In late 2022, several AI tools became available, sparking a wave of transformation that is reshaping our daily routines and offering a new era of possibilities within professional domains. Among those, transfer pricing (TP) is profoundly impacted, making it increasingly crucial to grasp both AI’s potential and its limitations.
TP often relies on the review of multiple sets of data, such as accounting records, legal documentation, extensive market and sector analysis, and comparable transactions, all within complex regulatory frameworks. It often combines several disciplines encompassing tax, accounting, and corporate finance, together with the requirement of a deeper understanding of each client’s business activities, environment, and transactions.
AI can also facilitate rapid review, classification, and interpretation of purely quantitative data, such as multiple accounts, key terms extracted from complex agreements, or internal client data. It can prepare summaries or forecasts, allowing TP practitioners to quickly spot issues and inconsistencies, and determine the right approach. Furthermore, AI can streamline certain aspects of economic analysis, including data extraction for benchmarking and initial filtering. Finally, it can generate certain parts of product services deliverables, such as appendices or highly standardised and descriptive sections of documentation.
However, while it clearly offers support and excels in quantitative analysis, AI cannot replace human TP expertise. Human intervention remains essential, particularly in areas requiring professional judgement, such as in the context of intricate value chains. For instance, an analysis of intellectual property assets would entail not only the examination of quantitative data and facts but also the mapping of key functions and responsibilities, which frequently extend beyond mere job descriptions. In this domain, human judgement and discussions with each involved party often result in conclusions that could have been different from those derived solely from a quantitative or mechanical approach embedded in AI.
The selection and determination of comparable transactions
Even for relatively simple analyses, the ultimate selection and determination of comparable transactions still require professional discretion, particularly when the criteria extend beyond quantitative characteristics to include legal precedents. In today's world, courts and tax authorities not only closely examine data sets and require increased comparability from those retained but also increasingly rely on anti-abuse provisions such as those concerning the abuse of law. Those concepts involve subjective assessment and interpretation of a transaction.
As a succinct example, consider an investment fund that acquires an entire business group from an existing multinational company. The fund creates a dedicated property company to acquire, finance, and hold the business. Tax authorities in the business's source country may want to review and understand the economic context of these transactions. This is where a TP practitioner must demonstrate, using client-specific business and structural arguments, that the property company has valid business reasons to exist and is provided with sufficient substance to satisfy those enquiries. This is something that AI, relying solely on existing algorithms, cannot handle on its own.
Interpretation of regulations and the potential for mistakes
Although AI development is progressing rapidly, making it difficult to fully gauge its potential, interpreting key regulations involving abstract language, such as the arm's-length principle, remains a domain where professional judgement is crucial. AI remains a probabilistic model, which lacks the ability to think independently.
Contractual, legal, and ethical risks constitute a challenge on their own. TP heavily relies on third-party and proprietary databases, where intellectual property plays a fundamental role. Furthermore, legal and contractual considerations, such as client data protection, currently raise multiple questions about the ability of AI to fully comply with these legal constraints.
Finally, AI can, and does, make mistakes, whether due to errors in the data it processes or what is sometimes termed ‘hallucination’. While detecting errors in databases or provided data may sometimes be nearly impossible, human control and oversight can prevent cases where AI generates invented or inaccurate conclusions.
Final thoughts on the role of AI in transfer pricing
The integration of AI into TP opens a new era of efficiency and accuracy. However, it is imperative to strike a balance between automation and human intervention. As the OECD has affirmed, TP is not an exact science. Unlike AI, which relies on mathematics and algorithms, TP requires interpretation and judgement. Thus, while AI can enhance multiple TP processes, it is essential to recognise the central and indispensable role of the TP professional.
