The Bottleneck Doctrine

The law

Value lives at the weak link. The weak link is the tacit task. Tacit knowledge has decodable layers. So you build a system by walking the bottleneck inward, decoding the decodable tacit layers and encoding them deterministically, and what remains is the collective-tacit floor, which rises in worth as everything around it falls.

Why it holds, theory by theory

Six ideas, from six thinkers, that lock into one law. None of them is mine. The synthesis is.

Value lives at the weak link. Chad Jones's weak-links model of growth says output is paced not by what you do best but by what is essential and hard to improve. Automate seventeen of twenty steps in a process and the whole thing is still gated by the three slow human ones, so infinitely cheap software raises total output only by software's small share of the work. William Baumol saw the same thing decades earlier as cost disease: a task that resists productivity gains, a string quartet plays the same notes with the same four musicians in 2020 as in 1800, rises in relative cost as everything around it gets cheaper. And Michael Kremer's O-ring theory, named after the single failed part that destroyed the Challenger, models production as a chain whose output is the product of its links, so the weakest link alone caps the result. Three angles, one conclusion. Value and constraint concentrate at the hardest-to-improve task, and automating the easy parts only relocates the bottleneck, it never removes it.

The weak link is the tacit task. Michael Polanyi's phrase, "we know more than we can tell," names a real thing: the skill you have but cannot fully put into words, like balancing a bike, recognizing a face, or knowing a deal is off. David Autor turned this into Polanyi's paradox for automation: a task resists being automated precisely when its knowledge is tacit, because to automate something you must write down its rules, and you cannot write rules you cannot articulate. This is the piece that explains the first. The task that is hard to improve is hard because it is tacit. Tacitness is the cause of the bottleneck.

Tacit knowledge has decodable layers. Harry Collins broke tacit knowledge into three kinds by how hard each is to codify. Relational tacit is unspoken only by circumstance, nobody bothered to write it down, but it could be, so it is fully decodable. Somatic tacit is tied to the body as a physical system, like balance, hard for a person to explain but reachable by a machine doing it a different way, through learned patterns. Collective tacit is embedded in a shifting social context, like knowing when it is acceptable to break a rule, and it is the genuinely hard floor. This is the hinge of the whole doctrine. It turns the bottleneck from a fixed wall into a graded stack, some of it decodable, some not, which is the only reason the bottleneck can be moved at all.

You can decode judgment, not just method. In 1954 Paul Meehl compared expert clinical judgment against simple statistical formulas and found the formulas usually won, a result replicated hundreds of times since. Robyn Dawes went further, showing even crude linear models beat the experts, because a model is consistent and a human is noisy, tired, and biased. Daniel Kahneman set the boundary: this holds in high-validity environments, where patterns genuinely repeat and there is feedback to learn from, and it fails in chaos. So decoding can reach past method, where to look, into judgment itself, what to conclude, by finding the observable signals that predict what the expert only feels, wherever the domain is stable enough to learn.

The residue rises in worth. Return to Baumol. As the explicit, automatable tasks fall toward zero cost, the tacit tasks that remain do not just stay valuable, they rise in relative value, because they are now the scarce and expensive part of the whole. So the human's remaining layer is not only where the bottleneck sits, it is where the money concentrates, and it grows more valuable the more you automate around it.

Externalization is the act itself. Ikujiro Nonaka and Hirotaka Takeuchi described how knowledge moves through an organization in a cycle, and named its hardest, most valuable step externalization, the conversion of tacit knowledge into explicit form. That is exactly the act of building one of these systems: taking what an expert knows but has never written down, and externalizing it into logic a system can hold and anyone can use.

Put together. Value sits at the bottleneck (Jones, Baumol, Kremer). The bottleneck is tacit (Polanyi, Autor). Tacit is layered (Collins). So decoding a layer and externalizing it (Nonaka) moves the bottleneck inward, judgment included wherever the patterns repeat (Meehl, Dawes, Kahneman), and the value follows it toward a collective floor that keeps rising as everything around it falls (Baumol). Six ideas, one law.

The design rules

1. Aim at the weak link, not the easy win. Automating what is already easy adds only that task's small share of the value. Point the system at the bottleneck, because that is where the value is.
2. Find the tacitness. The bottleneck is hard because it is tacit. The design question is always: what does the expert here know that they have never written down. That unwritten thing is the target.
3. Decode by degree. Sort the tacit. Relational (just undocumented, decode it now), somatic and method (learnable pattern, decode with data and AI), collective (accountable, social, leave it). Attack in that order.
4. Encode deterministically, keep it a glass box. Decoded knowledge becomes traceable rules, not a black box. A decoded rule you cannot inspect is a liability, not an asset.
5. Direct attention, do not replace judgment. Herbert Simon and Erik Brynjolfsson: point the human at the current weak link and let them do the residual tacit work. Augment, never imitate.
6. Move the bottleneck, then repeat. Each decode relocates the weak link one layer inward. The roadmap of any such system is just the sequence of its bottlenecks. You are never finished, you chip the next layer as data and models allow.
7. Let the human keep the floor. The accountable, in-the-room, socially embedded call is the permanent human role. Respect it. Systems that respect the floor are trusted. Systems that fake past it are the AI slop the market is learning to distrust.
8. Price sits at the residual. The human's remaining tacit layer is where value and fees concentrate, and it grows as you automate around it. The human is never removed, only relocated to a deeper, scarcer layer of their own judgment.

The floors are plural

The rules above treat collective-tacit human judgment as the floor, because in knowledge work it is usually the deepest one. But a real system has more than one floor, and the binding one is not always the human. Capital can bind. Compute and GPUs can bind. Regulation can bind, it is why radiology's AI adoption lagged for years even after the reads were solved. Data can bind, the corpus you need before a layer will decode. Trust and adoption can bind, the human's willingness to hand a task over. Distribution can bind.

Two ideas make this rigorous. Eliyahu Goldratt's Theory of Constraints says a system's throughput is set by its single binding constraint, and the discipline is to find it, relieve it, then watch the constraint move to the next one. Justus von Liebig's law of the minimum says growth is capped by the scarcest input, not the average, the barrel holds only to its shortest stave. Both echo the weak-link logic, but they insist the binding link can be any factor, not just the tacit one, and that it moves as you elevate each in turn.

So the granular version of the doctrine is this. Value and risk concentrate at whatever floor binds now. Decode a tacit layer and you may relieve the human bottleneck only to hit a capital or a regulatory one. The builder's job is not only to walk the tacit bottleneck inward, it is to know which floor is binding at each moment, because that is where the next dollar and the next risk both live, and it shifts. The tacit floor is the one this doctrine dwells on because it is the deepest and most durable, but a builder who ignores the capital, compute, and regulatory floors will relieve the wrong one and stall.

The honest limits

• It works where patterns repeat and feedback exists (Kahneman). In genuine chaos, the decode-judgment step fails, and you should not attempt it.
• Decoding needs data. Some layers cannot be decoded until you have run enough cases to see the pattern. That is the flywheel, and it gates the schedule.
• Beware overfitting and false precision. Keep decoded rules interpretable, and render them as signals with confidence carried, not verdicts.
• The floor is real. The collective-tacit residue does not decode with today's tools. Claiming to automate it is the failure mode, not the goal.
• This is a lens, not a theorem. A powerful, cited way to see system design, to be applied with judgment, not followed off a cliff.

A worked example, told through Chad Jones

Chad Jones uses this case to make the weak-links point, and it is the whole doctrine in one story.

In 2016 Geoffrey Hinton, the grandfather of deep learning, said we should stop training radiologists, because within five years AI would read scans better than they could. It was the most credible possible voice making the Turing Trap error, assuming the task and the job are the same thing, that automating the read replaces the radiologist.

The prediction failed, and it failed exactly the way the doctrine says it should. The read was one layer of a job made of many. AI did take that decodable layer, spotting the pattern in the image, and Hinton was not wrong about that part. But the O-ring held. A radiologist's output was still gated by the links AI never touched, the clinical context, the ambiguous case, the procedure, the accountable sign-off. You do not get the productivity win until the weakest link is handled, and those links stayed tacit and human.

Then Baumol, twice. Cheaper, faster reads did not shrink demand, they grew it, more imaging was ordered because it was cheaper, so the job expanded rather than vanished. And the human residual rose in relative value as the readable part commoditized. The result, years on, is more radiologists making more money than ever. The bottleneck moved inward and the money moved with it. Adoption also lagged for regulatory and liability reasons, the rebuild-the-floor point again, so the story is not one mechanism but the whole doctrine at once.

The lesson to keep: the most authoritative voice in AI predicted the expert would be replaced, and reality relocated the expert instead, and paid them more. Bet on the doctrine, not the headline.

Why it is universal

The doctrine is domain-agnostic. It tells you, in any expert field AI is now touching, where to point the machine and where to keep the human.

• Medicine: automate the labs and the imaging reads, decode the pattern recognition, leave the bedside judgment and the accountable call.
• Law: automate research and drafting, decode precedent matching, leave the strategy and the ethics.
• Underwriting and M&A: automate the math, decode the operator's eye for where the risk hides, leave the judgment and the read on the people.
• Engineering, audit, lending, diagnosis: same shape every time.

In each, the mistake is the same, pointing AI at the easy task or faking past the floor. The doctrine points it at the bottleneck and stops it at the floor.

One sentence

Value lives at the weak link, the weak link is the tacit task, tacit has decodable layers, so you walk the bottleneck inward by decoding them, and what remains is the collective floor that rises in worth as everything around it falls.

Brad Detlor has spent thirty years operating in lower-middle-market M&A, across roughly sixty deals. He is the founder of JCoBee.