AI Prompt Engineering and the U.N. Model Treaty

Introduction

Artificial intelligence (AI) prompt engineering is the kind of job your kids will chuckle about in the future (much like we laugh today about brick-sized Motorolas that used to be the talk of the town back in the late 1990s). Regardless, it is what the CEO of Nvidia, Jensen Huang, said will become the role of current programmers.¹ In June 2023 founder and former CEO of Stability AI, Emad Mostaque, predicted that human programmers will cease to exist in five years.²

In the spectrum of AI’s evolution into artificial general intelligence,³ we are now at a point at which our skills are being enhanced by the Claudes, the Geminis, and the ChatGPTs of this world. This will pass — this notion that we must teach AI to do stuff so that it can propel our lives forward,⁴ which is essentially what prompt engineers do. That said, it is just the sort of role that rekindles a debate about what is considered a “technical service” for the purpose of income tax treaties and, in particular, the widespread clauses of article 12A of the U.N. model convention.⁵

To put this into context, classifying a service as “general” or “technical” in nature is a pretty important thing for tax practitioners — depending on the makeup of the treaty in question, you might either be taxed at source and in your residence jurisdiction or in your residence jurisdiction only. Courts have labored over this topic for many years before the inclusion of article 12A in the U.N. model in 2017,⁶ but there has always been a debate over the level of know-how a person would need to be able to provide a technical service. AI adds fuel to this fire by harnessing the know-how of (a growing number of) activities and being maneuvered by a human agent, one with a diminishing degree of contribution to the end result. If the service is technical in nature, does it matter that the taxpayer (the person from which the payer would be required to collect a withholding tax) does not have the know-how to perform it alone? Would the answer to this vary depending on whether the AI system was trained by the taxpayer or by itself?

In this article, we investigate the qualification of AI prompt engineering as a technical service, and for that, we resort to article 12A of the U.N. model convention (“Fees for Technical Services”) as a reference point. As a preliminary comment, though the U.N. model also features a specific provision for so-called automated digital services (article 12B),⁷ and though some of our analysis will touch upon that distinction where appropriate, it is outside our scope because it has not gained traction in bilateral income tax treaties since its inclusion in the model in 2021.

General vs. Technical Services in Tax Treaties

You kind of have to define general services by exclusion, because the cynical definition of technical services is just services that carry a degree of technical expertise. On its face, this would exclude a limited range of services, and we are only interested in the ones for which people actually pay, which means they become relevant for tax purposes. If what you are paying for carries zero technical expertise, perhaps the best question is why you are paying for it in the first place.

The distinction between general and technical services is first and foremost a creation of domestic tax policy. You do it because ideally you would like to tax imported technical services in a way that either encourages people to hire locally or nonresidents to set up shop in your jurisdiction (and bring their research and development with them).⁸ The other component is base erosion: Subsidiaries of large multinational enterprises that do not incur local expenditure with technical services will have to obtain that from some other place, typically other group entities abroad, which in turn empties the pot of corporate income left over for local taxation.⁹ This is also a concern expressed by developing countries in relation to intangibles, which is why several manifested positions on article 12(1) of the OECD model advocating for taxation of royalties at source¹⁰ and why they opt to lump royalties and technical services (as well as technical assistance), at least in the treaties before, or not inclusive of, article 12A of the U.N. model.¹¹

In practice, you will find that developing jurisdictions are not a monolith. Their views vary on how much know-how a service must require in order for it to be technical. For instance, on December 21, 2017, the Federal Revenue of Brazil issued a consultation ruling in which they said that brokerage or sales intermediation are not technical services in nature.¹² In June 2019 and again in May 2022, Colombia’s National Directorate of Taxes and Customs stated that “educational services” may be technical in nature depending on a detailed assessment of whether the provider transfers some sort of know-how to the local institute administering exams and offering certificates.¹³ Finally, in the memorandum of understanding between the United States and India,¹⁴ the one that accompanies the 1989 treaty,¹⁵ it is said that technical services include engineering, architectural, and computer software development services.

The MOU between the United States and India is a nice segue into the qualification of technical services in tax treaties (and their segregation from general services, which are usually within the scope of article 7(1) of the OECD and U.N. models if not permanent establishment-based). Article 12(4) of the 1989 India-United States tax treaty categorizes “fees for included services” into two buckets. First, (4)(a) refers to those that are “ancillary and subsidiary to the application or enjoyment of the right, property or information” for which royalties are received,¹⁶ the sort of thing other treaties would qualify as “technical assistance” (like in the 1991 treaty between Mozambique and Portugal).¹⁷ On the other hand, (4)(b) refers to services that “make available technical knowledge, experience, skill, know-how, or processes, or consist of the development and transfer of a technical plan or technical design.”¹⁸ That is the bucket of technical services in nature, so much so that, in the MOU, it is said that “the fact that the provision of the service may require technical input by the person providing the service does not per se mean that technical knowledge, skills, etc. are made available to the person purchasing the service, within the meaning of technical services.”¹⁹

That is not the position of Brazil in Normative Instruction 1,455 of 2014,²⁰ a position that the country individually applies on tax treaties signed with other countries, some of them developing jurisdictions like South Africa²¹ and Turkey.²² The Normative Instruction says that technical services are those that rely on “specialized know-how” or that involve administrative assistance or consultancy, or even those derived from “automated structures with clear technological content.”²³ Reliance on “a certain degree of complexity” and “a special knowledge of a science or art” is what the Chilean IRS used to define technical services in their Ruling 186 of 2022.²⁴

You would be right to point out that this brings us back to where we started, in the sense that technical services seem to be those that are reliant on technical expertise. The point we are making, though, is that this quasi-definition is the outcome of domestic tax policy preferences that find their way into tax treaties, meaning that there is a reason why countries choose to treat technical service providers differently from others. They are trying to differentiate between providers that perform activities dependent on little to no expertise (for example, sales intermediation as defined by Brazil, though one could think of a number of professionals in that space that only succeed because of their experience and know-how) from others that have devoted their time and resources to become the designers, the engineers, the mathematicians, and the AI developers that they are today. If you are not satisfied with the rationale of countries using tax policy to lure these experts in, you can look at it from the viewpoint of equality, perhaps the most basic principle of tax law in democratic societies. If “persons in equal circumstances [are to] be treated equally,”²⁵ there is an argument to be made that tax treaties should not offer the same standard of treatment to unequal service providers, who are the actual taxpayers in a withholding tax scenario. The legerdemain we use when we speak of taxation of services hides the fact that taxes are imposed on persons (individuals or legal entities incorporated for the benefit of individuals), and it is their ability to pay them that ought to serve as a litmus test for whether they are fair and legitimate in our legal systems.

Technical Services in Models’ Commentary

This issue revolves around article 12A of the U.N. model, but the concern that produced it has been sparsely discussed by the OECD in the commentary to their own model convention. As expected, the OECD and the U.N. approach the subject in different ways.

Because the OECD model does not contain an article 12A (or its equivalent), it is under article 12 (royalties) that you will find comments on the treatment of technical services, albeit quite limited. Section 11 of the commentary to article 12(2)²⁶ of the OECD model — the provision that defines royalties under article 12 — is a mixture of observations on know-how, general or technical services, and technical assistance. If you squint hard enough, you might glean a definition of (general or technical) services as those “in which one of the parties undertakes to use the customary skills of his calling to execute work himself for the other party”²⁷ or as those that require “special knowledge, skill and expertise but not [its] transfer,”²⁸ the transfer being qualified as what they call “the supply of know-how.”²⁹ The OECD lists a few payments that would not be defined as royalties under article 12, one of which is “payments for pure technical assistance,”³⁰ but the import of the commentary is really the separation between supplying know-how and providing general or technical services without the supply of know-how, the former being the only viable candidate for the royalties provision.

The other place where the OECD provides some input on technical services is in a 2002 report titled “Treaty Characterisation Issues Arising from E-Commerce,”³¹ a part of volume II of the full version of the 2017 OECD model. This is the source of the adjustments made to the commentary in article 12 (which include those in its section 11). The report states, unsurprisingly, that “services are of technical nature when special skills or knowledge related to a technical field are required for the provision of such services.”³² It adds, more in an observational than a definitional tone, that “the provision of knowledge acquired in fields such as arts or human sciences would generally not [qualify as a technical service],”³³ which is likely as tenable as you are thinking. Engineering services would be technical in nature, but the services of a psychologist would not.³⁴

The U.N. model is much more concerned with the definition of fees for technical services as a category of payments subject to tax treaties. For comparison, while the OECD refers only eight times to technical fees and 15 times to technical services in its entire model convention (13 of which are just references to reservations or positions presented by members and nonmembers), the 2021 U.N. model cites the expression 327 times. Article 12A(2) preserves the taxing rights of source jurisdictions on income from technical services, and in article 12A(3), the U.N. simply says that these are payments “for any service of a managerial, technical or consultancy nature,” except for a few exceptions.³⁵ It is noteworthy — and this was highlighted by the Delhi High Court in India back in 2008, though in reference to their domestic law³⁶ — that technical services are grouped together with those of a “managerial [or] consultancy nature,” which points to their reliance on human ability. That is to say, the specialized know-how that a technical service must carry (for it to be labeled as technical) has to be possessed and applied by a human provider.³⁷

The U.N. commentary fleshes out the meaning of technical services subject to article 12A. It says in section 62 that technical services “must involve the application by the service provider of specialized knowledge, skill or expertise on behalf of a client or the transfer of knowledge, skill or expertise to the client” and that services of a routine nature “that do not involve the application of such specialized knowledge, skill or expertise” are not within the scope of article 12A.³⁸ The U.N. lists technical services including regulated professions (for example, accounting, architecture, engineering), but adds that other professional services (apart from those described in article 14(2) of the model) can constitute technical services as long as they entail “the provision of specialized knowledge, skill and expertise.”³⁹

Among the numerous examples of transactions potentially qualifiable as fees for technical services in the U.N. commentary, we would like to highlight the example that appears in sections 90 and 91. In the example, the U.N. refers to a company R, a resident of State R, that provides to a company S, residing in state S, the service of access to one of its various databases (which R collects, organizes, and maintains).⁴⁰ The U.N. says that the payment for this service would not be within the scope of article 12A(3) because, even though R “used its knowledge, skill and expertise in creating the database, the services that [R] provides to [S] — access to the database — are routine services that do not involve the application of [R]’s knowledge, skill and expertise for the benefit of [S].”⁴¹ The flip side of that is the remark in section 91 that, if R “created a specialized database customized for [S]’s use from information supplied by [S] or collected by [R],” then the payment made for that service would be qualified as a fee for a technical service.⁴²

What we learn from this example (and the rest of the U.N. commentary) is that service providers must apply their know-how to their deliverables (whatever the client pays them for) to be considered technical services. True, access to a database is premised on that database having been built and maintained for commercial use, which is undeniably technical in nature (it requires the application of technical expertise). But if the technical work that went into the database is not a part of the service agreement itself (the database is ready for use, and it is not tailor-made to fit the client’s specific demands), then offering it as a service is an activity lacking specialized know-how. This is tied to the notion of technical services relying on the application of human expertise — the separate treatment that they command under article 12A is based on that, not on how much technology went into a routine service that is now a finished product in and of itself.

AI Prompt Engineering: General or Technical Service?

Before we discuss prompt engineering, perhaps we should take a step back and explain prompts in the realm of generative AI. Generative AI is a form of technology that leverages deep learning models to generate human-like content (images, words) as a response to instructions.⁴³ These instructions are what we call prompts, the sort of thing you would type on Discord for Midjourney to draw an image of a capybara paying taxes or the OECD and the U.N. pulling the same rope in a literal tug-of-war. A prompt is the input users give to AI, which is trained on large datasets and learns to generate text, or images based on them, as its output.⁴⁴

Prompting is just giving instructions to AI — and recently launched large language models enable users to give those instructions to AI in different ways, like typing, showing images, or even using their voice.⁴⁵ Prompt engineering, on the other hand, is described as a technique for enhancing the capabilities of large language models: It involves the strategic design of task-specific instructions to guide model output (or the desired output) without altering parameters.⁴⁶ It is, in the words of James Phoenix and Mike Taylor, “the process of discovering prompts that reliably yield useful or desired results.”⁴⁷ Instead of asking ChatGPT for a list of names for a given product, you specify whether the name you are looking for needs to be in English, what sort of name would be a good name in your opinion, what style or attributes the name you are looking for should have, and so on.⁴⁸

From a distance, prompt engineering looks like art.⁴⁹ It is a sort of linguistic lock-pick that people use to get AI to do exactly what they want. Commercially, however, the sort of thing you see presented as prompt engineering feels like common sense at times (so closer to just prompting). If you type “prompt engineering” into Amazon’s Kindle Store search box, over 1,000 results will appear. Some, like the instant classic The Only ChatGPT Prompts Book You’ll Ever Need, written by the author GPT Penguin, teach you that one should start prompts “in a way that places the most important and attention-grabbing information first, followed by progressively less critical details.”⁵⁰ The author later highlights that you can “assign roles or personas to ChatGPT, encouraging it to provide a deeper answer from the function, perspective, and point of view of the assigned persona,”⁵¹ which is literally what you find in OpenAI’s ChatGPT prompting guide.⁵²

This has an effect on the qualification of prompt engineering as a service, which is not the main focus of this article — we are more concerned about the effect that prompt engineering has on the qualification of other services in light of article 12A of the U.N. model. That said, if someone provides you with a prompt engineering service that is no more than just prompting, aside from the more important question of why you are paying for that in the first place, it is hard to qualify your payment as a fee for a technical service. The service would have to carry some sort of specialized knowledge so that it would become in-scope for article 12A, and even then, it would have to be exclusively focused on instruction design, meaning someone who prepares a batch of prompts that you will then use to interact with AI yourself. Section 65 of the commentary to article 12A exemplifies technical services as those provided by academics, as long as they “involve the provision of specialized knowledge, skill and expertise,” which proper value-adding prompt engineering would likely do.⁵³ This is not to say that prompt engineers are similar to academics, but the tone of section 65 seems to be that the provision of academic knowledge (which includes linguistic knowledge) can be a technical service. Viewed in a different light, AI like Midjourney can be quite limited in terms of the images it produces if it is used by someone without effective prompting skills — if that person hired a professional graphic designer to help them materialize their vision through Midjourney, that batch of tailor-made, client-specific instructions would naturally be technical in nature (otherwise the user would never have hired that person with that skill set).

If AI users find themselves in a position to hire someone to help them with their prompts, it is likely that the provider of this solution will be involved in the end product — the deliverable the user is expecting to receive from AI. To give an example, the recently released ChatGPT-4o (o standing for omni model) can code entire games from a still image on the web.⁵⁴ Suppose that a person without any coding capabilities wishes to build a sophisticated game using ChatGPT-4o. They watch videos on YouTube about it then try prompting the platform themselves, but to little avail. What they can do instead is hire someone who will prompt tailor-made instructions on their behalf so that the end result, the thing they hired the provider for, is the game itself. This would not be prompt engineering as a service, but a genuine form of game development — the provider used ChatGPT-4o instead of other software, but they were hired to deliver to the client a game, not a batch of specialized instructions to build the same game.

We will come back to the game development example later, but it should be clear by now that, in the context of a tax treaty featuring article 12A of the U.N. model, if someone pays for mere prompting (even disguised as prompt engineering, and good luck trying to find a bright line between the two), that should not be considered a payment for a technical service. Its appropriate treatment depends on other factual circumstances, but if it is provided by an enterprise in the other contracting state, it might be subject to article 7(1) (thus denying taxing rights to the jurisdiction of source unless a PE is present). If, on the other hand, someone pays for the specialized craft of tailoring the instructions given to AI so that it produces the exact output desired by the client, and if that payment is specifically made for the batch of instructions as opposed to the end product resulting from them, that would be just the sort of payment covered by article 12A(3), given that it is premised on the application of knowledge, skill, or expertise as outlined in section 62 of the respective U.N. commentary.

Embedded AI Prompt Engineering

The treatment of prompt engineering in the context of tax treaties starts to become more complicated when it is in the backstage of the service offered by the relevant provider. For instance, think of someone who sells their services as an architect to clients in many countries. They receive the appropriate instructions via email, and unbeknownst to their clients, they just feed those instructions as prompts to a powerful AI platform. It produces the output required by the client, the client pays the architect for the service, and that ends their business relationship. Just for the sake of comparison, suppose the same client hired an actual architect that employed their own skills and expertise to produce the deliverables they were looking for and paid the same price for their services.

Inequality alone should make you bristle at the idea of taxing these two sorts of payments as if they were the same. The AI “architect” and the actual architect are not in the same position, have not invested the same time and resources to present themselves as architects to the public, and do not have the same expertise — and here we are referring to the actual taxpayers in question, not to the AI used by the first provider, which is not a taxpayer (to date).⁵⁵ If you look at it from the broader angle of tax policy, of trying to attract certain kinds of skilled professionals by erecting strategically placed tax walls, then equal taxation of these two providers becomes even less tenable because the skills used by the first professional are not their own — the skills were borrowed from AI which, in this example, the professional did not develop.

The challenge is that prompt engineering can be a nuanced expression referring to everything from just typing commands to essentially teaching AI to do precisely, exactly, what you want it to do (we would not call prompting the same as prompt engineering, but here we are alluding to its commercial applications as we see them today). Below we walk through some examples that explore these nuances, all with a focus on whether they modify the character of the service provider as a technical service provider under article 12A of the U.N. model.

AI Incapable of Providing the Service on Its Own

If AI is used as a tool more than as the engine that produces the end result the client is looking for, then its participation in the service should not alter the character of the human service provider. In other words, if the human professional in question has the know-how to provide the service the client hired them for, the fact that they used AI as no more than just a tool should not change their technical service into a general service.

Consider this example. Suppose that the second architect we referred to above notices that they are losing market share to “architects” who are no more than AI-powered prompt engineers. The architect denounces these individuals to the relevant regulatory board, but they take the longest time to act — the only choice left for them is to compete. They start using Midjourney to produce the first drafts of floor plans and 3D renderings that they then refine to provide to their clients. In this scenario, Midjourney is used as just a tool; the architect has to filter the information provided by the client and use their own expertise to give accurate instructions to Midjourney, which then generates images that are maybe 50 percent of the complete deliverable that the client expects from the architect (and they work on the images produced by Midjourney to arrive at the 100 percent they were hired to provide).

This is a clear situation of a human professional using technology to enhance their offering to their clients. AI in this case is not capable of providing the service on its own, and here we distinguish between “a” service that is typically required of an architect and “the” service that the architect using Midjourney was able to provide. Midjourney is unable to provide the exact service (or standard of service) that the architect provided, if nothing else because it did not have an interaction with the client — its only access to the client’s wishes and guidelines has been filtered by the person using it as a tool. You could say that the architect acted as a prompt engineer, but one, prompt engineering is not what the client hired them for, and two, they would have been able to do the thing they were hired for with or without AI.

We would qualify this as a technical service under article 12A(3), unless of course the client is an individual and the services they are paying for are for their personal use (considering the exception in article 12A(3)(c)).⁵⁶ Notice that architecture is listed as a technical service in section 64 of the U.N. commentary, but the taxable subject — always an important reminder — is not the service but the service provider. That is why section 62 states that the service must involve the application of specialized knowledge, skill, or expertise “by the service provider,” the person from whom the source has to collect a withholding tax if enabled by article 12A(2). If we go back to the 2002 OECD report on e-commerce, it says in section 38 that while “special skill or knowledge may be used in developing or creating inputs to a service business,” the fee for the provision of a service “will not be a technical fee . . . unless that special skill or knowledge is required when the service is provided to the customer.”⁵⁷ In our example, the special skill is possessed and applied by the human provider — not by AI, and certainly not by AI without the architect’s efforts — for the benefit of the client.

AI Provides the Entire Service on Its Own

If the example above is one end of the spectrum, this is the other. If AI is the sole provider of the service hired by the client, and if the human professional that uses AI simply pastes the client’s instructions onto its platform and copies the result back from the platform to them, their service would not be a technical service under article 12A of the U.N. model.

Suppose someone offers their services to a client as a game developer, but they themselves have no expertise in the field. The client sends them detailed instructions of what kind of game they are looking for via email. The game developer receives it, shares it with ChatGPT-4o, and the engine produces a beautiful, functional little game that is exactly what the client was hoping to receive. If there is a treaty featuring article 12A of the U.N. model or its equivalent between the country of residence of the developer and the country of residence of the client, and if there is a discussion on whether the payment for the game should be subject to article 7(1) of the OECD or U.N. model or its equivalent (as a general service) or 12A (as a technical service), we find that it should be subject to article 7(1).

This seems simple enough, but it is premised on the client paying for a thing that in theory they could attain on their own (or pay a license fee to an AI platform that is bound to be far cheaper than what the developer charged). Should it matter that the client believed they were paying for an expert to produce something using their own skill and expertise? Not really — at least not for tax purposes. What matters is whether the provider applied their own specialized knowledge, skill, or expertise to provide the service that the client (well-informed or otherwise) paid for, and in our example, they did not. What about legal entities that provide technical services? They are not using their own knowledge or expertise, and they are not human professionals themselves. Well, if the human professionals that carry out the service are employed or otherwise hired by the provider as a legal entity, the fact remains that human expertise is being used “when the service is provided to the customer.”⁵⁸ So it should not matter that the person invoicing the client (the legal entity provider) is not the actual human that carried out the service for which it is being paid.⁵⁹

There are a couple of counterarguments one could pose to this assessment. A first counter is that the large language models we use as AI today learn from human knowledge. In other words, if we say to ChatGPT-4o that we want it to work as a game developer, its notion of what a game developer is will be entirely based on its dataset, a dataset that was curated by humans. As a response to that, we could invoke section 38 of the 2002 OECD report, which we already cited. The better response, though, is that payment for a capsule containing or manifesting human knowledge is not a payment for a service, and it should not matter whether the knowledge in question is specialized or not. The taxpayer of our example is the provider, the human user of the AI engine that produced the game — they have contributed nothing to the end result of the service they sold to the client, and the know-how that went into it has no connection whatsoever to them. They might be taxed by their residence jurisdiction or at source via other provisions of the relevant treaty — but it is not defensible to use article 12A of the U.N. model for that purpose.

The other counter is that the human professional in our example might have built the whole AI engine together with hired game developers so that it basically automates tasks. The impulse here is to answer that, in this case, the client is hiring the service of “a” human developer (not the provider) and that the AI engine is just a tool — but this would be premature. First, if the provider does not have specialized know-how and has not applied it to the service in question, the payment received should not be a payment made for a technical service. Second, unlike the example with the architect, here the “developer” is not guiding the AI engine while providing the service to the client. The “developer” is sitting back, relaxing, and watching the robot do all the work from start to finish — in our example the “developer” is not fiddling with the result, just copy-pasting it to an email, sending it to the client, and collecting the money.

Finally, even if you are not a fan of article 12B of the U.N. model (and you would be in good company),⁶⁰ it is fair to say that there should be a theoretical segregation between technical services and automated digital services.⁶¹ Some Indian court rulings have distinguished between the two,⁶² while Brazil, as we indicated earlier, lumps both kinds of services together.⁶³ That said, if we go back to the tax policy rationale of taxing nonresident technical service providers, it makes sense to separate those from the ones that merely pay a license fee to an AI engine and make money off of its work. If your taxes lure the former providers into your country, they might bring expertise with them — that rationale makes no sense if the providers you are bringing in are in fact clients of (and wholly dependent on) AI.

AI Trained by the Prompt Engineer

This is the midpoint between the two scenarios we analyzed above, meaning that here the prompt engineer is a significant contributor to the capabilities of the AI that are then employed to provide the services. The prompt engineer trains the AI to become the service provider, in a sense. Unlike the first scenario, the AI does most, if not all, the work. Unlike the second scenario, the AI can only provide the service because it was trained by the human professional selling it to the client.

Training is a broad term, so let’s start the analysis with the type of training that places the human service provider as close as possible to the client. Think of a tax professional that offers the service of preparation of corporate tax returns. A client with offshore bank accounts and crypto wallets approaches the provider and asks for their income tax return preparation for the year. The provider then trains an AI platform like Claude 3 Opus⁶⁴ to read and allocate financial information received from the client into buckets of offshore earnings, earnings subject to foreign taxes, cryptoassets treated as currency or financial investments, and so on (all with references to legal provisions in the relevant jurisdictions, which are furnished to Claude by the human professional). The professional also shares the template of the corporate income tax return in the jurisdiction of the client’s residence with Claude so that it can fill out the appropriate fields with the data that the client provided. The resulting tax return is prepared by Claude and delivered by the professional to the client, which then is expected to pay a fee.

In this case, if the client is in one treaty jurisdiction and the professional is in another, and if the treaty in question has a provision like article 12A of the U.N. model, it makes sense for the payment to be treated as a fee for a technical service. There is little doubt that in our example the tax professional used expertise for the benefit not of any client, but of this client and taking into account their individual data and situation. This is not a service of a “routine nature”⁶⁵ — it involves, to borrow an expression used by the U.N. in the commentary to article 12B (listing exceptions to automated digital services), “significant customization.”⁶⁶

The reason this example deserves the same treatment as the example explained above is that, while here the AI engine is not acting as just a tool, its capabilities to provide the required service were engineered by the human provider and in connection with the client’s specific guidelines (“when the service [was] provided to the customer”).⁶⁷ Treating this as a provision of a general service just because AI was involved would be tantamount to qualifying software-based services as general simply because of their reliance on software.

To make the example a bit more complex, consider this: Two weeks after the first client receives the corporate tax return from the human professional, a second client approaches asking for the same service. Suppose the new client has strikingly similar assets domestically and abroad, including crypto wallets, to the point that the provider just says to Claude to repeat the work it did for the first client, which it does. The trouble with this modified example is that now we are pushing the human professional further away from the client, even if you could argue that the data and situation are similar to those of the first client. This is a case in which qualification under article 12A would have to rely on a facts and circumstances analysis: If it can be proven that the tax professional somehow monitored Claude’s work and “tweaked” the corporate tax return that it produced, perhaps to correct a mistake or strengthen the tax position of the second client, then this could be a candidate for article 12A. If that is not present in the interaction between the professional and Claude (as evidenced by the chat history between the two),⁶⁸ then we could be talking about article 7(1) or, as we discuss in the next example, article 12 of the relevant treaty.

To really test the reach of article 12A of the U.N. model, we could think of an even more complex example. Forget Claude and suppose that the tax professional in question learns how to code an AI engine. The tax professional feeds links to updated tax legislation, jurisprudence, and scholarly contributions to the point that the AI becomes a skilled tax professional. Of course, if the tax professional sells direct access to the AI engine, the corresponding payment should be treated as a license fee and likely be subject to article 12 of the OECD or the U.N. model, assuming that it is featured in the relevant tax treaty. But what if the tax professional sells the tax return preparation services as if AI were not involved in providing the service? Would the level of AI use be enough to displace the tax professional from the scope of article 12A? Notice that unlike the earlier scenario, here the only person who developed the AI is an expert in the field and is the formal service provider.

In our view, the short answer is yes. There is no synchronicity between the efforts of the tax professional who developed the AI engine and the services that it (the AI engine, indirectly) provides to clients. In substance, the client is paying for the right to use some software, which should be a royalty and not a fee for a technical service (and it does not matter — for tax purposes — whether the client believed it was hiring a service from a human provider). This could emerge as a tax controversy if the treaty in question differentiates between the tax treatment of royalties and fees for technical services (for example, if its article 12 or equivalent imposes a higher or lower withholding tax on royalties than it does on fees for technical services in 12A), and then tax authorities and the withholding agent would have to discuss how far removed from the service the AI operator actually is for the service provider status to be either validated or compromised under the terms of the treaty.

Final Remarks

Taxes are imposed on taxpayers, either human beings or entities led by human beings. The qualification of services as technical services in tax treaties must be mindful of the human beings who provide them. Those in a similar position should be treated in a similar manner. The point of this article is to highlight that the usage of AI can affect that status, that notion of equality between different service providers. It is our concern for fairness and for the taxpayers’ ability to pay that guides our assessment of how to accommodate AI in our tax systems.⁶⁹

It is noteworthy that article 12A is featured in the U.N. model and not in the OECD model — we discussed how this has to do with developing countries’ positions on preserving source taxing rights on payments for technical services. Importers of technical services tend to be developing countries anyway, but what this means is that it is probably the case that the country in which the AI-powered provider will reside if or when such a conflict on qualification of services emerges is also going to be a developing country (it is far more common to see developing countries signing treaties featuring article 12A of the U.N. model among themselves, rather than with their developed country counterparts).⁷⁰ One of the chief complaints of U.N. member states that favored the insertion of article 12A in the model was that they lacked the administrative capacity to control or limit the base erosion caused by payments for technical services, so it will be interesting to see whether they are able to handle cases in which AI is used to provide a service offered to the public by a resident person.

Though it could be argued that most AI engines or platforms will be offered by companies residing in developed countries, the examples and analyses we laid out in this article are agnostic as to the residence of the company that owns the AI engine used by the service provider (if, of course, AI is in the backstage of the service offered by the human professional). It could very well be the case that a person offering architectural, game development, or tax preparation services will reside in one country and pay a license fee to an AI platform owned by a company someplace else — this should not alter the individual’s qualification as a technical service provider, which should still be dependent on the level of expertise the service provider offers for the benefit of the client.