Minimal Human Involvement

Two invoices are issued in the same month, both for work done across a border.

The first goes from an advisory firm to a client abroad, for an opinion: forty pages on whether a reorganization the client is planning will hold up in both countries it touches. The second goes to the firm itself, from a software company in another country whose platform the firm used to research and draft that opinion. For most of the history of international tax, nobody would have thought to put the two side by side. One is professional advice, bought from people. The other is a digital service, bought from a system. They live in different chapters of every textbook.

The United Nations is drafting a treaty that would put both kinds of service to the same test. For a service delivered online, the test comes down to a single question: how much human involvement did it take to deliver this? The answer helps decide which country may tax the income. I think it is a reasonable test. I also think it rests on an assumption about the first invoice that is quietly ceasing to be true.

---

A Line Drawn Through Services

The process began with a General Assembly resolution in December 2023, which set up a committee to draft terms of reference; the Assembly adopted them a year later. They call for a framework convention on international tax cooperation and two early protocols, with final texts due back to the Assembly in 2027. The first protocol covers the taxation of income from cross-border services. Its latest published draft, released by the co-lead in July ahead of the negotiating committee’s fifth session in New York, is a working text in the full sense. Its rates are blank, and it has no commentary. When the session closed in August, delegates were still divided over, among other things, whether physical presence should matter at all. The next session meets in Nairobi at the end of November.

Before going further, I should say that neither of the two countries I work between is likely to be bound by any of this soon. The United States walked out of the process on the first day of the organizational session in February 2025. Israel voted against the 2023 resolution and, at the organizational session, objected to deciding tax matters by majority vote, while saying it remained committed to engaging. So this essay is not about my clients’ tax bills. It is about a definition, and about what the definition assumes.

The draft sorts cross-border payments into categories, and two of them carry the argument. Fees for services covers any payment in consideration for any service, other than an employee’s pay. Income from automated digital services covers income from any service provided over the Internet or another electronic network “requiring minimal human involvement from the service provider.” Its list of examples is the digital economy as it looked a few years ago: online advertising, search engines, intermediation platforms, social media, digital content, gaming, cloud computing, standardized online teaching.

The categories are not just labels. Each gets its own cap on the tax the source country may charge on the gross payment, both caps still blank, and each gets its own rule for where the income arises. For ordinary services the rule starts from where the work is physically performed, and turns to where the customer lives only if that first test is not met. For automated digital services it starts from the customer. And where a payment could fit both, the digital rule wins, because the services article makes its own source-country rule expressly subject to it.

So the draft makes human involvement the hinge. A service that people deliver is placed, first, where the people are. A service that needs almost none of them is placed where the customer is. How that ordering applies to a professional who sits in one country and advises a client in another is part of what the August debate over physical presence left open. That human involvement decides which of the two rules applies is not open at all. It is written into the definition.

---

Where the Words Came From

The phrase is not new. The protocol copies the definition word for word from Article 12B of the UN Model Tax Convention, which the UN’s tax committee added in 2021. The commentary written for that article explains the phrase in a way the protocol, which has no commentary of its own yet, does not.

A service is automated, the commentary says, when the customer obtains it through equipment and systems already in place, rather than through a bespoke interaction with the supplier. The test looks only at the supplier, and only at delivery. Human effort spent building the system, maintaining it, or fixing its errors does not count, and neither does the customer’s own effort: a user who enters parameters to get a customized result is still using an automated service. What counts is the human response each individual request requires. The commentary is candid that the point is economic. It is trying to pick out businesses for which the marginal cost of serving one more customer, measured in additional human involvement, is nil or almost nil, as against businesses whose per-unit costs rise in proportion to each new client.

Then, to avoid uncertainty, it names what falls outside. First on its list are customized professional services: legal, accounting, architecture, engineering, medical, financial and other specialized expert consultancy. Such services may be delivered online, the commentary concedes, and its examples include legal advice sent by email. They are excluded for three linked reasons: each must be tailored to the client through professional judgment and bespoke interaction; that takes more than minimal human involvement; and, in the commentary’s words, they “would also not be scalable without additional human involvement.”

The protocol carries the definition but not, as yet, the commentary. Its text says nothing about professional services. The exclusion that kept my profession out of the digital category lives in the commentary to a different instrument, written five years ago. That commentary did anticipate change; it said the general definition could accommodate new types of automated services as technology produced them. What it does not discuss, and in 2021 had less reason to, is an established type of service becoming automated.

---

The Assumption Underneath

Read the three reasons together and notice what they share. Each assumes that tailoring advice to a client is something only a person can do, and that a person has to do it again for every client. The same commentary had already accepted that a system can produce a customized result without ceasing to be automated. What kept advice on the other side of the line was not that it is customized. It was that people did the customizing, one client at a time. That is less a definition of professional work than a description of how professional work was produced.

That was true for the whole of my career. An earlier essay in this series, on proposals to tax AI, described the first cross-border opinion I signed: eleven pages, four sets of hands, about three weeks. A second client meant a second set of hours from the same people, or more people. The pyramid the profession complained about as a cost was also, in the commentary’s terms, the evidence that advice was not automated.

It is less true now, and I know because I built one of the things making it less true. On my own matters, the system I built has cut meaningful time from drafting an opinion. The retrieval, the cross-referencing, the first draft, the work that used to occupy the lower floors of the pyramid, is now largely the system’s. The human cost of each additional opinion has fallen. It has not fallen to nil. I still read every opinion against the file, test its authorities, change what I disagree with, and sign it. The judgment is still mine, which is what the commentary’s first reason requires. Its last reason is different. It asks how much additional human involvement each additional client requires, and that is exactly the quantity this generation of tools is designed to reduce. A practice can keep the first condition and drift away from the last.

The commentary even describes the edge of its own threshold. The line is not crossed, it says, where delivery involves only very limited human response to individual requests, or where the system routes the occasional complex problem to a member of staff. Describe an AI-assisted advisory practice in those terms, with a system that does most of the work and hands the hard questions to a professional, and it begins to sound like the commentary’s picture of a service that is automated after all. This is the pattern the essay on schemas traced in software: a definition made carefully for one world, meeting another.

---

Why the Drafters Chose This Test

None of this is an argument that the drafters chose badly, and I want to give their case its full weight.

The commentary’s reasoning is strong. Digital businesses can serve a market at speed and in volume without any physical presence there, and permanent establishment rules built around a fixed place of business often leave the market country with nothing to tax. A withholding tax on gross payments is something a tax administration with limited capacity can actually collect, and the commentary says so directly. It also stresses an approach that is simple for businesses and administrations alike, and a test based on how a service is delivered is one that both can apply without first agreeing on where value was created. Those are real advantages, and the countries pressing for this protocol have serious reasons to want them.

Whether the market country should be able to tax services at all is a question on which governments reasonably disagree, and the commentary itself records a sizeable minority of its own committee who argued that a market’s location does not, by itself, generate profit. I am not going to take a side in that argument. My point is narrower, and it holds whichever side is right.

A test anchored in human involvement is administrable only as long as human involvement is easy to see. For digital platforms in 2021 it was. Either a person answered your request or a server did. For professional services in 2026 it is becoming a question of degree, and of kind, inside a single engagement. The same opinion can involve a great deal of human judgment or very little, and from the outside, on the invoice, the two look identical.

---

What Moves When the People Leave

Here is what the line does, if it holds. A service that requires real human involvement is placed, first, by where the people performing it are. As the people’s share of the work shrinks, that anchor weakens. Push a professional service far enough toward minimal and it changes article. With the article, the starting point for where its income arises moves from the country where the work is done to the country where the client sits.

The caveats from the start of this essay apply here with full force, and there is one more: more than one text for this protocol is on the table. I would not claim to know how any of this resolves.

But the direction is visible in the structure. The essay on taxing AI argued that when a firm’s work stops passing through people, the domestic tax base their payslips produced goes with them. The same thing happens here, one level up. When the people leave, the place of performance leaves with them, and a treaty that locates services by their people has to find something else to hold on to. The draft has already chosen what that something is: the customer.

---

The Line Runs Through Our Own Product

I should apply this to myself, because the argument is not neutral for me.

ktAi sits on both sides of the line. The platform a firm would use to research and draft an opinion looks far more like the cloud computing services on the draft’s list than like advice. The opinion the firm signs is advice. If a protocol like this applied to the countries involved, I suspect the two invoices at the beginning of this essay would land in different articles, even though more and more of the work inside the first one is done by what the second one pays for.

What keeps the opinion on the professional side is the part this series has been defending since its first essay. Every formal report the system drafts reaches a client only after an adviser approves it. The system enforces the approval. It cannot enforce what the approval consists of. An adviser who engages with the substance is doing the kind of human work the commentary meant when it excluded professional services. An adviser who approves a file they have not read is much closer to the very limited human response the commentary says does not cross the line, and to the thin approval that the essay on the European guidance showed a regulator refusing to count as review. Until now I have argued that the difference matters for accountability. This draft suggests that one day it may also matter for which country taxes the fee.

That is not a planning point, and I would not want it read as one. A review added to change how a fee is classified invites exactly the question the commentary asks: what did the person actually contribute to this client’s result? Nor is any of this a tax position on anyone’s facts, including ours. It is a description of where a line falls, and of what is moving across it.

---

One Opinion at a Time

The Nairobi session is expected to concentrate on the convention itself and on the disputes protocol, and the definition at the center of this essay may get little attention there. It is a definition carried over from 2021, and definitions carried over tend to survive by default.

It deserves more attention than that, because it asks the question this series has been asking from the profession’s side: how much of a service has to be human for it to be what it says it is. Earlier essays asked it about accountability. This draft asks it about jurisdiction. Both turn on the same fact, and it is a fact no definition can settle in advance: what the person at the end of the work actually did.

A treaty can draw the line, and a commentary can explain it, but neither can see inside a single engagement, and a system can record an approval without knowing what went into it. That fact is decided in each engagement, by whoever signs. I intend to keep the human part of my work substantial, and to keep showing what it consists of, one opinion at a time.