## tarea2025054-print-pdf

## Source details

**Canonical URL:** [tarea2025054-print-pdf](https://www.imf.org/-/media/files/publications/tar/2025/english/tarea2025054-print-pdf.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/tar/2025/english/tarea2025054-print-pdf.pdf.md)
- [Structured JSON version](/-/media/files/publications/tar/2025/english/tarea2025054-print-pdf.pdf.json)

---

### Mission, scope, and timeline
- IMF Fiscal Affairs Department capacity development mission visited Yangzhou and Beijing, China, from November 4 to 15, 2024, to provide guidance on advancing risk management for large taxpayers.
- First week: seven-day workshop at the Yangzhou Training Academy delivering comprehensive training and sharing presentation materials in advance.
- Second week: technical discussions in Beijing and submission of the draft report; final aide-mémoire presented to officials on November 15, 2024.
- Mission team led by Ms. Cindy Negus (Senior Economist, FAD) and included Messrs. David Hadwick, Cameron Smith, and Ms. Maureen Kidd (external experts).
- Key STA interlocutors met by the team include: Mr. Bu Xianglai; Ms. Meng Yuying; Ms. Li Na; Mr. Yue Song; Mr. Wang Dan; Mr. Li Bin; Mr. Du Jianwei; Ms. Yang Xiao; Ms. Liu Yingjie; and other central and provincial taxation officials.
- Acknowledgement of support from Mr. Liu Xin, Office of Academic Affairs of the National Tax Academy, Yangzhou Campus.

### Executive summary — strategic thrusts and analysis
- Modernization: STA has modernized operations with digital solutions and advanced analytics to strengthen risk assessment capabilities.
- Objective: Use data analytics to improve compliance, reduce compliance costs, and strengthen relationships with the large business tax community.
- Emphasis areas covered during the mission: organizational governance, application of artificial intelligence (AI) and big data, Compliance Risk Management (CRM), risk assessment approaches, strategic planning, treatment strategies, Compliance Improvement Plans (CIP), and capability building for large taxpayer management.
- Two priority areas for recommendations:
  - Enhancing Risk Management: move from an entity-by-entity approach to an aggregate population-level approach to identify and prioritize the highest risk cases; establish a proactive framework anticipating data and technology needs; identify emerging data sources and leverage advanced analytical tools.
  - Strengthening Governance and Accountability: designate the Large Enterprise Tax Administration Department (LETD) as the primary functional department at headquarters to centralize coordination; establish a dedicated CRM unit and a senior-level committee for CRM strategy and governance.

### Key statistics and facts
- STA total tax revenues collected in 2023: 17.7 trillion RMB.
- 2023 year-on-year increase: 10.1 percent.
- Effective increase after adjustments for excess input VAT credit refund policy: 2.3 percent.
- Domestic VAT increase from 2022 to 2023: 42.2 percent.
- Corporate (enterprise) income tax (CIT) change 2022–2023: decreased by 5.9 percent.
- Large Taxpayer segment contribution to total revenue: 45.70 percent.
- USD equivalent noted in source: Approximately 2.478 trillion USD.
- STA estimates close to 10,000 staff engaged in large taxpayer business management across the country.
- LETD HQ staff: about 40 staff.
- There are 2,000 large taxpayers managed through large taxpayer management, with LETD oversight.
- Among them, 140 are considered particularly important and are jointly managed by the General Taxation Department of the State Taxation Administration and the provincial and lower large taxpayer management departments (headquarters-to-headquarters management).
- Of these 2000 taxpayers, the top 1000 generate about 40 percent of STA revenue while the entire 2000 generates over 45 percent 910.
- STA mandate delivered in close to 4000 locations (600 manage large taxpayers) as well as 41,000 township level divisions.

### Current large taxpayer administration model and challenges
- LETD created in 2008 to oversee large taxpayer management; hybrid role combining headquarters and operational responsibilities; LETD does not include debt management functions.
- LETD coordinates with other functional departments (e.g., VAT Department, Income Tax Department) and is supported by TBDRM (Tax Big Data and Risk Management Bureau) with responsibility for the risk management model.
- Annual LETD process: formulate an annual risk management plan, consult in advance with Big Data and Risk Administration Bureau and Tax Inspection Department, then send audit referrals to the Audit Department.
- Large taxpayers are distributed nationally and managed by provincial and lower tax authorities, highlighting the need for clear methods, procedures, and adequate reporting to ensure consistent and fair administration.
- Current STA practice: entity-by-entity risk assessments can cause duplication and fragmentation.

### Large taxpayer population, segmentation, and data sources
- Classification criteria:
  - Qualitative criteria examples: whether the enterprise is a state-owned enterprise of central government; whether it is listed among the Fortune 500.
  - Quantitative criteria: six indicators — business revenue, annual tax payment, asset scale, industry influence, business region, and organizational structure; a comprehensive score based on these indicators and weighting ratios determines inclusion.
- Large taxpayer composition includes state-owned enterprises (SOEs), privately-owned companies, foreign-owned entities, and a significant number of multinational enterprises (MNEs).
- Sector composition (China’s Fortune 500 Industries):
  - Energy - 20%
  - Telecommunications - 15%
  - Banking and Financial Services - 13%
  - Real estate - 10%
  - Automotive - 9%
  - Retail - 8%
  - Manufacturing - 7%
  - Construction - 6%
  - Technology - 5%
  - Pharmacueticals and Healthcare - 4%
  - Others - 3%
- Recommended external data sources to enhance population understanding and cross-checking:
  - Tax return information and forms detailing cross-border related party transactions.
  - State registry offices for corporate ownership information.
  - Department of statistics.
  - Financial intelligence unit tracking cross-border flows.
  - Studies from institutions (universities, media).
- Recommended development of industry profiles including: Name of industry; Industry definition; Industry characteristics; Nature of the competition; Industry skill levels; Industry key players; Industry business models; Industry association; Industry profit margins and cost structures; Industry regulators; Industry financial performance (3 years); Industry tax performance (3 years, four compliance pillars for CIT and VAT); Tax issues, key risks, hot spots.

### Risk assessment frameworks and recommended approach
- Two-tiered approach (common international practice):
  - First-tier: screening all taxpayers to identify tax compliance risks, convert identified risks into risk filters/indicators, apply to large taxpayer data, and rank taxpayers by risk level.
  - Second-tier: in-depth risk assessment of selected complex, high-risk taxpayers using complete corporate structures, financial analyses, third-party data, and internal specialist workshops to identify priority risk areas for audit focus.
- Recommended shift to holistic, group-level risk management focusing on major events (equity transfer, mergers and acquisitions, related-party transactions) and employing business-chain and structural penetration analysis.
- Robust risk assessment should evaluate both likelihood and consequence to determine overall risk level (conceptual risk rating matrix with categories Unlikely, Possible, Likely producing High/Moderate/Low risk designations).
- Appendix III referenced as providing a list of typical tax risks for large taxpayers with topic descriptions and summaries for tax officials.

### Transition to digital risk management — core pillars and systems
- Pathway to digital risk management broken into three core pillars: systems integration, data governance, and AI.
- Systems integration:
  - Integrated Tax Administration System (ITAS) modernizes tax administration across registration, assessment, collection, enforcement and supports analytics and decision-making.
  - ITAS core segments include e-service, IT, Registration, External integration & reporting, Workload management, Accounting Processing, Headquarters services.
  - Effective integration requires simultaneous communication of over 100 discrete services (hardware and software).
  - Centralized nationwide access portal with individual dashboards connecting to APIs and analytics systems recommended to streamline analysis across diverse data sources and promote end-to-end institutional planning, modelling, and statistical learning.
  - Individual dashboards enable visualization, business intelligence, social network analysis, and integration of external reporting and 3rd party data into a single access point for a 360-degree view of taxpayers.
  - Advanced economies frequently use commercial-off-the-shelf (COTS) solutions for single access points and dashboards.
- Data governance and data pipelines:
  - Architecture depends on a cohesive data governance standard and a data pipeline for integration, processing, and storage into a data warehouse.
  - Data harmonization (data integration) brings together data from different sources into a cohesive data model to ensure correctness of statistical inferences.
  - A unique data format, generally JSON or XML, helps ensure data harmonization and structuration into a data warehouse and significantly improves AI and analytical tool application.
  - Standard model for data pipelines: gather data from tax returns, financial institutions, external databases; clean and transform data; store in centralized databases/data warehouses; apply advanced analytics (data mining, machine learning); make processed data accessible via dashboards and reporting tools.
- Artificial intelligence:
  - AI and statistical learning are core components for advanced analytics and real-time detection within integrated architectures (examples include crawlers, scrapers, GenAI, social network analysis, metrics builders, data matching, info management).
  - AI enhances detection, modelling, and risk prioritization when supported by harmonized data and integrated ITAS components.

### AI functions, usage, and comparative findings
- Six primary functions of AI systems essential to the CRM process:
  - e-service
  - data collection
  - risk detection
  - risk management (predictive and prescriptive)
  - nudging
  - internal assistance
- Example system types by CRM phase:
  - E-Service: Chatbot, Adaptive Communication, Generative AI.
  - Data Collection: Webscraping, Crawling, Machine Vision, Data Matching (Real-time).
  - Risk Detection: Outlier Detection, Data Matching, Visualization.
  - Risk Management: Predictive Analytics, Prescriptive Analysis.
  - Nudging: ‘Friendly Alert’, Adaptive Communication.
  - Internal Assistance: Information Management, Dashboards, Visualization.
- Quantitative utilization shares for assessment and assistance functions in advanced economies:
  - data collection: 20 percent
  - risk detection: 25 percent
  - risk management: 20 percent
  - chatbots: 10 percent
  - internal assistance: 20 percent
  - nudging techniques: 5 percent
- Comparative findings:
  - STA strengths: extensive e-service capabilities—on par with or exceeding current state in advanced economies.
  - STA gaps: not yet integrated AI systems performing web crawling, prescriptive analysis tools, machine vision, and real-time data matching.
  - Supervised learning classifiers:
    - In the European Union, supervised learning classifiers account for nearly 20 percent of all observed AI systems.
    - Among OECD members, over 90 percent report using these classifiers and business intelligence tools to categorize taxpayers by risk.
  - STA has chosen to outsource web scraping to an external service provider, differing from advanced economies that typically use open access systems and adaptable scraping/crawling frameworks.

### Organizational design, governance, and accountability
- Headquarters (HQ) role and principles for high performance:
  - (i) a strong HQ;
  - (ii) a function-based organization design;
  - (iii) minimal management layers;
  - (iv) streamlined field offices;
  - (v) an organization that reflects key taxpayer segments.
- HQ responsibilities (high-level role) include: developing procedures and methodology; preparing manuals and instructions; determining business processes and IT requirements; developing risk management strategies; providing technical advice to operational units; allocating resources; developing operational plans and national programs; distributing revenue targets; setting performance measures; establishing reporting requirements; monitoring operational performance.
- Organizational challenges:
  - STA does not have a single locus for HQ functions; 17 functional departments are grouped as HQ with a hybrid of tax-type and function-based departments.
  - TBDRM handles risk management and data analytics but is not in the HQ grouping.
  - Field offices deliver operations per HQ policies; a symbiotic HQ–field office relationship is critical for consistency and credibility.
  - Some positions have two lines of accountability: local management and HQ functional departments (e.g., audit teams).
- Accountability and performance management:
  - Accountability defines which division is responsible for delivering parts of STA’s mandate; CRM success requires clear roles across departments and well-coordinated action.
  - A performance management framework identifying key indicators and linking them to strategic and operational plans is integral.
  - The 2023 FAD report provided advice on enhancing performance management for corporate income tax administration.

### CRM department mandate, governance, and operations
- CRM mandate and core activities:
  - Typically an HQ function, possibly located in a planning/programming or a risk management division.
  - Key activities:
    - (1) analyzing internal and external environments;
    - (2) analyzing the current tax administration data for indications of non-compliance;
    - (3) collaborating closely with risk owners, who are the business leaders for the various tax functions e.g. audit;
    - (4) working collaboratively with risk owners to develop treatment strategies for identified risks;
    - (5) developing CRM management methodologies and practices and documenting them for future reference.
  - CRM will require new skillsets related to data analytics, pattern recognition and early identification of outliers.
- CRM focal points and center of excellence:
  - CRM focal points should be named in all key departments.
  - As CRM develops, involved departments must be formalized and coordinated so CRM products and direction are integrated into daily operational work across the STA.
  - Naming focal points beginning with key departments such as audit will help create a culture of CRM and centers of CRM excellence across the STA.
- Resources and short-term actions:
  - Resources devoted to CRM should be reviewed with a view to possible reallocation.
  - Assess existing CRM development across STA to determine new work and resource needs; resource reallocation may be possible, but additional resources may be critical.
  - Short-term analysis of resource needs recommended.
- Senior CRM committee:
  - Establish a senior CRM committee to bring all intelligence into one place and ensure senior leadership authority over CRM direction and resources.
  - Composition: usually chaired by the most senior executive and include senior leadership/risk owners across STA functions; staffed by the CRM department with support from CRM focal points.
  - Meeting frequency and mandate:
    - The committee would usually meet annually (and perhaps more frequently early on).
    - Mandate includes:
      - (1) evaluate progress in addressing existing risks;
      - (2) confirm whether existing risks are still the highest priority risks for treatment;
      - (3) identify any emerging risks to the tax system;
      - (4) approve – on an annual basis – the CIPs for the upcoming year, including resource allocations aligned to the treatments identified;
      - (5) provide guidance to the CRM department as needed;
      - (6) review the performance results of the various treatment strategies and adjust as required.
    - The committee would oversee the compliance risk register (CRR) that is a historical record of the risks and approved strategies to address the risks.

### Technical guidance on data quality, predictive accuracy, and ML maturity
- Data quality dimensions and definitions:
  - Data accuracy: “the degree to which information is correct, unambiguous, meaningful, believable, and consistent”.
    - Error rates formula: error_rates = (number_of _erred_fields)/(number_of _total_fields)
  - Data representativity: “the degree to which the sample accurately corresponds to the phenomenon studied or the target population measured”.
    - Key biases: representation bias; training-deployment skew or ‘data drift’.
  - Data currency: “the degree to which information is up to date, or the degree to which the information precisely reflects the current state of the world that it represents”.
    - Data currency is inversely correlated to data drift and training-deployment skew.
- Monitoring and constraints:
  - Field audit is reserved to a selected sample of taxpayers (CRM), limiting capacity to verify reference values for all taxpayers or entire data lakes.
  - Open-source monitoring tools (example: EvidentlyAI) perform recurrent audits of data quality within data pipelines.
  - Recurrent monitoring improves predictive and prescriptive AI accuracy and signals when models should be updated.
- Proxy checks and stability metrics:
  - Feature range compliance for anomaly detection (example: low-value goods generating exceptionally high revenue; a car reported as weighing one kilogram).
  - Characteristic Stability Index (CSI) and Population Stability Index (PSI):
    - PSI formula: Σ (Actual% –   Expected%) *Ln (Actual%/Expected%)
    - CSI formula: Σ (Actual% –   Expected%) *feature predictive score
    - Interpretation:
      - PSI and/or CSI value below 0.1 or ranging from 0.1 to 0.2 indicates no major changes or moderate changes.
      - PSI and/or CSI value above 0.2 should alert model supervisors to data or model drift and indicate retraining and dataset quality assessment.
- Classifier performance and thresholding:
  - Common metrics: Receiver Operating Characteristic (ROC); Area Under the Curve (AUC).
  - Definitions: Sensitivity and Specificity.
  - Threshold trade-offs: lower threshold increases sensitivity and false positives; higher threshold increases specificity and fewer false positives.
  - Examples:
    - Threshold set at 1 selects taxpayers whose likelihood of non-compliance is 100%.
    - Threshold set at 0.6 generates more false positives, increasing audit selection and administrative burden.
  - AUC examples and interpretation:
    - AUC calculation method: Rule Numerical Integration method: (x_i+1 – x_i) * ( y_i + y_i+1 ) / 2
    - Results obtained: Model A = 0.91, Model B = 0.794, Model C = 0.59, with the optimal AUC = 1.
- Real-time data matching and machine learning maturity:
  - VAT refund fraud context:
    - VAT refund fraud in the EU is estimated to account for an estimated 80 to 120 billion in revenue losses every year.
    - 2019 EU implementation of Transaction Network Analysis (TNA) connected tax IT platforms of EU Member States for real-time transactional processing when VAT returns are filed.
    - EU Fiscalis network reported a €12 to 15 billion increase in revenue collected annually following implementation.
  - Three maturity steps for real-time data-matching or predictions:
    1. Batch Prediction — predictions computed in batch at intervals; sufficient for annual plans and segmentation.
    2. Session-based Predictions — online data collected in real-time with features computed after receipt (near real-time); reduces inference latency to seconds.
    3. Online Predictions — feature computation at moment of prediction request; requires continual self-learning capability and carries higher risk of feedback loops.
  - Current status: no tax administration has implemented a true online prediction model; data is matched at interval rates of seconds which is sufficient for CRM detection of non-compliance and anomalies.

### Explicit recommendations (preserve wording and numeric precision)
- Pathway to Digital Risk Management:
  - II.1 The STA should shift from an entity-by-entity approach to risk assessment to a more comprehensive aggregate perspective that targets the highest risk cases within the large taxpayer population.
  - II.2 Establish a framework using ITAS and CRM principles that anticipates the data and technology needs necessary for effective risk management. This framework should incorporate strategies for identifying emerging data sources, leveraging advanced analytical tools and innovative technologies.
- Governance and Accountability:
  - III.1 Designate LETD as the primary functional department at headquarters responsible for coordinating all large business programs and supporting functions from a central focal point.
  - III.2 Establish a dedicated CRM unit with a senior level committee to lead the STA's CRM strategy and governance framework.
- Additional organizational recommendations (implied across governance section):
  - Consider mandating large business management to LETD (or designate a lead functional department) working closely with VAT, income tax, audit, taxpayer service.
  - Review how large business population priorities are set and managed within TBDRM.
  - Assign a dedicated organizational focus for CRM (to an existing department such as TBDRM or a new department/division) to coordinate CRM across STA departments.
- Operational and implementation considerations:
  - Implement comprehensive system integration (nationwide portal, individual dashboards, APIs).
  - Establish strong data governance (data harmonization, JSON/XML formats).
  - Build robust data pipelines feeding a data warehouse for analytics and AI applications.
  - Develop industry profiles and holistic group-level risk assessments.
  - Establish clear KPIs and performance management to align staff with strategic goals.

*This report is the final version of the aide-mémoire presented to STA officials on November 15, 2024 (IMF Technical Assistance Report).*

### PREFACE __________________________________________________________________________ 6

### PREFACE

### Mission, scope, and timeline
- IMF Fiscal Affairs Department capacity development mission visited Yangzhou and Beijing, China, from November 4 to 15, 2024, to provide guidance on advancing risk management for large taxpayers.
- First week: seven-day workshop at the Yangzhou Training Academy delivering comprehensive training and sharing presentation materials in advance.
- Second week: technical discussions in Beijing and submission of the draft report; final aide-mémoire presented to officials on November 15, 2024.
- Mission team led by Ms. Cindy Negus (Senior Economist, FAD) and included Messrs. David Hadwick, Cameron Smith, and Ms. Maureen Kidd (external experts).
- Key STA interlocutors met by the team include: Mr. Bu Xianglai; Ms. Meng Yuying; Ms. Li Na; Mr. Yue Song; Mr. Wang Dan; Mr. Li Bin; Mr. Du Jianwei; Ms. Yang Xiao; Ms. Liu Yingjie; and other central and provincial taxation officials.
- Acknowledgement of support from Mr. Liu Xin, Office of Academic Affairs of the National Tax Academy, Yangzhou Campus.

### Executive summary — strategic thrusts and analysis
- Modernization: STA has modernized operations with digital solutions and advanced analytics to strengthen risk assessment capabilities.
- Objective: Use data analytics to improve compliance, reduce compliance costs, and strengthen relationships with the large business tax community.
- Emphasis areas covered during the mission: organizational governance, application of artificial intelligence (AI) and big data, Compliance Risk Management (CRM), risk assessment approaches, strategic planning, treatment strategies, Compliance Improvement Plans (CIP), and capability building for large taxpayer management.
- Two priority areas for recommendations:
  - Enhancing Risk Management: move from an entity-by-entity approach to an aggregate population-level approach to identify and prioritize the highest risk cases; establish a proactive framework anticipating data and technology needs; identify emerging data sources and leverage advanced analytical tools.
  - Strengthening Governance and Accountability: designate the Large Enterprise Tax Administration Department (LETD) as the primary functional department at headquarters to centralize coordination; establish a dedicated CRM unit and a senior-level committee for CRM strategy and governance.

### Key statistics and facts (preserve numeric precision)
- STA total tax revenues collected in 2023: 17.7 trillion RMB.
- 2023 year-on-year increase: 10.1 percent.
- Effective increase after adjustments for excess input VAT credit refund policy: 2.3 percent.
- Domestic VAT increase from 2022 to 2023: 42.2 percent.
- Corporate (enterprise) income tax (CIT) change 2022–2023: decreased by 5.9 percent.
- Large Taxpayer segment contribution to total revenue: 45.70 percent.
- USD equivalent noted in source: Approximately 2.478 trillion USD.
- STA estimates close to 10,000 staff engaged in large taxpayer business management across the country.
- LETD HQ staff: about 40 staff.

### Current large taxpayer administration model and challenges
- LETD created in 2008 to oversee large taxpayer management; hybrid role combining headquarters and operational responsibilities; LETD does not include debt management functions.
- LETD coordinates with other functional departments (e.g., VAT Department, Income Tax Department) and is supported by TBDRM (Tax Big Data and Risk Management Bureau) with responsibility for the risk management model.
- Annual LETD process: formulate an annual risk management plan, consult in advance with Big Data and Risk Administration Bureau and Tax Inspection Department, then send audit referrals to the Audit Department.
- Large taxpayers are distributed nationally and managed by provincial and lower tax authorities, highlighting the need for clear methods, procedures, and adequate reporting to ensure consistent and fair administration.

### Recommendations (as presented in the report)
- Pathway to Digital Risk Management
  - II.1 The STA should shift from an entity-by-entity approach to risk assessment to a more comprehensive aggregate perspective that targets the highest risk cases within the large taxpayer population.
  - II.2 Establish a framework using ITAS and CRM principles that anticipates the data and technology needs necessary for effective risk management. This framework should incorporate strategies for identifying emerging data sources, leveraging advanced analytical tools and innovative technologies.
- Governance and Accountability
  - III.1 Designate LETD as the primary functional department at headquarters responsible for coordinating all large business programs and supporting functions from a central focal point.
  - III.2 Establish a dedicated CRM unit with a senior level committee to lead the STA's CRM strategy and governance framework.

_This report is the final version of the aide-mémoire presented to STA officials on November 15, 2024 (IMF Technical Assistance Report)._

### 8.      There are 2,000 large taxpayers managed through large taxpayer management, with LETD

### There are 2,000 large taxpayers managed through large taxpayer management, with LETD oversight.

### Large taxpayer population and oversight
- There are 2,000 large taxpayers managed through large taxpayer management, with LETD oversight.
- Among them, 140 are considered particularly important, and are jointly managed by the General Taxation Department of the State Taxation Administration and the provincial and lower large taxpayer management departments (headquarters-to-headquarters management).
- Of these 2000 taxpayers, the top 1000 generate about 40 percent of STA revenue while the entire 2000 generates over 45 percent 910.
- The large taxpayer population includes state-owned enterprises (SOEs), privately-owned companies, foreign-owned entities, and a significant number of multinational enterprises (MNEs).

### Mission scope, strategic recommendations, and organizational context
- The 2023 IMF Technical Assistance Report, Selected Issues in Administering Corporate Income Tax recommended enhancements for organizational effectiveness and improved CRM 11.
- The report found the STA has established a clear strategic plan with strong outcomes and recommended:
  - Develop a comprehensive strategic management framework incorporating operational work plans informed by risk assessments, resource availability, and emerging issues.
  - Strengthen governance processes and organizational support, including a strong headquarters function to set policy direction and ensure adequate functional support.
  - Implement a structured framework for CRM to systematically identify, assess, rank, and address tax compliance risks to improve decision-making and revenue outcomes.
  - Consider new compliance strategies as the number of large taxpayers continues to rise.
  - Establish a performance management framework with clearly defined key performance indicators (KPI)s to facilitate progress assessment and alignment with strategic goals and objectives.
- This report focuses on enhancing risk management for large taxpayers and transitioning to digital risk management employing artificial intelligence and big data (Section II), and bolstering governance and accountability (Section III). Outcomes and recommendations were developed via the Yangzhou workshop and consultations with LETD officials from STA HQ.

### Defining the large taxpayer population and data sources
- STA uses both qualitative and quantitative criteria to classify large enterprise taxpayers:
  - Qualitative criteria: e.g., whether the enterprise is a state-owned enterprise of central government and whether it is listed among the Fortune 500.
  - Quantitative criteria: six indicators — business revenue, annual tax payment, asset scale, industry influence, business region, and organizational structure; a comprehensive score based on these indicators and weighting ratios determines inclusion.
- Additional external data sources recommended to enhance understanding and cross-checking of the large taxpayer population:
  - Tax return information and forms detailing cross-border related party transactions.
  - State registry offices for corporate ownership information.
  - Department of statistics.
  - Financial intelligence unit tracking cross-border flows.
  - Studies from institutions (universities, media).

### Importance of sectoral segmentation and industry profiling
- Taxpayers in the large taxpayer segment are categorized by industry sectors to identify largest industries and assess industry-specific risks.
- Figure (China’s Fortune 500 Industries) composition by sector (as stated):
  - Energy - 20%
  - Telecommunications - 15%
  - Banking and Financial Services - 13%
  - Real estate - 10%
  - Automotive - 9%
  - Retail - 8%
  - Manufacturing - 7%
  - Construction - 6%
  - Technology - 5%
  - Pharmacueticals and Healthcare - 4%
  - Others - 3%
- Recommended development of industry profiles to document industry determinants and enhance understanding. Box 1 lists elements typically contained in industry profiles, including:
  - Name of industry; Industry definition; Industry characteristics; Nature of the competition; Industry skill levels; Industry key players; Industry business models; Industry association; Industry profit margins and cost structures; Industry regulators; Industry financial performance (3 years); Industry tax performance (3 years, four compliance pillars for CIT and VAT); Tax issues, key risks, hot spots.

### Risk assessment processes and frameworks
- Current STA practice: entity-by-entity risk assessments can cause duplication and fragmentation.
- Recommended shift to holistic, group-level risk management focusing on major events (equity transfer, mergers and acquisitions, related-party transactions) and employing business-chain and structural penetration analysis.
- Two-tiered approach (common international practice):
  - First-tier: screening all taxpayers to identify tax compliance risks, convert identified risks into risk filters/indicators, apply to large taxpayer data, and rank taxpayers by risk level.
  - Second-tier: in-depth risk assessment of selected complex, high-risk taxpayers using complete corporate structures, financial analyses, third-party data, and internal specialist workshops to identify priority risk areas for audit focus.
- Robust risk assessment should evaluate both likelihood and consequence to determine overall risk level. Conceptual risk rating matrix presented with categories: Unlikely, Possible, Likely across consequence levels producing High/Moderate/Low risk designations (Figure 3).

### Transition to digital risk management — core pillars
- The pathway to digital risk management is broken into three core pillars: systems integration, data governance, and AI.
- CRM processes have traditionally been manual; digital technologies can enhance efficiency, improve data accuracy, and enable real-time risk assessments.

Systems integration
- Integrated Tax Administration System (ITAS) modernizes tax administration across registration, assessment, collection, enforcement and supports analytics and decision-making.
- ITAS core segments include e-service, IT, Registration, External integration & reporting, Workload management, Accounting Processing, Headquarters services.
- Effective integration requires simultaneous communication of over 100 discrete services (hardware and software).
- Centralized nationwide access portal with individual dashboards connecting to APIs and analytics systems is recommended to streamline analysis across diverse data sources and promote end-to-end institutional planning, modelling, and statistical learning.
- Individual dashboards provide protocols for software applications to exchange data and features; they enable visualization, business intelligence, social network analysis, and integration of external reporting and 3rd party data into a single access point for a 360-degree view of taxpayers.
- Advanced economies frequently use commercial-off-the-shelf (COTS) solutions for single access points and dashboards.

Data governance and data pipelines
- The architecture depends on a cohesive data governance standard and a data pipeline for integration, processing, and storage into a data warehouse.
- Data harmonization (data integration) brings together data from different sources into a cohesive data model to ensure correctness of statistical inferences.
- A unique data format, generally JSON or XML, helps ensure data harmonization and structuration into a data warehouse and significantly improves AI and analytical tool application.
- Standard model for data pipelines: gather data from tax returns, financial institutions, external databases; clean and transform data; store in centralized databases/data warehouses; apply advanced analytics (data mining, machine learning); make processed data accessible via dashboards and reporting tools.

Artificial intelligence
- Artificial intelligence and statistical learning are noted as core components for advanced analytics and real-time detection within integrated architectures (Figure examples include crawlers, scrapers, GenAI, social network analysis, metrics builders, data matching, info management).
- AI enhances detection, modelling, and risk prioritization when supported by harmonized data and integrated ITAS components.

### Key implementation considerations and outcomes
- Implementing digital risk management requires:
  - Comprehensive system integration (nationwide portal, individual dashboards, APIs).
  - Strong data governance (data harmonization, JSON/XML formats).
  - A robust data pipeline feeding a data warehouse for analytics and AI applications.
  - Development of industry profiles and holistic group-level risk assessments.
  - Clear KPIs and performance management to align staff with strategic goals.
- Appendix III (referenced) provides a list of typical tax risks for large taxpayers with topic descriptions and summaries for tax officials.

*IMF Technical Assistance Report*

### 33.        Currently, AI systems in advanced economies serve six primary functions within tax

### Currently, AI systems in advanced economies serve six primary functions within tax administrations

### Functional roles and taxonomy of AI in tax administrations
- Six primary functions of AI systems essential to the CRM process:
  - e-service: systems meant to aid taxpayers at their volition.
  - data collection: systems meant to collect taxpayer online data.
  - risk detection: systems meant to collect non-compliance signals.
  - risk management: systems meant to segment taxpayers in risk categories (predictive) and systems meant to suggest follow-up actions (prescriptive).
  - nudging: systems meant to aid taxpayers, at the administrations’ volition.
  - internal assistance: systems meant to aid tax officials, at their own volition.
- Example system types mapped to CRM phases (pre-assessment, assessment, post-assessment):
  - E-Service: Chatbot, Adaptive Communication, Generative AI.
  - Data Collection: Webscraping, Crawling, Machine Vision, Data Matching (Real-time).
  - Risk Detection: Outlier Detection, Data Matching, Visualization.
  - Risk Management: Predictive Analytics, Prescriptive Analysis.
  - Nudging: ‘Friendly Alert’, Adaptive Communication.
  - Internal Assistance: Information Management, Dashboards, Visualization.
- Integration note: Figures indicate which systems are integrated by the STA and which are not integrated by the STA.

### Quantitative usage of AI functions in advanced economies (assessment and assistance)
- Current utilization shares for tax administrations (assessment and assistance functions):
  - data collection: 20 percent
  - risk detection: 25 percent
  - risk management: 20 percent
  - chatbots: 10 percent
  - internal assistance: 20 percent
  - nudging techniques: 5 percent
- Observed patterns:
  - Tax administrations primarily utilize AI for assessment functions (data collection, risk detection, risk management).
  - AI is also used for internal and external assistance in pre- and post-assessment phases (chatbots, internal assistance, nudging).

### Comparative findings: STA versus advanced economies / OECD
- STA strengths:
  - STA has extensive e-service capabilities—on par with or exceeding current state in advanced economies.
- STA gaps relative to advanced economies and OECD:
  - STA has not yet integrated AI systems performing key CRM functions: web crawling, prescriptive analysis tools, machine vision, and real-time data matching.
  - These functions have been underpinned by AI in OECD and advanced economies for almost a decade.
- Supervised learning classifiers:
  - STA officials noted lower utilization of supervised learning classifiers compared to advanced economies.
  - Role of classifiers: trained on large historical datasets to predict likelihood of non-compliance, process a larger variety of features, help devise risk profiles, inform annual planning, and develop diversified treatment strategies.
  - Comparative metrics:
    - In the European Union, supervised learning classifiers account for nearly 20 percent of all observed AI systems.
    - Among OECD members, over 90 percent report using these classifiers and business intelligence tools to categorize taxpayers by risk.
- Outsourcing practice:
  - STA has chosen to outsource web scraping to an external service provider, differing from advanced economies that typically use open access systems and adaptable scraping/crawling frameworks (e.g., Python frameworks such as Django, Scrapy).

### Organizational scale and governance implications
- STA network and scale:
  - STA mandate delivered in close to 4000 locations (600 manage large taxpayers) as well as 41,000 township level divisions.
- Governance and accountability priorities as STA transitions to digital CRM:
  - Need for robust governance framework as operations grow in complexity with digital CRM systems.
  - Coordinated action across functional departments at STA headquarters is essential to establish clear roles, responsibilities, and processes.
  - Effective governance ensures alignment with strategic objectives, transparent oversight, data integrity, and security to protect sensitive taxpayer information.
- Headquarters (HQ) role and principles:
  - Key organizational design principles for high performance:
    - (i) a strong HQ;
    - (ii) a function-based organization design;
    - (iii) minimal management layers;
    - (iv) streamlined field offices;
    - (v) an organization that reflects key taxpayer segments.
  - HQ responsibilities (high-level role):
    - Developing procedures and methodology.
    - Preparing manuals and instructions.
    - Determining business processes and IT requirements.
    - Developing risk management strategies.
    - Providing technical advice to operational units.
    - Allocating resources and distribution in operational offices.
    - Developing operational plans and national programs.
    - Distributing revenue targets.
    - Setting performance measures.
    - Establishing reporting requirements.
    - Monitoring operational performance.
- Organizational challenges:
  - STA does not have a single locus for HQ functions; 17 functional departments are grouped as HQ with a hybrid of tax-type and function-based departments.
  - TBDRM handles risk management and data analytics but is not in the HQ grouping.
  - Field offices deliver operations per HQ policies; a symbiotic HQ–field office relationship is critical for consistency and credibility.
  - Some positions have two lines of accountability: local management and HQ functional departments (e.g., audit teams).

### Accountability and performance management
- Accountability:
  - Accountability defines which division is responsible for delivering parts of STA’s mandate; CRM success requires clear roles across departments and well-coordinated action.
  - Leadership and senior managers are essential to define collective action, execute strategic plans, and renew the organization.
- Performance management:
  - A performance management framework identifying key indicators and linking them to strategic and operational plans is integral.
  - Individual clarity on roles and expectations is necessary because individual efforts aggregate to overall STA performance.
  - The 2023 FAD report provided advice on enhancing performance management for corporate income tax administration.

### Recommendations (explicit)
- Shift from entity-by-entity risk assessment to a more comprehensive, aggregate perspective targeting highest risk cases within the large taxpayer population.
- Establish a framework using ITAS and CRM principles to anticipate data and technology needs necessary for effective risk management; incorporate strategies for identifying emerging data sources, leveraging advanced analytical tools and innovative technologies.
- Organizational recommendations (implied across governance section):
  - Consider mandating large business management to LETD (or designate a lead functional department) working closely with VAT, income tax, audit, taxpayer service.
  - Review how large business population priorities are set and managed within TBDRM.
  - Assign a dedicated organizational focus for CRM (to an existing department such as TBDRM or a new department/division) to coordinate CRM across STA departments.

*IMF Technical Assistance Report*

### 56.        The mandate of a CRM department will be one of leadership and coordination. Typically, it

### The mandate of a CRM department will be one of leadership and coordination.

### CRM mandate and core activities
- Typically, it would be an HQ function and could be located in a planning/programming or a risk management division.
- Key activities include:
  - (1) analyzing internal and external environments;
  - (2) analyzing the current tax administration data for indications of non-compliance;
  - (3) collaborating closely with risk owners, who are the business leaders for the various tax functions e.g. audit;
  - (4) working collaboratively with risk owners to develop treatment strategies for identified risks; and
  - (5) developing CRM management methodologies and practices and documenting them for future reference.
- CRM will require new skillsets related to data analytics, pattern recognition and early identification of outliers.

### CRM focal points, culture, and center of excellence
- CRM focal points should be named in all key departments.
- HQ functions are found in many STA functional departments and in the TBDRM.
- As the CRM framework develops and gains traction, all of these departments will make some contribution to CRM; this involvement needs to be formalized and coordinated so CRM has needed inputs and CRM products and direction are integrated into daily operational work across the STA.
- Business knowledge about existing compliance risks is important; staff with experience in auditing will likely be aware of patterns of non-compliance and service officers will know trends in questions that could direct CRM treatments in education and awareness.
- CRM can with time become a center of excellence.
  - Naming focal points, beginning with key departments such as audit and eventually extending to any department with some CRM interest, will help create a culture of CRM and centers of CRM excellence across the STA.
  - Given the size of the STA and its office network, this would be an especially important achievement.

### Resources and short-term actions
- Resources devoted to CRM should be reviewed with a view to possible reallocation.
- Some CRM is under development across the STA; this should be reviewed and assessed to understand what new work will be needed to properly undertake CRM.
- Building an awareness of risk and identifying priorities and treatments will in many cases go beyond current working methods in place in STA and will need a resource response.
- Resource reallocation may be possible if some areas of work are no longer needed but there may be many cases where additional resources are critical to success.
- This analysis should be undertaken in the short term.

### Senior CRM committee: governance, membership, and staffing
- A senior CRM committee should be established to support management awareness and decision-making.
  - This would bring all intelligence into one place and would mean that senior leadership is the STA’s authority over CRM direction and resources.
  - The committee would ensure close linkage to the STA’s overarching strategic priorities and goals.
- Composition and staffing:
  - Such a committee would usually be chaired by the most senior executive and its membership would include the senior leadership i.e. risk owners from across the STA who are responsible for key tax administration functions.
  - The work of the committee would be staffed by the CRM department with support from CRM focal points.

### Committee meeting frequency and mandate
- The CRM committee would meet regularly and have a clear mandate.
  - The committee would usually meet annually (and perhaps more frequently in the early months after its creation).
  - Its mandate would include:
    - (1) evaluate progress in addressing existing risks:
    - (2) confirm whether existing risks are still the highest priority risks for treatment;
    - (3) identify any emerging risks to the tax system;
    - (4) approve – on an annual basis – the CIPs for the upcoming year, including resource allocations aligned to the treatments identified;
    - (5) provide guidance to the CRM department as needed; and
    - (6) review the performance results of the various treatment strategies and adjust as required.
  - The committee would oversee the compliance risk register (CRR) that is a historical record of the risks and approved strategies to address the risks.

### Recommendations (as listed)
- Designate LETD as the primary functional department at headquarters responsible for coordinating all large business programs and supporting functions from a central focal point.
- Establish a dedicated CRM unit with a senior level committee to lead the STA's CRM strategy and governance framework.

*IMF Technical Assistance Report | Excerpt*

### Appendix IV.  Technical Guidance on Enhancing

### Appendix IV. Technical Guidance on Enhancing Data and Predictive Accuracy

### Data quality dimensions and definitions
- Data accuracy: “the degree to which information is correct, unambiguous, meaningful, believable, and consistent”.
  - Error rates formula: error_rates = (number_of _erred_fields)/(number_of _total_fields)
- Data representativity: “the degree to which the sample accurately corresponds to the phenomenon studied or the target population measured”.
  - Key biases of concern: representation bias; training-deployment skew or ‘data drift’.
- Data currency: “the degree to which information is up to date, or the degree to which the information precisely reflects the current state of the world that it represents”.
  - Data currency is inversely correlated to data drift and training-deployment skew.

### Practical constraints and monitoring approaches for tax administrations
- Constraints:
  - Field audit is reserved to a selected sample of taxpayers (CRM), limiting capacity to verify reference values for all taxpayers or entire data lakes.
  - Some scientific solutions are impractical; open-source monitoring tools (example: EvidentlyAI) perform recurrent audits of data quality within data pipelines.
- Recurrent monitoring benefits:
  - Improves accuracy of predictive and prescriptive AI systems.
  - Signals when a statistical model should be updated.
  - Can be embedded into standard statistical procedural requirements to monitor data quality and verify predictive models or individual features.

### Proxy values and automated checks when ground truth labels are unavailable
- Feature range compliance:
  - Tracks whether the value of a label remains within a minimum-maximum range.
  - Augmentable with logical inferences and expert knowledge.
  - Example use: detect anomalous reported values (e.g., low-value goods generating exceptionally high revenue, or a car reported as weighing one kilogram) in real-time during or prior to reporting.
- Characteristic Stability Index (CSI) and Population Stability Index (PSI):
  - Purpose: assess consistency/reliability of predictors over time and quantify model drift / deployment-skew.
  - PSI formula: Σ (Actual% –   Expected%) *Ln (Actual%/Expected%)
  - CSI formula: Σ (Actual% –   Expected%) *feature predictive score
  - Interpretation:
    - PSI and/or CSI value below 0.1 or ranging from 0.1 to 0.2 indicates no major changes or moderate changes.
    - PSI and/or CSI value above 0.2 should alert statistical model supervisors or process owners to data or model drift and indicate that the model should be retrained and dataset quality assessed.

### Monitoring predictive value and classifier accuracy
- Common metrics:
  - Receiver Operating Characteristic (ROC)
  - Area Under the Curve (AUC)
- Key definitions:
  - Sensitivity: The probability that the model predicts a positive outcome for an observation when the outcome is positive.
  - Specificity: The probability that the model predicts a negative outcome for an observation when the outcome is negative.
- ROC and thresholding for audit selection:
  - ROC plots true positive rate (TP) against false positive rate (FP) for various threshold values.
  - Threshold trade-offs:
    - Lower threshold minimizes likelihood of missing potential fraud cases (increases sensitivity, more false positives).
    - Higher threshold minimizes likelihood of auditing compliant taxpayers (increases specificity, fewer false positives).
  - Examples:
    - Threshold set at 1 selects taxpayers whose likelihood of non-compliance is 100%.
    - Threshold set at 0.6 generates more false positives, increasing the number of taxpayers selected for audit and potentially increasing administrative burden.
- AUC calculation example:
  - Rule Numerical Integration method: (x_i+1 – x_i) * ( y_i + y_i+1 ) / 2
  - Results obtained: Model A = 0.91, Model B = 0.794, Model C = 0.59, with the optimal AUC = 1.
  - Interpretation: A higher AUC indicates better overall classification performance; AUC corresponds to the probability the model can correctly classify positive and negative instances.

### Managing real-time data matching and machine learning maturity stages
- VAT refund fraud context:
  - VAT refund fraud in the EU is estimated to account for an estimated 80 to 120 billion in revenue losses every year.
  - 2019 EU implementation of Transaction Network Analysis (TNA) connected tax IT platforms of EU Member States for real-time transactional processing when VAT returns are filed.
  - EU Fiscalis network reported a €12 to 15 billion increase in revenue collected annually following implementation.
- Real-time (online) machine learning:
  - Definition: using machine learning systems to process data as soon as it becomes available to permit instantaneous predictions.
  - Use case relevance: primarily useful for improved VAT refund fraud detection.
  - Examples of implementations: TNA (EU), systems in Belgium and Poland to verify correspondence of financial transactions.
- Three maturity steps to attain real-time data-matching or predictions:
  1. Batch Prediction
     - Predictions computed in batch at intervals; sufficient for annual plans, treatment strategies, and taxpayer segmentation.
  2. Session-based Predictions (also called ‘online prediction with batch features’)
     - Online data collected in real-time; features computed after receipt with a time interval (near real-time).
     - Requires thorough annotation, pre-computed algorithms, and streaming infrastructure (real-time data streaming transport and streaming computation engines, e.g., Structured Query Language (SQL)).
     - Reduces inference latency (unit of latency is seconds) and enables detection of more anomalies compared to pure batch.
     - Challenge: requires high-quality embeddings of pre-computed algorithms to ensure matching rate between self-reported data and administration data warehouses.
  3. Online Predictions
     - Feature computation using newly generated data at the moment of prediction request.
     - Requires continual self-learning capability combining pre-computed algorithms and self-learning AI systems.
     - Higher risk of feedback loops introducing bias.
     - Currently, few if any CRM use cases require online prediction; data matching at interval rates of seconds is generally sufficient for CRM purposes.
- Current status:
  - Despite claims, no tax administration has implemented a true online prediction model; data is matched at interval rates of seconds which is sufficient for CRM detection of non-compliance and anomalies.

*IMF Technical Assistance Report — Appendix IV.*

---


_Source: https://www.imf.org/-/media/files/publications/tar/2025/english/tarea2025054-print-pdf.pdf_