Multinationals sharing proprietary AI agents among their subsidiaries risk the possibility of more disputes with tax authorities as questions arise about how to price that sharing.

The fact that any person with access to the technology can make or add to an agent raises issues about where the value of the agent should be assigned, especially because there is no law or international standard governing the transfer pricing — valuation of affiliate transactions — of AI.

Companies may find themselves in a situation where multiple countries claim they’re owed more tax because they think the value of an AI agent was created in their jurisdiction.

An AI agent is software coded to perform autonomous tasks such as collecting and analyzing data or researching and writing memos.

A fundamental issue is whether sharing the artificial intelligence agent is sale of a service, or sale or licensing of intellectual property, which is usually of higher value. The determination isn’t always easy, but is critical under transfer pricing rules, which govern intragroup transactions and have enormous implications for profit allocation and tax bills.

For example, Grant Thornton partner Samit Shah said, a sales team dispersed through 30 countries contributes to development of an agent that finds new leads to sell products.

“You’ve got a collective team that has built an agent that’s used around the world. Is this a shared service allocation? Is this a cost share?” he said, referring to ways companies allocate costs under transfer pricing rules.

“That’s one of my major questions,” he said. “Is this truly IP development or is this a service? Those are two different things.”

And even if that distinction is made, allocating the value, profit, and tax due to the correct jurisdiction is difficult.

Companies must be able to figure out how much profit each entity deserves, Shah said. If governments are left to their own devices, “they’re going to claim the biggest piece” of profit they can.

Back to Basics

In the absence of guidance, practitioners said they’re going back to transfer pricing basics: looking at companies’ AI investments, the functions the technology performs, and how it benefits the business.

Some of the investments companies are making in AI are routine and “enhance capabilities that the company may already have,” said Brian Burt, a leader in PwC’s transfer pricing practice. He cited an example of an agent that’s used to spot new consumer trends.

In that case, a services categorization is warranted, he said.

But other investments in AI may provide a “new channel to business” or “new channel to market,” Burt said, potentially warranting designation as an IP transaction.

The difference matters. Services — IT assistance, software use, and legal advice exchanged for a fee — are generally valued lower than IP, where a royalty payment or licensing fee is due to the affiliate designated as the IP’s owner.

Firas Zebian, a leader at Deloitte’s US transfer pricing and value chain alignment group, said a subsidiary’s ability to change the model impacts how transactions are characterized.

Like Burt, Zebian said many of the AI transactions he sees at the moment are SaaS-like — providing software as a service. The parent company shares a “locked” AI model with a subsidiary in another jurisdiction, but the subsidiary can’t retrain the agent.

But if affiliates share an agent that’s customizable, where coding can be changed, this activity “can be close to a license” for IP, he said.

Business Impact

Burt said it’s important for companies to look at the impact AI investment has on the business to accurately characterize the transaction.

“Is it increased revenue? Is it decreased cost? Is it just back-office efficiency? Is it speed to market?”

Answers will help a company figure out whether an agent is a separate asset, he said.

Anthony Pastore, a partner at Mayer Brown, pointed to 2025 US Treasury Department final regulations that, in the absence of AI-specific guidance, could help companies characterize the agent for pricing purposes.

The regulations, often referred to as cloud computing rules, detail the characterization of digital content transactions like the sale of books, articles, movies, and music for tax and transfer pricing purposes.

“I think there’s at least likely that the guidance that we have is flexible enough, detailed enough,” Pastore said.

AI agents may also affect the way a tax administration views a multinational company’s value chain, Shah said.

A company may contract out its research and development to a foreign affiliate, for example. If the affiliate creates an agent that can be used organization-wide, it’s performing an activity that’s outside its “routine” R&D work, Shah said. A tax administration could attribute more value to the foreign affiliate and expect a higher tax payment, he added.

To prevent disputes, both Zebian and Burt said, companies should always be aware of how AI is being used in their business.

“A lot of times I think someone might develop an agent or be using AI without the tax function knowing,” Zebian said. If an employee creates an agent with significant value, it puts the company at risk of audits, disputes, and an additional tax bill.