AI in Transfer Pricing

Transfer pricing rules require that transactions between related entities, such as intercompany sales of goods, services, loans and intellectual property licenses, be priced as independent parties would price them. This is the arm's length principle. It is set out in the OECD Transfer Pricing Guidelines and, in the United States, in Section 482 of the Internal Revenue Code and its regulations. Recognized methods include the comparable uncontrolled price method, resale price method, cost plus method, transactional net margin method (called the comparable profits method in US rules) and profit split method. Most routine entities are tested with a profit-level indicator, such as operating margin, compared against a set of independent companies. The result is usually expressed as an interquartile range. Building that set is slow. Analysts search databases such as Moody's Orbis or S&P Capital IQ by industry code and region. They apply quantitative screens such as independence, size and years of data. Then they read business descriptions and websites to reject companies with different functions. The qualitative review is where AI helps most. Models can read descriptions in several languages, suggest accept or reject decisions and draft rejection reasons. Documentation is the second use. Under the OECD's BEPS Action 13 framework, many countries require a master file and a local file. Large groups, generally those with consolidated revenue of at least 750 million euros, also file country-by-country reports. AI can draft industry overviews, functional analyses and transaction descriptions from source material. A common misconception is that AI can pick the comparables. Tax authorities expect a reproducible search and defensible human judgment. An unexplained model decision is hard to defend in an audit. Another misconception is that AI removes the need for good facts. A fluent functional analysis built on wrong interview notes is still wrong.