AI-generated content adds a few hurdles to transfer pricing. To illustrate them, let’s use a piece of software, developed in one region and then licensed to subsidiaries for international sale, as an example of the different approaches a company could take in determining a transfer price.
Sale price comparisons. What if a business sets the price based on what other people are selling it for? One problem is many software products are customized, so it’s hard to find two that do the exact same thing. The speed of AI-driven development makes this comparison even harder and it would require more frequent market analysis.
However, the proliferation of AI-driven development means more products to benchmark against as well. Consider also that if a business has two identical AIs but one is trained on a bigger and better dataset, the collateral it generates is more valuable.
Cost comparisons. What if you set the price based on how much it costs to produce? Unlike traditional development overhead like labor and material costs, AI coding platforms have different overhead costs to consider, such as the time developers spent on training, prototype review, and testing, for example. This cost/margin data is more closely guarded than sale prices, making it more difficult to find a fair benchmark.
Margin comparisons. What if you set the transfer price based on how much profit other similar products capture? Most software expenses need to be amortized because the value of the software will likely diminish over time. But an AI application’s useful life is more fluid than that of a spreadsheet or email app. AI-produced software could have value indefinitely because it could constantly rewrite itself. With more software being written by software, acquisition costs will change as well.
Economic substance. If the revised pricing moves more profit to lower-tax jurisdictions, comparisons and assumptions must be well documented to ensure that the profit assigned to each step in the chain matches the value created. It’s not enough to cross legal T’s and dot tax rule I’s. Transfer pricing must be based on the substantive contributions or risks assumed by the parties involved. Deciding which party bears which responsibilities becomes critical in determining economic substance.
Documenting AI transfer prices raises a host of questions: What other products or services did you scope out? Did your definition of gross or net margin match that of your benchmarks? How are you considering what each related entity is contributing to the finished product? Every step of the way must be clearly laid out.
Modernizing Transfer Pricing
While businesses grapple with transferring AI generated assets, laws need to adapt to the realities of the technology.
From one viewpoint, AI created materials are expensive.There are costs to develop a model, to obtain data used to train the model,to hire developers to oversee training, and to purchase the energy and hardware to run the technology. However, the cost can be negligible: Typing a prompt into a free ChatGPT account and getting an output might suffice.
What AI can produce, in traditional terms, could be worth a pittance or a fortune.
This means that similarly situated companies can set wildly different transfer prices that could all, arguably, be valid. This makes it difficult for businesses trying to do the right thing to know what to do.
It behooves tax authorities to update their laws not just to stop abusers, but to also make the correct approach clear.
As a start, Congress and the IRS should define how far back a taxpayer needs to go in capturing the cost to train the AI and provide more guidance on amortization. If an AI is “learning” and constantly getting better, is its value really diminishing, or is it just shrinking at a slower rate? These questions are easier to ask than to answer.
Absent specific rules on how to value AI-generated output, giving an honest estimate of where value is added in the chain, as well as thoroughly documenting every assumption and allocation, is the best course of action. These transfer pricing principles apply to both the work products of the past and future—however advanced they may be.
