The Assembly Line That Ends in a Signature

Henry Ford measured his triumph in minutes. A Model T that once took twelve hours came off the line in ninety. McKinsey opens its case for the AI assembly line on precisely that number, and the number is the whole argument: speed made cheap, complexity made cheap, and — by the end of the piece — judgment made cheap and abundant.

A tax opinion is measured in the opposite unit. Not minutes. Years.

An opinion is not finished when it is delivered. It is finished — if it is ever finished — two or three years later, in a room its author may never enter: an audit, an objection hearing, a courtroom. By then the position is being read by someone whose entire job is to find its weakest seam. That reader does not ask how many hours the work took, or how fast it arrived, or how clean the prose was. The reader asks three questions, and only three. Who reached this conclusion. On what authority. And whether it holds.

I have spent years on both sides of that table — scrutinizing other people’s positions on behalf of the tax authority, and building and defending them in private practice. The thing under examination was never the speed of the work. It was whether a person was willing to stand behind it.

That is the unit of value in this domain — not a document, but a defensible position with a name attached to it. Hold onto that unit, because everything that follows turns on whether it survives the machine.

---

The assembly line reaches knowledge work

The right thing to do with a strong argument is to take it seriously, and McKinsey’s is strong. So before I say where I part from it, let me say how far I travel with it — which is most of the way.

The piece’s central move is an old one made new. Ford did not win by building a better car; he won by changing the physics of unit cost — twelve thousand cars a year at over eight hundred dollars each became two million a year at two hundred and sixty. The claim is that agentic AI now does to cognitive work what the conveyor belt did to physical work: it takes tasks that once depended on a handful of scarce experts and makes them fast, cheap, and parallel.

The sharpest idea in the article is a single phrase: artisanal overhead. Each increment of product complexity, McKinsey observes, drives a disproportionate rise in the specialized human effort needed to design, integrate, and validate the result — the cross-functional meetings, the manual reviews, the thousands of hours that do not scale. Name a better description of a cross-border tax opinion and I’ll wait. Every new entity, every treaty article in play, every second jurisdiction does not add to the cost of the work; it multiplies it. The expertise that can hold all of it at once is scarce, expensive, and — as I argued earlier in this series — economically unreachable for the mid-market companies that need it most. McKinsey has given a precise name to the exact problem ktAi was built to solve.

And the structural prescription is right, too. An orchestration layer that behaves like a conveyor belt; specialized agents working inside their own domains while a higher layer manages the flow end to end; a hybrid model that processes the routine at scale and flags the hard cases for a human. By McKinsey’s own definition, ktAi is an assembly line — an orchestrator, a set of specialized agents for Israeli law, US law, the treaty, and red-flag detection, and a synthesis layer above them. I am not going to pretend I built something other than the machine they describe. I built the machine they describe.

And this is not one team’s enthusiasm for a single piece. It is McKinsey’s consistent counsel across its recent writing — that leaders should focus the enterprise on a few domains, rewire the organization around them, and let the line scale. On all of that, I agree.

So the disagreement is not about the machine. It is about one station on it.

---

Two kinds of assembly line

Here is where I stop traveling with the article, and the gap is narrow but it is deep.

McKinsey’s assembly line is built to diverge. Read the ambition in the piece literally: a line that scales to billions of virtual workers, sixteen million concurrent and individualized logic streams, each one resolving on its own and rolling off the belt as a finished output. A loan approved. A configuration priced. A demand forecast generated. The genius of Ford’s line was that a Model T came off the end of it complete — no further signature required, the car was the product. McKinsey’s cognitive version inherits that shape. The output is autonomous and terminal. It diverges into millions of parallel results, each one done when the line is done with it.

A cross-border tax opinion is built on a line that runs the other direction. It converges.

The specialized agents in ktAi look, from a distance, like McKinsey’s functional agents — one works Israeli law, one works the US side, one works the treaty, one hunts for red flags. They run in parallel, and to that point the two machines are identical. But they do not each produce a finished output that rolls off the belt. They feed a synthesis layer that pulls four partial readings into one position — and, by design, surfaces where they disagree rather than averaging the disagreement away. The line narrows as it runs. Many agents, much retrieval, parallel analysis, all of it converging toward a single defensible position that one named professional will adopt as their own.

That is not a difference in scale or sophistication. It is a difference in the direction of the machine. Ford’s line and McKinsey’s line widen toward output. The opinion line narrows toward judgment. You cannot build the second by tuning the first, because the thing the first optimizes for — autonomous, terminal, parallel output — is the exact thing the second must refuse to produce.

---

Where “the cost of judgment falls” breaks

The article’s closing promise is the line I want to hold up to the light: when the cost of judgment falls and its availability scales, McKinsey writes, personalization becomes standard and forecasting becomes continuous. For most of the cognitive work in most companies, I think that is simply true — I argued a version of it myself at the start of this series: the mechanical layer migrates to agents, and that is good.

But “the cost of judgment” hides a second cost inside it, and in a regulated profession the two come apart. Part is cognitive — the hours of reasoning, retrieval, and cross-referencing the assembly line genuinely can drive toward zero. The other part is not cognitive at all. It is responsibility: a licensed professional adopting the position as their own and answering for it, to the tax authority, to the standards board, to the client. Who Steers the Tool, the opening essay in this series, made that case in full, and I take it as established here — responsibility cannot transfer to a system that has no name to put on the file. The cognitive cost can fall to nearly nothing. The accountability cannot fall at all, because it is not a cost of producing the opinion. It is the opinion.

What I want to add is what this does to the machine, not to the professional.

McKinsey’s hybrid model already has the right shape and reaches it from the wrong reasoning. The article describes a financial institution’s orchestration layer clearing ninety percent of loan applications on its own and routing only the most complex and ambiguous cases to a human — efficiency preserved, accountability preserved. The structure is correct. But it casts the human as an exception handler: the line runs under its own authority, and the person catches the residue it cannot. That is not what happens on a converging line. The professional is not handed the hard cases while the machine ships the rest under its own name. The line hands over one hundred percent — every position it assembles — as a draft for adoption, and ships nothing autonomously at all. The contrast is exact: McKinsey’s line clears ninety percent on its own authority; the opinion line clears zero. The human is not the catch at the end of the belt. The human is the point the belt converges toward.

None of this is theoretical at ktAi; it is the design constraint the whole line is built around. A diverging line optimizes to produce more. A converging line optimizes to produce something a named person can stand behind — which sometimes means producing nothing, and saying so. What that looks like in practice is the last station on the line.

---

The last station

Calling the professional “the point the line converges toward” is an abstraction until you say what actually arrives at that point. It is not a finished opinion. It is a position with its confidence honestly located, its authority traceable to source, and its weak seams left visible rather than smoothed over. Three properties do that work. None of them is a safety feature bolted onto an autonomous line. Each is what a converging line is for.

The first is honest confidence. A cross-border tax position is not “right” or “wrong”; it sits somewhere on a ladder the profession has used for decades — will, should, more likely than not, substantial authority, reasonable basis. An auditor recognizing a position under ASC 740 cannot book the benefit at all until it clears more likely than not. The single most dangerous thing an automated system can do in this domain is round up — to dress a substantial authority position in the confident prose of a should opinion. A diverging line is rewarded for producing clean, finished-looking output; fluency is the goal. A converging line has to do the opposite: state the position at the confidence the authority actually supports, and refuse to let the prose outrun the law. I will be plain about where this stands. Anchoring the system’s internal signals to that ladder, in the vocabulary an ASC 740 reviewer expects, is engineering we are still doing. The principle is fixed; the calibration is being earned.

The second is traceable authority. Every citation a professional signs beneath is a representation: that the source exists, that it says what the opinion claims, and that it was in force at the relevant time. A model left to write citations freely produces ones that are fluent, plausible, and sometimes wholly invented — the failure that put a major firm in front of a government client with a refund. On a converging line this cannot rest on the model’s good behavior. Citations are not free text the system emits; they are references it requests against the authority actually retrieved, and renders from that record. A reference that cannot be resolved is dropped or flagged — never dressed up as a plausible cite. The professional at the convergence point inherits a position whose every authority walks back to its source, which is the only form in which a position is worth adopting.

The third is preserved disagreement. When the Israeli reading and the US reading of one fact pattern pull apart — when a treaty residence tie-breaker resolves one way and a domestic anti-avoidance provision pulls the other — the wrong response is to average them into a single confident number. Averaging is how a machine manufactures a position no competent professional would sign: a blend faithful to neither analysis. A converging line carries the disagreement intact to the convergence point and shows it, because the disagreement is frequently the most important thing on the page. The judgment the named professional is paid for is precisely the resolution of that tension — and they cannot resolve what the machine has already dissolved.

Naming what the line does not yet do is part of that professionalism, not a retreat from it. Two kinds of gap live here, and they should never be confused. Some are unfinished engineering. The user’s own facts do not yet resolve into formal corporate structure — ownership chains with dates that would answer “show me the chain for this subsidiary” directly. The system does not yet watch an incoming document and raise an unprompted compliance flag; analysis runs when asked. A fact stated in chat is searchable but not yet distilled into the user’s graph as a durable fact later work can rely on. And the confidence calibration above is, as I said, still being mapped to ASC 740 vocabulary. These close over time. We track them, and we say what is open.

One is not a gap at all, and never will be: the signature. The line is engineered so that no position it assembles is ever issued under its own authority — not because the capability is hard, but because issuing it would destroy the only thing that makes the output an opinion. The first kind of gap shrinks with every sprint. The second is load-bearing. A system that someday “closed” it would not be a better tax engine. It would be a different and worse thing — a machine producing documents shaped like opinions that no one is answerable for.

---

The signature

Return to the room from the beginning — the audit, the objection hearing, the courtroom, two or three years after the work was done. Everything McKinsey promises the assembly line will deliver is real and worth having: artisanal overhead compressed, cross-referencing driven toward zero, a position assembled in minutes that once took a team weeks. ktAi is built to deliver exactly that, and across this series I have argued it should.

But notice what the room never asks. It does not ask how fast the opinion was produced, or how many concurrent logic streams resolved to reach it, or whether the line behind it could scale to billions of virtual workers. It asks who concluded this, on what authority, and whether it holds. McKinsey closes by urging leaders to redesign the organization so that, in its words, “the speed of operations matches the speed of intelligence.” In this domain the binding constraint is the other one. The work can be made as fast as intelligence allows. The moment a licensed professional puts a name to it and agrees to answer for it cannot be accelerated, automated, or averaged — and an assembly line that forgets this does not produce cheaper opinions. It produces faster documents that turn out, in the room, to be worth nothing.

So the line is built to run the other way: many agents narrowing to one position, prepared in full and handed to a person to own. Whether that architecture holds up under real client work, against real authority, in the room where it eventually gets read — that is not a question a manifesto can settle. It is an answer this company earns one report at a time.