine learning models inherently carry a margin for false positives, Taiwan Mobile has instituted rigorous remediation and recourse workflows to protect legitimate businesses and benign websites from erroneous classification:
To safeguard database integrity against arbitrary modifications, phone numbers erroneously flagged as scam, high-risk, or telemarketing require a formal dispute process. The legitimate subscriber must submit an appeal via the dedicated email address (opscambuster@taiwanmobile.com) or by calling the 188 customer support hotline. The process mandates strict evidentiary compliance: applicants must provide the complete phone number, proof of ownership (e.g., subscription agreement or telecom bill), and matching identity documents or business registration certificates. Upon thorough human review and verification by specialized operations personnel, the misclassification is purged from the database and updates are synchronized across client-side applications.
When a benign webpage is misclassified as a security risk by the AI algorithms in conjunction with the Trend Micro threat database, users or website administrators can submit a dispute via Settings > Feedback within the app. Upon successful manual verification, Taiwan Mobile updates the system-wide whitelist. Furthermore, to uphold the principle of ultimate user agency, the client interface provides an override capability, allowing users to manually designate specific URLs as "Trusted" to directly bypass AI-enforced blocking rules.
Under the oversight of its ESG Steering Committee, Taiwan Mobile leverages rigorous quantitative metrics to demonstrate the positive environmental (E) and social (S) impacts of its AI implementations:
Mobile network operations—particularly 5G—account for a substantial share of a telecom operator's overall carbon footprint. Taiwan Mobile deploys AI/ML models to forecast traffic fluctuations and dynamically evaluate base station loads, replacing legacy fixed-schedule power-saving schemes. This intelligent energy-saving architecture has improved overall efficiency by 1.5 times. Specifically, this AI initiative yielded 2.11 million kWh in tangible electricity savings in 2024. In 2025, AI-powered energy-saving technologies optimized base station and data center operations, achieving electricity savings of more than 18 million kWh.
By implementing multiple efficiency measures and AI-driven workload orchestration across its IDC cloud data centers, Taiwan Mobile reduces approximately 24,930 metric tons of CO₂ emissions annually. To make this metric intuitive for non-technical stakeholders and the public, the report benchmarks this reduction as equivalent to the annual carbon sequestration of 65 Daan Forest Parks. Furthermore, MyCharge, the company's smart EV charging service integrated deeply via data APIs, was estimated to reduce approximately 760,830 kg of CO₂ emissions (approximately 760.83 metric tons of CO₂) in 2025.
AI deployment delivers equally significant metrics in enhancing public safety and combating cybercrime. To date, Taiwan Mobile's AI-powered anti-fraud system, OP Scam Buster, has proactively inspected more than 310 million suspicious domains, identifying and neutralizing over 80,000 fraudulent websites and malicious apps. In parallel, the Number Masking Service privacy-preserving routing mechanism has protected the personal call data of more than 300 million transactions/interactions, preventing sensitive data leakage.
In employee training and awareness capacity-building, Taiwan Mobile demonstrates a robust, defense-in-depth posture centered firmly on information security.
In routine training, Taiwan Mobile strictly mandates that all employees complete 3 hours of mandatory cybersecurity coursework annually. For IT personnel with privileged administrative access and operational business personnel, the company enforces an additional 9 hours and 3 hours of advanced elective courses, respectively, ensuring their technical expertise keeps pace with emerging threats.
Addressing novel threat vectors in the AI era, the company conducts systematic social engineering simulations utilizing AI-generated, multi-themed, hyper-realistic phishing email templates. These simulations mimic real-world threat scenarios—including malicious links, spoofed invoice notices, and fake job applications—to rigorously evaluate employee alertness. Post-drill remediation combines big-data behavioral analytics with targeted, expert-led training sessions, measurably hardening workforce resistance against sophisticated AI-assisted attacks.
Taiwan Mobile has officially designated 20 cybersecurity seeds across key business units who have completed specialized training in the Artificial Intelligence Management System (AIMS) standard (ISO/IEC 42001). These specialists possess practical expertise in executing AI risk identification, AI impact assessments (AIA), and cross-system compliance audits. This dedicated governance tier ensures that the enterprise maintains robust trust, ethics, and operational safety as it accelerates the deployment of generative AI and intelligent business applications.
