Implementing AI in Transfer Pricing: Will It Work the Same Way as in Financial Statement Auditing?

Recent publications from PwC, EY and KPMG describe how AI can reduce routine finance work and support deeper, potentially higher-quality audits.

Will AI affect transfer pricing similarly, or will its impact differ?

Transfer pricing differs because it spans decision and review points before and long after the transaction (T):

1. Planning and preliminary benchmarking (before T)
2. Price setting and transaction (T)
3. Recording actuals in ERP (during the year)
4. TP analysis, Local File and adjustments (at or after year-end)
5. CIT return (after year-end)
6. Pillar Two reporting (during the following reporting cycle)
7. Audit, MAP or litigation (years later)

At each stage, more information becomes available. However, more information does not automatically mean more relevant precedents, reliable evidence or reliable conclusions. Its effect depends on whether it helps distinguish between plausible versions of the case.

What kinds of uncertainty remain?

For this purpose, three forms of uncertainty can be distinguished in transfer pricing:

1. Uncertainty about the facts and circumstances of the transaction. This concerns what the parties actually did, who performed the functions and controlled the risks, and what amounts existed. Two companies may record the same transaction differently in their ERPs due to different accounting standards or policies, requiring reconciliation.
2. Uncertainty among analytical versions. A version is a complete or partial path from facts and assumptions through the delineation of the transaction, method selection, comparables and calculations to a transfer pricing conclusion. Different versions may follow different paths and reach the same result, diverge only on particular elements or produce entirely different conclusions. Where AI generates outputs that differ materially in meaning, this divergence may be examined through semantic entropy. Different paths leading to the same conclusion, however, do not necessarily imply high semantic entropy. Recent AI research uses this concept to distinguish uncertainty across meanings from mere differences in wording.
3. Legal and procedural uncertainty. This concerns how the facts, evidence and analytical version will be assessed by a tax authority, competent authorities or a court, and how an audit, MAP or litigation will ultimately be resolved.

What does this mean for AI and automation in transfer pricing?

These uncertainties do not necessarily move in the same direction. ERP may clarify actual amounts and reduce factual uncertainty, while inconsistencies found in the data create additional analytical versions. A Local File formalises one version but does not establish that it is correct. During an audit, additional evidence may clarify the facts, while a competing position introduced by the tax authority increases legal and procedural uncertainty. Subsequent evidence review or MAP may narrow it again.

For AI, movement through stages 1–7 can be viewed as progressive enrichment of context—an in-context learning effect rather than retraining of the underlying model. Within this expanding context, versions can be represented as complete or partial branches connecting facts, assumptions, methods, calculations and conclusions. New supporting evidence strengthens particular branches; contradictory evidence weakens or eliminates them; and new assumptions, rules and competing positions create additional branches.

The central risk is that AI may select one branch too early. Later stages may then treat that version as an established premise, allowing an initial mistake or unsupported assumption to propagate through the analysis. A coherent and increasingly detailed result may therefore create greater confidence without a proportionate increase in reliability.

The progression through stages 1–7 is consequently not a simple, continuous reduction of uncertainty. New assumptions, rules and competing positions create additional branches, while relevant facts, evidence and verification strengthen, weaken or prune them.

This also clarifies a separate source of AI’s value. Complexity concerns the number of records, analytical layers and interdependencies that must be processed; it is not uncertainty itself. AI can make this work more manageable by structuring information, tracing relationships and comparing versions, even when it cannot determine which final conclusion is most reliable. The human role shifts towards assessing context, comparability, assumptions and evidential weight.

AI’s effectiveness in dealing with uncertainty is likely to improve as more complete analytical cycles are performed with human involvement. The most valuable input may not be limited to the intermediate options selected by the expert, but may also include the reasons why one option was preferred over another. Capturing these explanations would allow AI to learn from the structure of professional judgement, rather than only from final conclusions.

What changes if AI makes retrospective verification faster and less expensive?

The focus may then shift further towards Operational Transfer Pricing: building reliable processes before and during the transaction rather than reconstructing them years later.

However, this shift should not be assumed to happen automatically. AI may play very different roles in planning and preliminary benchmarking, price setting, tax reporting, and audit or dispute resolution.

• A system trained or configured using well-designed and validated price-setting examples may help structure future transactions.
• A system trained mainly on examples of ex post adjustments will instead be better suited to identifying and correcting deviations after they occur. That capability will not necessarily produce reliable ex ante price setting.

Sound methodology, clear processes and, ideally, automation are therefore essential for creating suitable training examples and directing AI towards the intended task.

One possible direction is the creation of multiple, secure and well-governed Transfer Pricing Data Spaces for sharing complementary, properly anonymised knowledge among MNE groups, among advisory firms and, on a bilateral or multilateral basis, among tax authorities. By broadening the relevant experience available within each community, such arrangements could help test and eliminate weaker versions earlier—rather than waiting years for audits, MAP procedures or litigation to run their course.

Sources

• PwC, How AI Agents Help Drive a New Finance Operating Model: What CFOs Need to Know.
• EY, EY Launches Enterprise-Scale Agentic AI to Redefine the Audit Experience for the AI Era.
• KPMG, FY25 Audit Quality Report.
• Stephen Foley, AI Has Arrived in Auditing. Are Regulators Ready?, Financial Times, 6 April 2026.
• Sebastian Farquhar et al., Detecting Hallucinations in Large Language Models Using Semantic Entropy, Nature, 2024.