AI Audits Now Read Every Word of Your Transfer Pricing Files

More than seven in ten tax authorities in OECD countries now use AI — and that single shift has changed what makes a multinational tax audit start. For most of the last decade, the safest part of a transfer pricing report was the part nobody read. Now an algorithm reads every word, in every country, in minutes.

The stakes are considerable. The world's largest groups filed a brand-new annual return for the first time in June 2026, setting out profits and taxes paid country by country. From the end of December, those returns start passing automatically from one tax authority to another. The volume of shared documentation has reached a point where weak sentences — once buried in footnotes — now surface in seconds.

What actually changed?

Transfer pricing is what one company in an international group charges another — for goods, services, brands, know-how or loans. The rule is simple: those prices should match what two unrelated businesses would have agreed. Proving it is not. Every year, in every country it operates in, a group files a long transfer pricing report setting out its prices, backed by intercompany agreements.

For years, a busy inspector skimmed that paperwork for the headline numbers. The footnotes, the vague clause, the paragraph carried over from last year's report — most of it went unread. There was too much of it and too few people to read it.

Countries have shared information through country-by-country reporting for years. What is new is the sheer volume. Every new filing is one more place for a weak sentence to hide — and one more document to check it against. As the source puts it bluntly: "The risk to multinationals hasn't grown because the rules changed. It has grown because the reading changed."

How is AI used in tax audits?

Most people's experience of AI is a chatbot that answers questions. The next step, often called agentic AI, is different. It carries out a series of tasks on its own, like a tireless team of analysts. It gathers the filings, compares them, flags what looks wrong and drafts the first letter. A person still makes the final decision, because someone has to be accountable for it.

For a tax authority, that changes how many companies it can check. Checks that once took a team of inspectors weeks can now run in hours, across every company in a group.

It also changes what gets flagged. These tools surface anomalies — and an anomaly is not the same thing as a mistake. Many of the questions that follow will be about matters that are not black and white.

The AI's role breaks down into four concrete functions:

• Choosing who to audit. AI looks for groups whose story doesn't add up: profits booked in one place, with the people and the real activity somewhere else.
• Reading the paperwork. AI reads transfer pricing reports and intercompany agreements, then checks what they claim about who does the work, who owns what and who carries the risk against what the accounts actually show.
• Joining the dots. Tax authorities increasingly link their own records with information from other government bodies, from banks and from other countries. A gap that used to sit unnoticed between two filings now shows up.
• Writing to you. AI drafts the first questions. Letters arrive sooner, ask sharper questions, and often quote your own words back to you.

Why is the weakest sentence now the biggest risk?

AI won't argue about your prices. It will find where your paperwork disagrees with itself. It doesn't need to prove a price is wrong — only to find a sentence your numbers don't support, or two documents that contradict each other.

Consider a group's regional sales company. Its transfer pricing report calls it a limited-risk distributor. In plain terms, the parent carries the risk if stock doesn't sell or customers don't pay, so the sales company earns a small, steady profit. The contract between the two says the same. Then a bad year comes. The sales company's own accounts show it writing off unsold stock and unpaid bills — and making a loss.

An inspector working through hundreds of pages might never connect that one sentence with a note in the accounts. AI connects them in seconds. The first letter quotes the company's own words back at it and asks one simple question: why is a company that "carries limited risk" carrying the losses?

The cases that cause the most trouble are rarely about aggressive pricing, according to the analysis. They are about inconsistency, and it tends to show up in the same few places:

• wording carried over from previous years after the business has moved on
• an intercompany agreement that describes one way of working while the business runs another
• vague descriptions of the services group companies charge each other for
• comparisons with independent businesses that are years out of date
• the same company described one way in one country and differently in another

These gaps are also becoming visible sooner. A growing number of countries now require invoices to be sent to the tax authority electronically as they are issued. Transactions are seen as they happen, rather than once a year.

What does AI get wrong?

Plenty — and none of it should reassure finance leaders. Transfer pricing depends on judgement about how a business really works. AI can take a sentence out of context. It can read a sensible commercial decision as a red flag. Tax authorities' own guidance treats the data they exchange as a starting point for a risk check, not as proof of anything. Oversight of how these tools are used is still catching up.

The asymmetry of effort is the real problem. A query takes a tax authority seconds to send and a company weeks to answer. Even when the company is right, it pays in management time and adviser fees, and the dispute can run for years. And if a computer model picked your company for review, you may never be told why.

Being right is no longer enough. You have to be able to show it, quickly and in writing. Your own paperwork is the only part of this you fully control.

How can finance leaders prepare?

The source sets out five concrete steps for preparing transfer pricing documentation for AI scrutiny:

• Read your own small print. Go through reports and contracts line by line, the way an AI will.
• Make your numbers match. Every report the group files, in every country, should tell the same story.
• Check intercompany agreements against reality. Agreements between group companies should describe how the business works today, not how it worked when they were signed.
• Stop copying and pasting. Update descriptions every year, and keep comparisons with independent businesses current.
• Check yourself first. Find the outliers before the tax authority does. Where the risk is concentrated, consider an advance pricing agreement (APA) — agreeing your prices with the tax authority in advance.

The rule hasn't changed — the reading has

The principle remains what it has always been: group companies should trade with each other at fair market prices. What has changed is that every word written to prove it will now be read and scrutinised.

The groups that come through this well won't be the ones with the cleverest pricing. They will be the ones whose paperwork tells the same true story everywhere it is read.