## tnmea2025013

## Source details

**Canonical URL:** [tnmea2025013](https://www.imf.org/-/media/files/publications/tnm/2025/english/tnmea2025013.pdf)

## Other formats

- [Markdown version](/-/media/files/publications/tnm/2025/english/tnmea2025013.pdf.md)
- [Structured JSON version](/-/media/files/publications/tnm/2025/english/tnmea2025013.pdf.json)

---

### Purpose and scope
- Describes generative AI’s (GenAI) ability to support compliance risk analysis in tax and customs administration.
- Written for a technical audience to raise awareness by explaining and demonstrating GenAI capabilities.
- Structure of the note:
  - Conceptual overview of GenAI technology.
  - Four generalized use cases where GenAI can augment the work of risk analysts.
  - Practical demonstrations (proofs of concept) showing how GenAI can be used to better understand taxpayer and trader compliance.
  - Basic guidelines to help administrations considering operational implementation.
  - Forward-looking statements on likely developments.
- Intentionally limited in scope to compliance risk analysis for demonstrating emerging but practical, high-value applications.

### Key observations on GenAI’s transformational potential
- Three capabilities with transformational potential for risk functions:
  - natural language research,
  - analysis,
  - analytics.
- GenAI enables direct interaction in natural language and can:
  - inform analysis of technical tax matters,
  - generate code for analytics,
  - interpret results.
- Risks and caveats: inaccuracy, biases, AI hallucination; potential for AI utility to eclipse human analysts if model size and sophistication grow rapidly.

### Technical and conceptual foundations
- GenAI: pattern-based systems producing humanlike content (text, video, images, audio, or data) using deep neural networks and learned relationships from training data.
- Historical building blocks referenced: neural networks (1980s), deep learning (2000s), Generative Adversarial Networks (2014), Transformer-based models (subsequent years).
- Compound systems and concepts included: RAG, RLHF, foundation models, LLMs, sequenced agents, autonomous agents.

### Executive summary and timeframe
- A 25% devaluation of Country X’s currency, precipitated by significant declines in key commodity prices, will present multifaceted challenges for the tax administration over the next five years.
- Cross-cutting considerations:
  - Timing and interaction of different tax effects.
  - Sector-specific impacts beyond commodities.
  - Secondary effects on the broader tax base.
  - Practical challenges for tax administration capacity and operations.
  - International tax implications (profit shifting, transfer pricing, withholding regimes).

### I. Impact on large multinational taxpayers — tax revenue issues and timing
- Immediate and medium-term corporate tax revenue effects:
  - Decreased Corporate Income Tax Revenue
    - Immediate Effect (Year 1–2):
      - Fall in commodity prices reduces MNC profitability → lower corporate income tax payments.
    - Medium Term (Year 3–4):
      - MNCs may delay or cancel expansion projects → slowdown in economic growth and further decline in corporate tax revenues.
  - FX-Related Issues (Immediate Tax Revenue Impacts Years 1–2):
    - Foreign exchange exposures and translation effects can alter reported local-currency profits for MNC subsidiaries or branches.
- Compliance risk dynamics and timing:
  - Short-run (Years 1–2):
    - Increased incentives for profit-shifting via royalties, marketing hubs, interest-bearing related-party debt, and other intercompany charges.
    - Immediate scrutiny areas: effective tax rate differentials, deferred tax asset movements, unremitted earnings, economic substance of marketing hubs.
  - Medium-run (Years 3–4):
    - Changes in transfer pricing patterns (increased leverage, altered royalty rates) → need enhanced transfer pricing and thin-capitalization detection.
  - Longer-run (Year 5):
    - Structural reorganization could lead to persistent base erosion requiring international cooperation or legislative changes.
- Practical compliance actions:
  - Evaluate economic substance of marketing hubs, royalties, and intercompany debt.
  - Calculate and compare effective tax rates across jurisdictions.
  - Use natural language analysis and GenAI-assisted review of annual reports to identify textual signals (deferred tax anomalies, unremitted earnings).
  - Consider bow-tie risk frameworks for prevention, detection, and corrective controls.

### II. Impact on small and medium taxpayers — revenue and compliance pathway
- Immediate effects (Years 1–2):
  - Exchange-rate pass-through to input costs and local-currency liabilities increases operating costs for SMEs → squeezed margins and taxable profits.
  - Risk of underreporting or noncompliance as margins tighten.
- Medium term (Years 3–4):
  - Business failures or exits reduce taxpayer universe and shrink tax base.
  - Increased informality and cash transactions complicate audit selection and enforcement.
- Operational responses:
  - Reprioritize audit selection using risk-scoring incorporating exchange-rate exposure and cash-intensive sectors.
  - Deploy targeted taxpayer education and temporary relief measures to preserve filing coverage and compliance culture.

### III. Divergent effects across import- and export-exposed industries
- Export-exposed industries (commodity-linked):
  - Immediate decline in profits → lower corporate income tax receipts (Years 1–2).
  - Increased tax avoidance incentives; medium-term investment contraction (Years 3–4).
- Import-exposed industries:
  - Cost increases from weaker currency → reduced profitability and possible noncompliance.
  - Postponed investments reduce VAT and import duty revenues over medium term.
- Cross-sector interactions:
  - Reallocation of activity alters payroll tax and indirect tax bases over the five-year horizon.

### IV. GenAI and analytics-enabled compliance tools — operational considerations and performance metrics
- Use-cases illustrated:
  - Natural language risk-review of large taxpayer annual reports (Figure 5).
  - Dataset-based predictive analytics for audit prioritization with median tax audit adjustment = 27,633.13 used to create binary STRIKE target variable.
  - ROC curve reported with area = 0.65 for a predictive model example.
  - Clustering and SQL-driven agent workflows; agent roles: Administration, SQL coder, SQL Executor, Data Analyst, Data Scientist; integration patterns with ASYCUDA.
- Operational guidance:
  - Validate GenAI outputs carefully; foundation models may hallucinate or lack tax-specific training.
  - Prefer on-premises and open-source configurations where legal and data-sovereignty constraints require; 102 countries operate ASYCUDA and many choose on-premises operation.
  - Augment LLMs with RAG and curated domain data to improve accuracy and contextual relevance.
  - Combine GenAI natural-language capabilities with traditional statistical risk models (confusion matrices, ROC analysis) and domain expertise.

### V. Detection algorithms, audit prioritization, and resource allocation implications
- Detection strategies:
  - Prioritize audits with profit-shifting signals, large effective tax rate differentials, significant deferred tax asset movements, or unremitted earnings.
  - Use GenAI-assisted document review to flag textual indicators, then follow with quantitative validation.
  - Prioritization by expected aggregate adjustment or by turnover depending on policy objective.
- Audit selection metrics and trade-offs:
  - Median-adjustment threshold strategy: median adjustment = 27,633.13 defines STRIKE = “1” labels.
  - Evaluate model performance with ROC area = 0.65 and confusion-matrix diagnostics before deployment.

### VI. Policy recommendations and corrective measures
- Prevention and legislative/regulatory options:
  - Strengthen transfer pricing documentation and country-by-country reporting.
  - Review thin-capitalization rules and controlled-foreign-corporation (CFC) provisions.
  - Consider temporary measures to support vulnerable SMEs to preserve compliance behavior.
- Detection and analytics measures:
  - Integrate natural language analytics with structured data models (text-to-SQL, vector databases) for rapid, context-aware queries.
  - Adopt agentic workflows (Administration, SQL Coder, SQL Executor, Data Analyst, Data Scientist) with human oversight.
  - Continuously validate and monitor model performance using confusion matrices and ROC metrics (example ROC area = 0.65).
- Corrective controls and cooperative tools:
  - Use risk reviews, targeted audits, Advance Pricing Agreements, and cooperative compliance arrangements.
  - Apply corrective measures proportionate to detected noncompliance and aligned with international best practice.

### VII. Scenarios over the five-year horizon (illustrative sequencing)
- Scenario A — Short shock, partial recovery (Years 1–2 immediate shock; Years 3–5 gradual recovery)
  - Year 1–2: Corporate tax revenue declines; increased profit-shifting signals; SME stress rises.
  - Year 3–5: Commodity prices recover partially; tax revenues rebound but structural avoidance patterns persist.
- Scenario B — Prolonged commodity weakness and currency depreciation (multi-year)
  - Persistent low profits among exporters; investment contraction; shrinking taxpayer base; need for legislative reform and international cooperation.
- Scenario C — Rapid policy response and analytical modernization
  - Early adoption of enhanced analytics, RAG-augmented GenAI, and targeted relief measures stabilizes compliance rates and mitigates long-term revenue loss.

### Royalties and other deductions — specific observations and risks
- Observations from demonstration (Multinational Mining Co):
  - Marketing activities reported a 10% target effective tax rate vs a 25% rate for industrial activities.
  - Deferred tax assets valued at $2 billion related to tax losses carried forward.
  - Unremitted earnings of $50 billion retained in subsidiaries for reinvestment.
  - Around 85% of readily marketable inventories are financed by debt, total of $18 billion.
  - Overall net debt reduced to $30 billion during [report period]; capital raised via bonds and debt facilities.
  - Adjusted tax calculations based on EBIT reduced overall applicable tax rate to 17%.
- Tax compliance risks:
  - Profit shifting via marketing hubs, inflated intra-group royalties and service fees, debt-financing to shift interest expenses, and use of royalties as deductible payments to related parties.
- Draft preventative and legislative suggestions:
  - Tightened transfer pricing regulations; require BEPS Action 13-style country-by-country reporting.
  - Substance-over-form legislation requiring economic substance in subsidiaries.
  - Interest deduction limits or thin capitalization rules (e.g., cap interest deductions at fixed percentage of EBITDA).
  - Limitations on deductibility of intra-group royalties and service fees; require APAs or transfer pricing documentation.
- Detection approaches:
  - Effective tax rate discrepancy analysis.
  - Jurisdictional risk scoring for subsidiaries with high revenue but limited personnel/assets.
  - Debt-to-equity ratio tracking.
  - Transaction-based risk engine to flag large intra-group payments.

### Analytics demonstration outputs (selected metrics)
- Data preparation and model setup:
  - Median tax audit adjustment = 27,633.13.
  - STRIKE target variable: adjustment > median = “1”, <= median = “0”.
  - Data split: training dataset with 1,582 cases; verification dataset with 679 cases.
- Model and cross-validation:
  - Random Forest classifier with 500 trees.
  - Fivefold cross-validation scores:
    - Fold 1: 66.25%
    - Fold 2: 65.93%
    - Fold 3: 68.35%
    - Fold 4: 60.13%
    - Fold 5: 70.25%
  - Average cross-validation accuracy: 66.18%.
- Verification results:
  - ROC curve AUC reported as approximately 0.72.
  - Separate ROC annotation shows "area = 0.65".
  - Confusion matrix:
    - True Negatives (TN): 216
    - False Positives (FP): 124
    - False Negatives (FN): 132
    - True Positives (TP): 207
- Targeted ranking outcomes:
  - Top 100 by predicted likelihood:
    - Strike rate: 66%
    - Total aggregated tax adjustments for TP cases within these 100: 23,162,801
  - Top 100 by descending turnover among positive cases:
    - Strike rate: 65%
    - Total aggregated tax adjustments for TP cases within these 100: 41,632,577
- Comparative implication:
  - Similar strike rates (66% vs 65%) but different aggregated adjustments → prioritizing by turnover yields higher total adjustments recovered.

### Strike rate comparison and operational implication
- Strike rates:
  - 66% for likelihood-based ranking.
  - 65% for turnover-based ranking.
- Aggregate adjustments:
  - Top 100 turnover cases: 41.63 million.
  - Top 100 likelihood cases: 23.16 million.
- Policy recommendation:
  - Prioritize audits by turnover when the primary goal is to maximize total adjustments recovered and per-case audit inputs are comparable.

### Mandating human accountability, governance, and implementation guidance
- Human responsibility principle:
  - Risk analysts must remain accountable for decisions that AI influences; must explain procedures, defensibility, boundaries of AI use, and how AI operates.
- Governance requirements:
  - Up-to-date frameworks covering scrutiny of inputs, AI processing, outputs, and monitoring/evaluations.
- Staff training priorities:
  - Ethical and responsible AI use; individual duty of care; information security; predictive analytics; prompt engineering; use of specific AI tools.
- Repository and data management guidance:
  - Build secure repositories (reporting databases, data warehouses, data lakes) hosting structured and unstructured data for RAG/fine-tuning artifacts.
  - Proactive management: data quality, security classifications, audit trails.
  - Controlled experimentation with strong top-down governance; prioritize small, risk-assessed use cases before scaling.
- Operational caution:
  - Special care for risk analysts due to sensitivity of information; ensure experimentation adheres to AI and security policies, especially with externally hosted GenAI services.

### Final thoughts and probable developments
- Central question: “How much AI can be responsibly administered?”
- Systemic impact:
  - GenAI can both improve compliance and facilitate new forms of tax planning, evasion, and fraud.
  - Leading administrations may open additional capacity but must reinvest savings into defensive thinking; low-capacity settings may see the digital divide widen.
- Probable developments:
  - Aggressive mainstream integration of GenAI into digital tools; customized solutions for tax and customs.
  - Medium-term momentum for open-source and on-premises approaches.
  - Certification and innovation may alleviate some data privacy concerns with externally hosted services.
  - Repositories of compliance information for GenAI integration will become central.
  - Business intelligence platforms will increasingly incorporate GenAI and NLP.
- Compliance fundamentals likely to persist:
  - (1) registering correctly, (2) filing on time, (3) reporting accurately, (4) paying on time; for foreign trade: (5) classification, (6) valuation, (7) other reporting including origin, (8) use of procedures.

*Source: tnmea2025013 — IMF Staff.*

### 1. Understand When to Use GenAI . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 

### 1. Understand When to Use GenAI

### Purpose and scope
- Describes generative AI’s (GenAI) ability to support compliance risk analysis in tax and customs administration.
- Written for a technical audience to raise awareness by explaining and demonstrating GenAI capabilities.
- Structure of the note:
  - Conceptual overview of GenAI technology.
  - Four generalized use cases where GenAI can augment the work of risk analysts.
  - Practical demonstrations (proofs of concept) showing how GenAI can be used to better understand taxpayer and trader compliance.
  - Basic guidelines to help administrations considering operational implementation.
  - Forward-looking statements on likely developments.
- The note is intentionally limited in scope to compliance risk analysis for the purpose of demonstrating emerging but practical, high-value applications of the technology.

### Key observations on GenAI’s transformational potential
- Three capabilities with transformational potential for risk functions in tax and customs administration:
  - natural language research,
  - analysis,
  - analytics.
- Historically, effectiveness of risk analysis constrained by ability to source, process, and make sense of data and information; human analysts and a collection of tools were central.
- GenAI reduces barriers by enabling direct interaction with AI in natural language, potentially prompting a rethink of the relationship between analysts and technology.
- With current GenAI tools, capabilities include informing analysis of technical tax matters, generating code for analytics, and interpreting results.
- If model size and sophistication continue to grow rapidly, AI’s utility may soon eclipse that of human analysts—raising questions about how AI can or should be best used.

### Technical and conceptual foundations
- Definition: GenAI is a category of AI designed to work with patterns—for recognition, synthesis, and modeling—to produce humanlike content (text, video, images, audio, or data).
- GenAI systems leverage deep neural networks and model relationships found in training data to generate novel, contextually appropriate outputs and to learn and interact in human languages.
- Current iterations of GenAI build from decades of research:
  - breakthroughs in neural networks in the 1980s,
  - deep learning in the 2000s,
  - Generative Adversarial Networks in 2014,
  - Transformer-based models in the following years.
- Integration into digital services and IT platforms is what makes GenAI potentially revolutionary; mainstream integration can blur lines between real and artificial intelligence and increasingly satisfy criteria of the seminal Turing Test.
- Important challenges remain: inaccuracy and biases in GenAI outputs, including AI hallucination.

### Practical and policy points mentioned
- Demonstrations in the note were developed with both commercial and noncommercial resources and do not specify the underlying technology, services, or suppliers used.
- As a matter of policy, the IMF does not endorse specific technology or suppliers.
- The note benefits from prior exposure to compliance risk management (CRM), analytics for CRM, and AI in tax and customs administration; it suggests reviewers unfamiliar with these topics consult respective IMF materials and the VITARA online course.

### Abbreviations (selected, as presented)
- AI — artificial intelligence
- CRM — compliance risk management
- GenAI — generative artificial intelligence
- ICT — Information and Communications Technology
- IT — Information Technology
- LLM — large language model
- NLP — natural language processing
- RAG — Retrieval-Augmented Generation
- RLHF — reinforcement learning with human feedback
- TNM — technical note and manual
- VAT — value-added tax
- VITARA — Virtual Training to Advance Revenue Administration

*Generative Artificial Intelligence for Compliance Risk Analysis — excerpt from tnmea2025013*

### BOX 1. Essential Terminology

### BOX 1. Essential Terminology

### Definitions
- Artificial Intelligence. The simulation of human  
  intelligence conducted using computer systems,  
  relying on statistics, data science, and machine  
  learning.
- Machine Learning. A major branch of artificial  
  intelligence,  the  use  of  data  and  algorithms   
  to enable computers to learn without explicit  
  programming.
- Algorithm. Logic, methods, or sets of specific 
  steps  for  solving  a  problem  that  may  or  may  
  not be codified for use in a computer program-
  ming language.
- Supervised Learning. A category of machine 
  learning  that  relies  on  humans  to  label  data  
  and   train   algorithms   to   make   conclusions   
  or predictions.
- Unsupervised Learning. A category of machine 
  learning  that  does  not  require  human  “training”  
  to identify patterns in data and make conclusions 
  or predictions.
- Structured  Data.  Data  with  clearly  defined  
  relationships  and  data  types  that  are  easily  
  searchable  (“normalized”  database  tables,  as  
  one example).
- Unstructured Data. Data without clear defi-
  nition, often stored in a format that requires 
  specialized  tools  for  analysis  (for  example,  
  invoices scanned as images).
- Training Data. The data, both structured and 
  unstructured,  used  to  train  machine  learning  
  algorithms to recognize patterns and inform AI 
  models.
- Generative Artificial Intelligence. A branch 
  of  AI  within  machine  learning  that  relies  on  
  pattern  recognition,  synthesis,  and  modeling  
  to generate content.
- Foundation  Model.  Essential to  GenAI,  a  
  general-purpose  machine  learning  model  
  trained   on   data   to   generate   humanlike   
  content and perform a variety of tasks.
- Foundational Model Parameter. An internal 
  parameter  in  a  model  that  is  learned  and  
  adjusted  during  training,  directly  influencing  
  a model’s output.
- Hyperparameter. A  parameter  external  to  a   
  foundation  model  that  governs  the  model’s  
  training  process  and  structure  and  is  not  
  determined from training data.
- Large  Language  Model.  A  category  of  
  foundation  model  trained  on  text  data  and  
  designed to process, interpret, and generate  
  natural language text.
- Compound  Systems.  A  combination  of  AI  
  models  or  complementary  technology  that   
  improve GenAI’s utility (for example, retrieval-
  augmented generation, or RAG).
- Sequenced Agents. Systems or entities that  
  perform  small  tasks  in  a  step-by-step  order  
  to complete a larger task, relying on outputs 
  of one step to inform the next.
- Autonomous  Agents.  Systems  or  entities  
  that  autonomously  process  tasks,  using  AI  
  to  make  decisions  and  communicate  with  
  humans and other agents. 

*Source: IMF Staff.*

### 4. Elaborate on your answer and structure it on how these tax revenue and tax compliance issues may play out

### 4. Elaborate on your answer and structure it on how these tax revenue and tax compliance issues may play out over time

### Executive summary and timeframe
- A 25% devaluation of Country X’s currency, precipitated by significant declines in key commodity prices, will present multifaceted challenges for the tax administration over the next five years.
- Key cross-cutting considerations:
  - Timing and interaction of different tax effects.
  - Sector-specific impacts beyond commodities.
  - Secondary effects on the broader tax base.
  - Practical challenges for tax administration capacity and operations.
  - International tax implications (profit shifting, transfer pricing, withholding regimes).
- Several detailed annexes are referenced for full evidence and extended analysis (See Annex 1, Annex 2, Annex 3, Annex 4, Annex 5).

### I. Impact on large multinational taxpayers — tax revenue issues and timing
- Immediate and medium-term corporate tax revenue effects (as structured in the source):
  - Decreased Corporate Income Tax Revenue
    - Immediate Effect (Year 1–2):
      - The fall in commodity prices directly reduces the profitability of MNCs in the export sector.
      - Lower profits result in decreased corporate income tax payments to the government.
    - Medium Term (Year 3–4):
      - MNCs may delay or cancel expansion projects due to reduced capital, leading to a slowdown in economic growth and a further decline in tax revenues from corporate activities.
  - FX-Related Issues (Immediate Tax Revenue Impacts Years 1–2)
    - Foreign exchange exposures and translation effects can alter reported local-currency profits for MNC subsidiaries or branches (See Annex 1 for Full Details).

- Compliance risk dynamics for MNCs (timing and signals)
  - Short-run (Years 1–2):
    - Profitability declines may increase incentives for profit-shifting to low-tax jurisdictions via royalties, marketing hubs, interest-bearing related-party debt, and other intercompany charges.
    - Immediate scrutiny areas: effective tax rate differentials across jurisdictions, deferred tax asset movements, unremitted earnings, and the economic substance of marketing hubs.
  - Medium-run (Years 3–4):
    - As investments are delayed, tax administrations may see changes in transfer pricing patterns (e.g., increased leverage, altered royalty rates) and thus need enhanced transfer pricing and thin-capitalization risk detection.
  - Longer-run (Year 5):
    - If commodity-price recovery is limited, structural reorganization by MNCs could lead to persistent base erosion and necessitate international cooperation or legislative changes.

- Practical compliance actions and detection approaches (for MNCs)
  - Evaluate economic substance of marketing hubs, royalties, and intercompany debt.
  - Calculate and compare effective tax rates reported in lower-taxing jurisdictions versus higher-taxing jurisdictions.
  - Use natural language analysis and GenAI-assisted review of annual reports to identify textual signals of profit shifting, deferred tax anomalies, and unremitted earnings (see Generative Artificial Intelligence for Compliance Risk Analysis examples and Figure 5 workflow).
  - Consider bow-tie risk frameworks to map prevention, detection, and corrective controls (Annex 2 contains values and examples).

### II. Impact on small and medium taxpayers — revenue and compliance pathway
- Immediate effects (Years 1–2)
  - Exchange-rate pass-through to input costs (imports) and local-currency denominated liabilities can increase operating costs for SMEs, squeezing margins and taxable profits.
  - Some SMEs dependent on imported inputs may shift towards underreporting or noncompliance as margins tighten.
- Medium term (Years 3–4)
  - Business failures or exits reduce taxpayer universe and shrink the tax base.
  - Increased informality and cash transactions complicate audit selection and enforcement.
- Operational responses for tax administration
  - Reprioritize audit selection using risk-scoring that accounts for exchange-rate exposure and cash-intensive sectors.
  - Deploy targeted taxpayer education and temporary relief measures to preserve tax filing coverage and compliance culture.

### III. Divergent effects across import- and export-exposed industries
- Export-exposed industries (commodity-linked)
  - Immediate decline in profits for exporters tied to commodity prices → lower corporate income tax receipts (Years 1–2).
  - Potential increases in tax avoidance incentives (profit shifting) as firms seek to preserve after-tax returns.
  - Medium-term investment contraction (Years 3–4) leading to protracted revenue weakness.
- Import-exposed industries
  - Cost increases from weaker currency raise risks of reduced profitability and possible noncompliance as firms struggle with liquidity.
  - Sectors reliant on imported capital goods may postpone investments, reducing VAT and import duty revenues over the medium term.
- Cross-sectoral interactions
  - Shifts in relative competitiveness can induce reallocation of economic activity between sectors, altering payroll tax and indirect tax bases over the five-year horizon.

### IV. GenAI and analytics-enabled compliance tools — operational considerations and performance metrics
- Use-cases illustrated in source material:
  - Natural language risk-review of large taxpayer annual reports (Figure 5).
  - Dataset-based predictive analytics for audit prioritization with a median tax audit adjustment of 27,633.13 used to create a binary STRIKE target variable.
  - ROC curve reported with area = 0.65 for a predictive model example.
  - Clustering and SQL-driven agent workflows for customs/tax data (Figures 7–10), demonstrating agent roles (Administration, SQL coder, SQL Executor, Data Analyst, Data Scientist) and integration patterns with ASYCUDA.
- Operational guidance (short bullets)
  - Validate GenAI outputs carefully; foundation models are general-purpose and may hallucinate or lack tax-specific training.
  - When feasible, use on-premises and open-source configurations to meet legal and data-sovereignty constraints; note 102 countries operate ASYCUDA and many choose on-premises operation.
  - Augment LLMs with RAG and curated domain data to improve accuracy and contextual relevance.
  - Combine GenAI natural-language capabilities with traditional statistical risk models (confusion matrices, ROC analysis) and domain expertise.

### V. Detection algorithms, audit prioritization, and resource allocation implications
- Detection strategies
  - Prioritize audits where profit-shifting signals, large effective tax rate differentials, significant deferred tax asset movements, or unremitted earnings are present.
  - Leverage GenAI-assisted document review to flag textual indicators in financial statements; follow with quantitative validation.
  - For limited audit resources, prioritize cases by expected aggregate adjustment or by turnover depending on policy objective (see dataset analysis conclusion: prioritizing businesses with higher turnovers yields higher total adjustments even if strike rate is marginally lower).
- Audit selection metrics and trade-offs
  - Use median-adjustment threshold strategy (median adjustment = 27,633.13) to define high-risk STRIKE = “1” labels in supervised models.
  - Evaluate model performance with ROC area = 0.65 and confusion-matrix diagnostics before operational deployment.

### VI. Policy recommendations and corrective measures (prevention, detection, corrective)
- Prevention and legislative/regulatory options
  - Strengthen transfer pricing documentation and country-by-country reporting to reduce profit-shifting opacity.
  - Review thin-capitalization rules and controlled-foreign-corporation (CFC) provisions to address leverage and profit diversion.
  - Consider temporary measures to support vulnerable SMEs to preserve compliance behavior.
- Detection and analytics measures
  - Integrate natural language analytics with structured data models (text-to-SQL, vector databases) to enable rapid, context-aware queries on customs and taxpayer databases.
  - Adopt agentic workflows (Administration, SQL Coder, SQL Executor, Data Analyst, Data Scientist) with human oversight to automate routine detection and data extraction tasks.
  - Validate and monitor model performance continuously using confusion matrices and ROC metrics (example ROC area = 0.65).
- Corrective controls and cooperative tools
  - Use risk reviews, targeted audits, Advance Pricing Agreements, and cooperative compliance arrangements where appropriate.
  - Apply corrective measures proportionate to detected noncompliance and aligned with international best practice for dispute resolution and mutual agreement procedures.

### VII. Scenarios over the five-year horizon (illustrative sequencing)
- Scenario A — Short shock, partial recovery (Years 1–2 immediate shock; Years 3–5 gradual recovery)
  - Year 1–2: Corporate tax revenue declines; increased profit-shifting signals; SME stress rises.
  - Year 3–5: Commodity prices recover partially; tax revenues rebound but structural avoidance patterns persist requiring sustained enforcement.
- Scenario B — Prolonged commodity weakness and currency depreciation (multi-year)
  - Persistent low profits among exporters; investment contraction; shrinking taxpayer base for both corporate and payroll taxes; heightened need for legislative reform and international cooperation.
- Scenario C — Rapid policy response and analytical modernization
  - Early adoption of enhanced analytics, RAG-augmented GenAI, and targeted relief measures stabilizes compliance rates and mitigates long-term revenue loss.

*Source: IMF Staff.*

### 2. Mandate Human Accountability for Outcomes

### 2. Mandate Human Accountability for Outcomes

### Principle of human responsibility for AI
- Human responsibility for AI outcomes must be clearly stated as a core principle in formal AI usage policy.
- Risk analysts using AI, including GenAI, must:
  - remain accountable for decisions that AI influences or determines; and
  - be capable to explain which procedures were applied, why they are defensible, the boundaries within which AI may be used, and how the AI operates (technically and functionally).
- The principle extends across all stages of AI development and use, starting from core data collection, management, and use in training AI systems.
- Requirement: up-to-date governance frameworks and procedures covering:
  - scrutiny of inputs into AI systems;
  - scrutiny of AI processing;
  - scrutiny of outputs that AI generates; and
  - monitoring and evaluations to determine whether AI is performing as expected.

### Anticipate changes and prioritize staff training
- Organizational implications of GenAI may be counterintuitive; develop a plan before adoption.
- GenAI’s accessibility with natural language makes it relevant to more staff than prior AI forms.
- Expect both potential efficiency gains and significant new increases in overhead costs to:
  - responsibly maintain the technology; and
  - make effective use of it.
- Common cost areas: enhancements to data and IT governance, and investments in staff.
- Training priorities for those engaged in risk analysis include exposure to:
  - ethical and responsible uses of AI;
  - principles of individual duty of care;
  - information security;
  - predictive analytics;
  - prompt engineering; and
  - use of specific AI tools.

### Build (and secure) a repository of compliance information
- GenAI’s utility in risk analysis is greatest when adapted to account for relevant domain knowledge and integrated with internal, confidential information.
- Administrations should compile, secure, and carefully monitor a repository of confidential compliance information because:
  - foundation models underlying GenAI are pretrained on content intended for general consumption;
  - taxpayer and trader compliance history, fraud, and investigations are confidential and must be protected.
- Adaptation of GenAI to internal repositories is critical and likely a significant hurdle for many administrations.

Features of a repository adapted for use by GenAI
- Secure infrastructure for “big data”:
  - often includes a combination of centralized (a) reporting databases, (b) data warehouses, and (c) data lakes;
  - platforms host both structured and unstructured data;
  - using unstructured big data is less common and may require new investments.
- Proactive management:
  - well-administered repositories incur significant overhead costs for data management and information security policies and controls;
  - priorities include ensuring the quality of data in the repository (requiring strong domain knowledge), introducing security classifications to manage access to different levels of information confidentiality, and mandating audit trails.
- Structured for augmenting GenAI services:
  - repository could include artifacts and data sets for RAGs and fine-tuning, such as legislation and procedures, environmental scans, industry benchmarks, prior rulings, data sets describing compliance history, risk profiles, declarations, results of large taxpayer risk reviews, case files (service and enforcement), and reports from investigations.

### Iterate adoption of GenAI slowly, with caution
- Encourage controlled experimentation; avoid directionless introduction.
- Strong top-down governance is a critical starting point.
- Prioritize small numbers of specific, risk-assessed use cases; scale up for operational purposes only with proven use cases.
- Special care required for risk analysts due to sensitivity of information processed:
  - ensure experimentation adheres to AI and security policies, especially when using externally hosted GenAI services.

### Conclusion: systemic impact, probable developments, and final thoughts
- Central question reframed: “How much AI can be responsibly administered?”
- GenAI can support all facets of compliance risk analysis and may take roles from short-lived project support to long-running autonomous handling of routine tasks, including suites of AI agents collaboratively advancing complex compliance activities.
- The onus for defensible, responsible applications will remain human far into the future.

GenAI’s systemic impact
- As GenAI improves compliance, it may also facilitate new and aggressive forms of tax planning, evasion, and fraud.
- Net global impact will likely be unbalanced:
  - leading administrations may open additional capacity with GenAI but may need to reinvest savings into defensive thinking;
  - administrations in low-capacity settings are likely to experience GenAI’s arrival differently and may see the digital divide widen.
- Example numeric context: median ICT expenditure in 2022 was 14.20 percent in major advanced economies (G7) but only 0.66 percent in the Middle East and Central Asia (note provided in source).

Probable future developments
- Expect aggressive, mainstream integration of GenAI into digital tools and managed agentic environments, and emergence of customized solutions for specialized domains, including tax and customs administration.
- In the medium term, open-source and on-premises approaches are likely to gain momentum.
- Certification and technological innovation may alleviate some data privacy concerns with externally hosted GenAI services.
- Establishing repositories of compliance information for integration into GenAI services will become a key fixation.
- Business intelligence platforms will increasingly incorporate GenAI and NLP, reducing technical barriers and democratizing access to data and information.
- Management and use of agents will increasingly feature in conversations among risk analysts.

Final thoughts on compliance fundamentals
- Core taxpayer responsibilities likely to persist:
  - (1) registering correctly, (2) filing on time, (3) reporting accurately, and (4) paying on time.
- For those in foreign trade, obligations will continue for:
  - (5) classification, (6) valuation, (7) other reporting, including origin, and (8) use of procedures.
- GenAI will be embedded throughout digital environments over time, not just as an add-on.
- GenAI’s arrival may accelerate work to get fundamentals right and lead to new conceptions of how fundamentals can be administered, particularly in data-rich subjects like compliance risk analysis.

*Source: tnmea2025013 - 2. Mandate Human Accountability for Outcomes*

### Executive Summary

### Executive Summary

### Key areas warranting deeper exploration
- The timing and interaction of different tax effects
- Sector-specific impacts beyond commodities
- Secondary effects on the broader tax base
- Practical challenges for tax administration
- International tax implications

### I. Additional Considerations for Multinational Corporations (MNCs)

#### A. Immediate Tax Revenue Impacts (Years 1-2)
- 1. Foreign Exchange (FX) Related Issues
  - Timing of FX losses recognition and deductibility
  - Treatment of hedging arrangements and their tax implications
  - Complex tax accounting for unrealized FX gains/losses
  - Impact on thin capitalization calculations due to revaluation of foreign debt
- 2. Supply Chain Restructuring
  - Tax implications of emergency supply chain reorganizations
  - Permanent establishment risks from temporary relocations
  - Changes in profit attribution among group entities
- 3. Working Capital Adjustments
  - Tax treatment of emergency capital injections
  - Deductibility of increased financing costs
  - Transfer pricing implications of intra-group financing

#### B. Medium-Term Effects (Years 2-4)
- 1. Strategic Corporate Restructuring
  - Tax implications of potential mergers and acquisitions
  - Treatment of business restructuring costs
  - Impact on existing tax incentives and holidays
  - Changes in effective tax rates due to altered business models
- 2. Investment Decisions
  - Tax treatment of abandoned projects
  - Recognition of impairment losses
  - Changes in R&D tax incentive utilization
  - Impact on capital allowance claims

#### C. Long-Term Considerations (Years 4-5)
- 1. Business Model Evolution
  - Tax implications of vertical integration strategies
  - Changes in intellectual property holding structures
  - Impact on existing advance pricing arrangements
  - Evolution of permanent establishment footprint

### II. Enhanced Analysis for Small and Medium Enterprises (SMEs)

#### A. Sector-Specific Impacts
- 1. Service Sector
  - Treatment of long-term contracts denominated in foreign currency
  - Impact on VAT/GST compliance for cross-border services
  - Tax implications of changed payment terms with international clients
- 2. Manufacturing Sector
  - Tax treatment of inventory value adjustments
  - Impact on R&D tax incentives
  - Treatment of equipment lease payments in foreign currency
- 3. Retail Sector
  - Changes in transfer pricing for franchisee arrangements
  - Impact on thin capitalization rules for retail financing
  - Treatment of terminated commercial leases

#### B. Compliance Challenges
- 1. Resource Constraints
  - Impact on tax function staffing and expertise
  - Ability to maintain proper documentation
  - Capacity to engage with tax authorities
- 2. Systems and Processes
  - Currency translation in tax reporting systems
  - Maintenance of dual-currency records
  - Integration of tax and accounting systems

### III. Cross-Cutting Issues

#### A. VAT/GST Considerations
- 1. Timing Issues
  - Treatment of deposits and advance payments
  - Impact on bad debt relief claims
  - Changes in input tax recovery patterns
- 2. Valuation Challenges
  - Treatment of related party transactions
  - Impact on customs value calculations
  - Changes in transfer pricing adjustments

#### B. Employment Tax Issues
- 1. International Assignments
  - Treatment of cost of living adjustments
  - Impact on shadow payroll calculations
  - Changes in social security obligations
- 2. Compensation Restructuring
  - Tax treatment of currency protection schemes
  - Impact on share-based payment arrangements
  - Changes in bonus and incentive structures

### IV. Administrative Challenges for Tax Authorities

#### A. Operational Issues
- 1. Resource Allocation
  - Need for specialized forex expertise
  - Training requirements for staff
  - Technology system updates
- 2. Audit Strategy
  - Development of new risk assessment criteria
  - Changes in audit focus and methodology
  - Resource allocation between taxpayer segments

#### B. Policy Responses
- 1. Guidance and Rulings
  - Need for real-time guidance on complex issues
  - Development of practical compliance approaches
  - Balance between enforcement and support
- 2. International Cooperation
  - Enhanced information exchange requirements
  - Coordination on cross-border enforcement
  - Mutual agreement procedure capacity

### V. Timeline of Critical Issues
- Year 1
  - Emergency response to forex losses
  - Initial supply chain disruptions
  - Immediate working capital challenges
  - Initial compliance challenges
- Year 2
  - Business model adjustments
  - Transfer pricing modifications
  - Initial restructuring activities
  - Systems and process updates
- Year 3
  - Strategic repositioning
  - Enhanced compliance frameworks
  - International structure changes
  - Policy response implementation
- Years 4 - 5
  - Long-term business model evolution
  - Stabilization of tax base
  - Enhanced administrative capacity
  - International cooperation framework

### VI. Recommendations for Tax Administration

#### A. Short-Term Actions
- Establish emergency response team
- Issue practical guidance on key issues
- Implement flexible payment arrangements
- Enhance technical capacity

#### B. Medium-Term Initiatives
- Develop specialized audit programs
- Enhance international cooperation
- Update risk assessment frameworks
- Implement technology solutions

#### C. Long-Term Strategy
- Reform tax legislation
- Strengthen administrative capacity
- Enhance international engagement
- Develop specialized expertise

### Demonstration example: GenAI review of a multinational mining company annual report (selected findings)
- Effective tax rate differentials
  - Marketing activities reported a 10% target effective tax rate compared to a 25% rate for industrial activities.
  - Lower tax rate in marketing suggests use of marketing hubs to book profits in lower-tax jurisdictions.
- Subsidiary structures and placements
  - Deferred tax assets valued at $2 billion related to tax losses carried forward.
  - Unremitted earnings of $50 billion retained in subsidiaries for reinvestment.
- Debt usage and leveraging
  - Around 85% of readily marketable inventories are financed by debt, a total of $18 billion.
  - Overall net debt was reduced to $30 billion during [report period]; significant capital was raised via bonds and debt facilities.
  - Adjusted tax calculations based on EBIT show the overall applicable tax rate was reduced to 17%.
- Royalties and other deductions
  - Government-imposed charges treated as income tax in some cases and structured to be deductible for tax purposes.

### Conclusion
- The currency devaluation creates complex and interrelated tax challenges that evolve over time.
- Success in managing these issues requires:
  - Balanced approach between enforcement and support
  - Enhanced technical capacity
  - Strong international cooperation
  - Flexible and responsive administrative systems
- Tax administrations must be prepared for both immediate challenges and long-term structural changes in the tax base and compliance landscape.

*Source: tnmea2025013 - Executive Summary.*

### 4. Royalties and Other Deductions

### 4. Royalties and Other Deductions

### Observation and Context
- Multinational Mining Co makes use of royalties, service fees, and other deductible payments, which may be structured to shift profits to lower-tax jurisdictions.
- Royalties are considered part of Multinational Mining Co’s income tax calculation (MULTINATIONAL MINING CO ---Annual-R...), which can directly reduce taxable income in higher-tax jurisdictions.
- Payments such as royalties and levies can be structured as tax-deductible expenses, especially when the recipient is a related party in a lower-tax jurisdiction.
- Royalties are mentioned in relation to resource extraction payments that are treated similarly to corporate income tax (MULTINATIONAL MINING CO ---Annual-R...).
- Operating expenses related to marketing could potentially include management and service fees paid to entities in lower-tax jurisdictions (MULTINATIONAL MINING CO ---Annual-R...).

### Summary of Tax Compliance Risks Identified
- Profit shifting through the use of marketing hubs in low-tax jurisdictions, particularly by achieving a much lower tax rate in marketing activities (10%) compared to industrial activities (25%) (MULTINATIONAL MINING CO ---Annual-R...).
- Subsidiary structuring and deferred tax assets, especially unremitted earnings totaling $50 billion, suggest strategic placement of subsidiaries in low-tax jurisdictions (MULTINATIONAL MINING CO ---Annual-R...).
- Debt-financing and significant use of intercompany loans to shift interest expenses to higher-tax jurisdictions, reducing the overall tax burden (MULTINATIONAL MINING CO ---Annual-R...).
- Royalties and service fees that could shift profits to low-tax jurisdictions through inflated or non-arm’s length transactions (MULTINATIONAL MINING CO ---Annual-R...) (MULTINATIONAL MINING CO ---Annual-R...).
- The inclusion of royalty payments in tax calculations raises the question of whether these deductions are being used to reduce tax liabilities beyond what would be expected in an arm’s length transaction.

### Risk Management Implications
- Royalties and service fees are often scrutinized by tax authorities because they can be used to strip profits from one jurisdiction and reallocate them to another.
- Key compliance concern: whether these payments reflect genuine business transactions at arm’s length or are inflated to shift profits artificially.
- These arrangements are typically reviewed in transfer pricing audits to ensure payments are commensurate with the economic value provided by the related-party entity.
- Tax authorities may need to review:
  - The economic substance of marketing hubs and related subsidiaries.
  - The appropriateness of intercompany pricing for royalties, service fees, and management fees.
  - The legitimacy of debt allocation and the scope of interest deductions across jurisdictions.

### Draft Preventative Measures and Legislative/Regulatory Suggestions
- Tightened Transfer Pricing Regulations
  - Enforce stricter rules on transfer pricing, particularly on the economic substance of marketing hubs and royalty payments.
  - Regulatory measure suggested: Implement Base Erosion and Profit Shifting (BEPS) Action 13 guidelines, mandating country-by-country reporting.
  - Detection approach: Create risk indicators based on effective tax rate discrepancies between jurisdictions where a company operates vs. their global footprint.
- Substance-over-Form Legislation
  - Require economic substance (real activities, personnel, infrastructure) in subsidiaries that report significant income.
  - Detection approach: Jurisdiction risk scoring system that flags subsidiaries in low-tax jurisdictions with disproportionately high revenue relative to workforce, assets, and physical presence.
- Interest Deduction Limits and Thin Capitalization Rules
  - Restrict the amount of interest a company can deduct based on the company’s earnings.
  - Regulatory measure: Cap interest deductions at a fixed percentage of EBITDA or enact thin capitalization rules limiting debt-to-equity ratios.
  - Detection approach: Risk algorithm analyzing debt-to-equity ratios of subsidiaries and flagging significant deviations from industry norms.
- Limitations on Deductibility of Royalties and Fees
  - Restrict deductibility of intra-group royalties and service fees, especially when paid to entities in low-tax jurisdictions.
  - Regulatory measure: Require Advance Pricing Agreements (APAs) or transfer pricing documentation for all intra-group service and royalty arrangements.
  - Detection approach: Transaction-based risk engine that flags large intra-group payments (e.g., royalties, fees) made to entities in tax-favorable jurisdictions where the recipient has limited operational activity.

### Detection Approaches and Risk Algorithms (as proposed)
- Effective Tax Rate Discrepancies: Analyze deviations in effective tax rates compared to average rates for similar companies or industries.
- Jurisdictional Risk Scoring: Identify subsidiaries in low-tax jurisdictions with limited personnel or assets relative to income.
- Debt-to-Equity Ratio Tracking: Monitor excessive use of debt financing in high-tax jurisdictions.
- Intra-Group Payments Analysis: Flag large intra-group payments to low-tax jurisdictions, especially where the receiving entity lacks economic substance.
- Transaction-based risk engine: Flag large royalties, management fees, and service fees paid to related parties in low-tax jurisdictions.

### Corrective Controls and Enforcement/Cooperation Tools
- Tax Audits Targeting Transfer Pricing
  - Conduct in-depth tax audits of companies with significant marketing hubs and intra-group transactions to verify economic substance and pricing.
  - Apply penalties and adjustments in cases of non-compliance.
- Advance Pricing Agreements (APAs)
  - Encourage binding APAs to ensure intra-group pricing (including debt arrangements, royalties, and service fees) complies with the arm’s length principle.
- Risk Reviews and Data Analytics
  - Implement ongoing risk reviews using tax data analytics to identify patterns consistent with tax avoidance schemes, such as sudden shifts in profit allocation to low-tax jurisdictions.
  - Increase monitoring and follow-up audits where red flags persist.
- Cooperative Compliance Programs
  - Offer voluntary disclosure programs or tax transparency initiatives where companies disclose tax structures and engage early with tax authorities.
  - Possible incentive: reduced penalties in exchange for cooperation and behavioral change.

### Bow-Tie Framing (Central Risk and Elements)
- Central Risk: Tax Avoidance through Profit Shifting, Debt Structuring, and Use of Low-Tax Jurisdictions.
- Left-side causes include:
  - Profit shifting via marketing hubs taxed at 10% vs industrial activities at 25%.
  - Subsidiary structures with unremitted earnings totaling $50 billion.
  - Excessive debt allocation including allocation of debt-financing for 85% of its readily marketable inventories.
  - Royalties and management fees structured as deductible payments to related parties in low-tax jurisdictions.
- Right-side consequences include:
  - Reduced tax revenue for higher-tax jurisdictions.
  - Distortion of competitive market conditions.
  - Reputational damage and legal sanctions.

### Analytics Demonstration Outputs (from the GenAI demonstration included in the source)
- Data preparation and model setup:
  - Median tax audit adjustment calculated: 27,633.13.
  - STRIKE target variable created where adjustment > median = “1”, <= median = “0”.
  - Data split: training dataset with 1,582 cases; verification dataset with 679 cases.
- Model and cross-validation:
  - Random Forest classifier with 500 trees used.
  - Fivefold cross-validation scores:
    - Fold 1: 66.25%
    - Fold 2: 65.93%
    - Fold 3: 68.35%
    - Fold 4: 60.13%
    - Fold 5: 70.25%
  - Average cross-validation accuracy: 66.18%.
- Verification results:
  - ROC curve AUC reported as approximately 0.72.
  - A separate ROC annotation shows "area = 0.65" in the figure text.
  - Confusion matrix:
    - True Negatives (TN): 216
    - False Positives (FP): 124
    - False Negatives (FN): 132
    - True Positives (TP): 207
- Targeted ranking and outcomes:
  - Top 100 by predicted likelihood (highest probability of STRIKE = “1”):
    - Strike rate: 66%
    - Total aggregated tax adjustments for TP cases within these 100: 23,162,801
  - Top 100 by descending turnover among positive cases:
    - Strike rate: 65%
    - Total aggregated tax adjustments for TP cases within these 100: 41,632,577
- Comparative comment (as rendered in the source):
  - The two selection approaches produced similar strike rates (66% vs 65%) but different aggregated adjustments (23,162,801 vs 41,632,577), implying different expected aggregate audit yields depending on case selection criteria.

*Source: tnmea2025013 - 4. Royalties and Other Deductions (tnmea2025013 - 4. Royalties and Other Deductions)*

### 1. Strike Rate Comparison:**

### 1. Strike Rate Comparison:

### Strike rate findings
- The strike rates are:
  - 66% for the likelihood-based ranking
  - 65% for the turnover-based ranking
- Interpretation:
  - Both methods are effective in predicting true positives.
  - The likelihood-based method has a slight edge in strike rate.

### Aggregate adjustments
- Total adjustments for the top 100 turnover cases: 41.63 million
- Total adjustments for the top 100 likelihood cases: 23.16 million
- Interpretation:
  - The turnover-based ranking significantly outperforms the likelihood-based ranking in terms of aggregated adjustments.
  - Businesses with higher turnovers tend to have larger tax adjustments when they are true positives.
- Operational implication:
  - If the objective is to maximize total adjustments recovered, prioritizing businesses with higher turnovers yields a higher rate of return, even though the strike rate is marginally lower.
  - This approach is particularly advantageous when audit resources are limited and audits require similar inputs across cases.

### Policy recommendation
- Prioritize audits by turnover (as in Step 9) when the primary goal is to maximize total adjustments recovered and audit resource constraints make per-case inputs comparable.

*Source: tnmea2025013 - 1. Strike Rate Comparison*

---


_Source: https://www.imf.org/-/media/files/publications/tnm/2025/english/tnmea2025013.pdf_