Taxing the Thing That Isn't There

Four Sets of Hands, Four Tax Events

The first cross-border opinion I signed my own name to ran to eleven pages. By the time it reached the client, four people had touched it.

An intern had spent two days pulling authorities, half of them Israeli Tax Authority circulars that were not indexed anywhere useful and had to be found by knowing where to look. A junior associate turned that pile into a first draft, badly, the way first drafts are. A senior went through it with a pen and sent it back twice. Then it came to me, and I read it against the file one more time, and I signed it.

Eleven pages, four sets of hands, about three weeks.

Every one of those people drew a salary for that work. Every salary produced income tax, withheld at source, in the month it was earned. Every salary produced social insurance contributions, employee side and employer side both. One opinion, four or five separate collection events, all of them triggered before the client had read a word of it.

Nobody designed that. It was simply what happened when professional work had to pass through people.

What the Legislators Are Reaching For

That structure is coming apart, and treasuries have noticed the hole before the profession has noticed the cause.

The proposals arriving now take several forms. Some would levy a charge on firms that deploy automated systems in roles previously held by employees. Some would tax computing capacity, or data centres, or the energy used to train and run large models. Several American states are moving faster than any of it, extending sales tax to business digital services, imposing excise charges on data collection, or stripping data centres of the exemptions ordinarily granted to production equipment.

Joseph Thorndike has been pointing out in Tax Notes this summer that none of this is new. In the depths of the Depression, a diffuse and largely leaderless movement pressed Washington to tax labour-saving machinery, and it found meaningful support before fading away without a single proposal coming to anything. Gregory Woirol, whose history Thorndike draws on, records that the professional judgment of economists at the time was that the idea had few merits. When Bill Gates revived it in 2017, Lawrence Summers dismissed it in almost the same terms. The machine changes. The argument does not.

What is worth noticing is what all of these proposals, across ninety years, are actually reaching for. They are reaching for the payslip that stopped being issued. Every one of them is a search for a proxy that will stand in for a wage that is no longer paid.

Displacement: Taxing a Comparison That Cannot Be Made

The first family of proposals asks a question that sounds answerable: did this system take a person’s place, and how many?

It is not answerable. Will Rinehart has made the point most economically, observing that a modest shift in method swings the estimated share of affected jobs from roughly forty-seven percent to roughly nine, depending on whether you count whole occupations or the tasks inside them. A recent independent proposal circulating in preprint tries to escape the trap by building a careful taxonomy of software agents graded by functional autonomy and displacement capacity. The elaborateness of the attempt is itself the evidence: you do not need a taxonomy that precise unless the underlying category refuses to hold still.

In professional services the problem is worse than definitional. There is no counterfactual headcount sitting behind the finished opinion. The work was not done by fewer people doing the same steps. It was done by a different arrangement of steps, and the question of how many juniors that arrangement replaced has no honest answer, because the arrangement was never a compressed version of the old one.

There is also a perverse incentive worth naming, because it points at something real. A displacement levy bites only when a firm concedes that the system occupies a role a person used to hold. So it rewards firms for not describing it that way.

As it happens, that is also the correct thing to do for entirely unrelated reasons. A randomised study published in Harvard Business Review this spring, run by researchers at BCG and Boston University across more than twelve hundred managers, found that when the same system was framed as an employee rather than as a tool, managers’ sense of personal accountability for the output fell by nine percentage points, while accountability attributed to the system rose by eight. Review quality fell. Escalation rose. Adoption did not improve at all.

So the fiscal logic and the professional logic point the same way, for reasons that have nothing to do with each other. That alignment is a coincidence, and coincidences in tax design do not survive the first serious round of drafting.

Data Processing: Taxing an Input That Is Now in Everything

The second family avoids the counting problem entirely by taxing throughput instead.

Jared Walczak of the Tax Foundation published an illustration of where that leads which is hard to improve on. He tracks a box of breakfast cereal from seed to checkout and finds fifteen distinct stages, every one of them running on processed data: varietal modelling, farm management software, moisture sensors on the combine, commodity pricing platforms, batch tracking in the mill, warehouse routing, shelf pricing, the point of sale itself. Tax the processing at each stage and the charge is embedded and re-embedded down the chain. The shopper pays a substantial hidden tax on a product that is nominally exempt, and nothing appears on the receipt.

Apply that to what I do. A signed tax opinion is, in the vocabulary of these statutes, the output of a data processing service. The cross-border adviser becomes a processor.

This base has the opposite failing to the first one. It is not too hard to measure. It is far too easy. It bites on volume, indifferent to whether any person stands behind the result. A firm that reviews carefully and a firm that reviews carelessly, pushing the same quantity of work through the same infrastructure, pay identically.

Why Both Proxies Fail: The Architecture Changed Shape

Both families assume that something was removed from a structure that otherwise stayed put. That is the error, and it is a structural error rather than a drafting one.

The old chain was not a staffing model. It was an architecture. Each layer produced work, and each layer was checked by the layer above it, and those were not two activities but one. The senior who marked up my junior’s draft was performing quality control and professional training in the same pass, with the same pen. The pyramid that firms complained about as a cost was also the mechanism by which the firm produced its own successors.

What replaces it is not that pyramid with fewer rungs. It is a different shape: a single pass, one accountable person at the end of it, and verification performed by systems rather than by the next person up. The layers were not thinned. They were designed out, and the checkpoints went with them, because the checkpoints were the layers.

This is why neither proxy can find what it is looking for. Displacement asks how many slots were emptied, of a structure that no longer has slots. Processing asks how much passed through, of a structure whose value was never throughput. Both are measuring a building that has been replaced rather than renovated.

McKinsey’s recent workforce study puts lawyers and accountants among the occupations it expects to be handled mainly by software agents, at the end of its spectrum with the highest automation potential, which accounts for about forty percent of current American jobs. That is a projection rather than a finding, and it should be read as one. But the direction is not in serious dispute among people who do this work.

One Pass, One Signature

Strip the chain down and ask what is left that is stable enough to attach a rule to.

Not headcount, which no longer maps onto output. Not volume, which now maps onto nearly everything and therefore distinguishes nothing. What remains, and what has not moved at all, is the signature: a licensed person taking personal responsibility for a specific conclusion given to a specific client on specific facts.

That is what the client is buying. Not the research, which is now cheap. Not the drafting, which is now cheap. The client is buying somebody who can be asked, under their own name and their own licence, why the position was taken, and who bears the consequence if the answer is wrong. That is the part that was never automated, and the part that sets the price.

What the Old Chain Carried Besides Revenue

Here is the thing I keep returning to. The training pyramid and the collection pyramid were the same pyramid.

Nobody built it that way on purpose. The structure through which a profession reproduced itself happened to be the structure through which the state collected from it. The same people carried both. Now both are collapsing at once.

The obvious objection is that the base has not gone anywhere. If a firm produces the same opinions with one person instead of four, the profit rises, and corporate tax rises with it, and so does tax on what the partners eventually take out. That is correct, and any argument that ignores it deserves to lose.

But it moves the base from the least mobile form of income to the most mobile. A salary paid to an associate in Tel Aviv does not relocate. Profit relocates, and an international tax practitioner is precisely the wrong person to try to persuade otherwise. Meanwhile the expense that replaced the salary, the purchase of computing capacity, is very often paid to a foreign supplier, so a portion of the value simply exits the taxing jurisdiction on its way through.

Two features do not migrate at all. The first is timing. Wages are taxed immediately and compulsorily, in the month earned. Profit is taxed in the entity, and the second layer is triggered on distribution, at a moment the taxpayer chooses. Integration doctrines are designed to equalise the eventual burden. They do nothing whatever to equalise the calendar.

The second is the insurance component. Social insurance contributions attach to the employment relationship. Profit does not carry them. They do not shift to some other base; they are deleted. And they are the specific contributions that fund the safety net which the displacement levies are being proposed in order to finance.

A Base That Attaches to the Signature

So I will state a position rather than hedge one.

For professional services, the right base is neither displacement nor processing. It is the signature.

Displacement is measured badly, and the measurement moves whenever the definition does. Processing is measured well and therefore lands on everything, producing exactly the cascade Walczak describes. Both tax an input. The signature is not an input. It is the thing the client actually consumes, and it is the only element of the old chain that survived the change of architecture intact.

A base tied to the signature does two things at once. It falls on the point where economic value is genuinely created rather than on the machinery upstream of it. And it makes the accountable structure the cheaper structure, which is the reverse of what both current regimes achieve.

I am not going to pretend this is easy to draft. Who qualifies as a signer, how advice given without a signature is treated, and what happens when a signature can be sourced from whichever jurisdiction treats it most kindly are all serious problems, and the last one is severe. Signatures are mobile too. I spend a good deal of my professional life on exactly how mobile things are, and I would not want to be understood as claiming otherwise. What I am claiming is narrower: of the available candidates, this is the only one that is measuring the right thing at all.

What We Built, and Why It Would Pay

I should be plain about where I stand in this, because I am not a neutral observer of it.

I build a system that produces cross-border tax opinions. It is organised so that a formal report does not reach a client until a human adviser has reviewed and approved it. The gate exists because the responsibility has to land somewhere, and a system cannot hold it.

Under a displacement levy, that system is taxable. Under a data processing tax, it is taxable. Under the base I have just argued for, it is also taxable, because there is a signature at the end of it and that is the whole point. The argument is not convenient for me, and I would distrust it if it were.

Here is the part we cannot solve, and I would rather write it down than let a reader discover it. A quality gate implemented in software checks output. It does not train anybody. We can build systems that catch errors reliably and prove that they caught them. We cannot build a system that produces the next person qualified to sign. The old chain did that as a by-product of doing something else, and no architecture I am aware of replaces it, including mine.

No One Is Coming Up Behind Me

I learned to check work because somebody checked mine, in red ink, twice, on an opinion that was not good enough the first time. That is the entire mechanism. There was never anything more sophisticated to it.

The proposals on the table are trying to recover the revenue that chain produced. None of them is trying to recover what else it produced, and I do not think the people drafting them know that there was anything else, because from the outside a professional pyramid looks like an expense.

A base that attaches to the signature would at least be pointing at the right place: at the person who is answerable, and therefore at the structure that has to keep producing such people. Whether that structure can be rebuilt deliberately, having previously existed by accident, I do not know. I know that the profession has perhaps one cohort’s worth of time to work it out, and that nobody is going to hand us the answer.

I sign my own reports. I intend to keep doing that, and to keep saying why it matters, one signed report at a time.