AI for transfer pricing: a practitioner read
Published 12 May 2026
I spent a substantial portion of my Big 4 years on transfer pricing. The work that delivered a £125m UK tax burden reduction for a major PLC across more than 1,000 group entities was, fundamentally, a documentation-heavy, comparables-heavy, judgement-heavy exercise. The shape of that work has not changed in twenty years. The tools the team uses to do it have changed a great deal.
The question worth answering in 2026 is not “will AI change transfer pricing.” It already has. The question is which parts of the work it changes, where the risk now sits, and what a head of tax or a CFO should be doing about it before the next documentation cycle.
What is actually happening in the TP-AI market
Two things in the last twelve months, both worth knowing about.
First, dedicated TP-AI vendors are now real. Aibidia closed a $28m Series B in June 2025 and launched TP Aurora, a dedicated AI research assistant for transfer pricing built in partnership with IBFD, on 22 May 2025. (Source: GlobeNewswire, 22 May 2025; Axios on the Series B.) Aibidia’s customer list now includes Unilever, Nokia, Dyson, and Delivery Hero. Reptune in Amsterdam has shipped an AI documentation assistant for Local File, Master File, and CbCR drafting since 2024. ArmsLength AI is marketing quantitative-screening and NACE-lookup tooling. The market has gone from a handful of advisory firms with internal scripts to a small but funded vendor ecosystem in eighteen months.
Second, the major LLM vendors have not entered transfer pricing directly. Anthropic’s 5 May 2026 finance agents launch shipped ten templates, and none of them is for transfer pricing or international tax. (Source: anthropic.com/news/finance-agents.) The valuation reviewer agent is the closest the launch comes to TP-adjacent scaffolding (it checks valuations against comparables and methodology), and it stops short. That gap is itself a buying signal. The TP vendors will keep their differentiation for at least the next 18 months because the general-purpose model vendors are not coming for this specific corner of finance yet.
Where AI is genuinely earning its keep
Three areas, in approximate order of maturity.
Documentation drafting. Master File, Local File, and CbCR commentaries are now produced with AI drafting assistance at most large in-house tax teams I have talked to. The agent ingests the controlled transaction schedule, the functional analysis, the comparables analysis, and produces a structured first draft of the Local File. The team reviews and substantively edits. The time saving on the production layer is real, often in the order of weeks per entity. The quality of the first draft is unsurprising. Fluent, structurally correct, sometimes substantively bland, occasionally misleading. As with the month-end close agent, the production gain is real and the review layer holds.
Comparables search and benchmarking. Aibidia, Reptune, and ArmsLength AI all market AI features that search Bureau van Dijk’s Orbis and Amadeus, apply quality screens, and produce ranges that are defensible to a tax authority. The published evidence on accuracy is vendor-supplied and not independently audited. CIAT’s 2025 paper on AI agents in transfer pricing observes that AI “tends to perform more stably than humans who may become fatigued by repetitive tasks” but that “consistency and reliability of AI performance can fluctuate.” (Source: CIAT, 2025.) That is a careful claim, and the right one. AI is better than tired humans at filtering a database and worse than focused experts at deciding which filters matter for this specific transaction.
Risk assessment and audit defence. EY’s 2025 Tax and Finance Operations Survey (Oxford Economics, fieldwork July to September 2025, 1,600 respondents in 32 jurisdictions) found that 70% of tax leaders have already implemented or are integrating at least one GenAI tool for tax controversy management, and 86% are prioritising data and GenAI investment overall. (Source: EY 2025 TFO Survey.) The same survey found that 80% of US respondents say their data is not ready for AI, and 68% of industrial tax functions consider themselves minimally or not prepared to deploy AI agents. The adoption is wide. The readiness is shallow.
Where the risk now sits
The first reported UK tax tribunal case involving AI-fabricated authorities is now precedent and worth knowing.
In Felicity Harber v HMRC [2023] UKFTT 1007 (TC), decided 4 December 2023, the appellant submitted nine fictitious First-tier Tribunal decisions in support of a reasonable-excuse argument against a CGT failure-to-notify penalty. The Tribunal found the cases had been generated by ChatGPT or a similar tool. The £3,265.11 penalty was upheld, and the decision explicitly warned that fabricated citations “waste time and public money,” reduce resources for genuine litigants, and “promote cynicism about judicial precedents.” (Source: BAILII full judgment.)
That is a 2023 case, in personal tax, brought by a litigant in person. It is not a transfer pricing case. It is also exactly the kind of case that will arrive in TP within the next few years, with much higher stakes. By the end of 2025, the Damien Charlotin AI Hallucination Cases Database had documented around 800 sanctions globally for AI-fabricated authorities across more than 25 jurisdictions. (Source: damiencharlotin.com/hallucinations.) Mata v Avianca in 2023 was the canonical legal example. There is no canonical TP example yet. There will be one before this decade ends.
The PCRT bodies have responded. On 19 January 2026, the seven professional bodies behind the Professional Conduct in Relation to Taxation framework (ICAEW, CIOT, ATT, ICAS, AAT, STEP, and IFA) issued binding topical guidance on the use of AI in tax work. It applies the five fundamental PCRT principles, integrity, objectivity, professional competence and due care, confidentiality, and professional behaviour, to AI tools used in tax practice. (Source: ICAEW summary, 19 Jan 2026; CIOT page.) The guidance does not single out transfer pricing. It applies. If you are a chartered tax adviser using AI for TP work, the guidance is binding on you from 1 January 2026.
Where the regulatory ground is moving
The TP regulatory environment in 2026 has its own pressures, independent of AI, that any practitioner read has to take seriously.
The OECD’s Side-by-Side Package, published 5 January 2026, restructured Pillar Two so that the Income Inclusion Rule and Undertaxed Profits Rule do not apply to US-headquartered MNE groups for fiscal years commencing on or after 1 January 2026. (Source: OECD Side-by-Side Package PDF.) That is a structural change in the global minimum tax framework. Twenty-two of twenty-seven EU member states have implemented the Income Inclusion Rule plus the Qualified Domestic Minimum Top-up Tax by 2025. (Source: Tax Foundation, Pillar Two Implementation in Europe.) The first GloBE Information Return filings are due 30 June 2026.
In Ireland, Revenue’s three-tier OECD documentation framework applies a Master File threshold of €250m of consolidated revenue and a Local File threshold of €50m. The penalty for missing a thirty-day Master or Local File request is €25,000 plus €100 per day. CbCR failure carries a €19,045 fixed penalty plus €2,535 per day. (Source: Revenue Part 35A guidance via PwC Ireland.) That is the operational reality your TP team is documenting against. AI helps with the volume of work that documentation requires. It does not lower the standard the documentation has to meet.
The OECD itself has not yet published TP-specific AI guidance. The closest piece is the OECD Due Diligence Guidance for Responsible AI of 19 February 2026, which is general AI governance and not transfer pricing. (Source: OECD AI.) That gap is a working assumption that will not last past 2027.
What I would do if I were running a TP function in 2026
Five things.
Get the documentation drafting tool now, but limit its use. Aibidia, Reptune, or a senior-pair-of-hands plus ChatGPT Enterprise will save your team weeks per cycle on Local File drafting. The skill the tool is following has to be reviewed by a chartered tax adviser. The output cannot be lifted into the file. The drafter on the team is the person responsible, not the agent.
Insist on a verification layer for any citation produced by an AI tool. Every authority cited in a defence brief, an APA submission, or an audit response has to be verified by a human against the underlying source. Harber makes this not a courtesy but a professional obligation. The PCRT topical guidance makes it the default standard for tax advisers.
Document the AI use in your TP file. When the tax authority asks (and they will, within three years), the function that can show what AI tool produced what part of the file, when, on what data, with what review, is the function whose documentation holds. The audit trail standard for AI-assisted TP work is the same standard as for any other AI-assisted finance work I have covered in the continuous attestation post and the CFO 11 questions post. Build it on day one.
Use AI for the work it is good at, not the work that looks impressive. Comparables filtering, NACE code application, draft commentary on routine controlled transactions, populating Local File templates. Not the functional analysis on the high-value transaction. Not the value chain analysis on the IP-rich structure. Not the litigation strategy. Those are still judgement work and the audit trail your team builds is the audit trail you defend in tribunal.
Watch the regulatory ground shift. The OECD will produce TP-specific AI guidance in 2027 at the latest. HMRC, Revenue, and the IRS will follow within twelve months of that. The tax authorities are themselves using AI for risk assessment of taxpayers, though none has published the methodology. The taxpayer that has documented its own AI use carefully will be in a stronger position than the one that has not, regardless of where the regulatory ground lands.
Where this lands
Transfer pricing has historically been a discipline whose technology stack lagged the broader finance function by five to ten years. The AI shift may close that gap faster than the previous waves did, because the work is dense in documentation, language, and structured data, and that is exactly where this generation of tools is strongest.
The TP function that adopts thoughtfully, builds the audit trail, and keeps the substantive judgement with chartered professionals will do more work faster, with the same quality, and the same defensibility. The TP function that lifts AI output into the file without review will have its Harber moment within the next eighteen months, and it will not be a £3,265 case.
Choose which one yours is going to be.
Maebh Collins is a Fellow Chartered Accountant (FCA, ICAEW) with Big 4 training and twenty years of operational experience as a founder and senior finance leader. Earlier in her Big 4 career she delivered a £125m UK transfer pricing reduction for a major PLC across more than 1,000 group entities.