The falling cost of reproducing industrial know-how

AI is changing what a successor can learn from a company's mistakes, and what a competitor may be able to reconstruct.

A general intelligence knows everything in general and nothing about you in particular. It has read most of what has been written and none of what happened in your specific corner of the world. I have a friend who built most of what he has before 1980. He doesn't know what a hyperparameter is and never cared. What he has instead is something that accumulated slowly and without intention, mostly through being wrong in expensive ways: a precise sense of who in his world understands what, where their edges are, and what they'll do when they reach them. His competitors know him, he knows them, and, more to the point, each of them knows what the others know. That's the whole game at that level. A year ago, over a lunch that stretched, Henry told me domain knowledge was the only thing that had ever actually mattered. The kind you can't delegate, because the judgment doesn't travel well. He said it the way you state something long settled, with no argument attached. I've been less certain ever since, mostly because I'm no longer sure knowledge means the same thing it did when he learned it.

What I keep returning to is what the education cost. Henry learned by making decisions and living with their consequences, and I believe him about how much of his judgment came from that. I am less certain that someone arriving now would have to pay the same price. They would start with whatever those years had left behind: rejected quotations, revised specifications, a supplier dropped after a run of problems. The record would be incomplete. It would also be larger than anyone who wrote it realised, because nobody writes a specification thinking of the person who will read it twenty years later to work out why it changed. Henry knows what it cost him to learn. That number, on its own, says nothing about what it would cost someone else to reach similar decisions. Those are two different prices, and the whole question of this essay is how far apart they have become.

A veto on a raw material

A smaller case than Henry's helps, one where the decision can be examined instead of admired. Take a mid-sized manufacturer of industrial coatings. It formulates products, adapts them to specific applications, and responds when something that worked in the laboratory fails on a customer's line. Purchasing proposes substituting a raw material. The alternative meets every specification on file and performs in the initial tests. The technical director, who has been there since the company had one reactor, asks that it not be approved yet for a particular application. He does not write a report. He says something like I would wait on that one for the marine range, and the room accepts it, because the room has learned that when he says this he is usually right.

The interesting part is working out what he just did. He may be recalling a documented failure that nobody has retrieved from the archive. He may have recognised a combination of conditions (substrate, cure temperature, a customer's habit of thinning the product beyond what the data sheet allows) that the others have not connected. He may also be generalising from one bad experience in 1998 that has nothing to do with this supplier. All three are possible, and from inside the meeting they look identical. That is the uncomfortable property of this kind of judgment: its correctness and its authority are established by different mechanisms, and the second one runs on reputation.

I will call what the director has company know-how, and keep domain knowledge for the broader thing Henry means. The difference is the difference between knowing coatings chemistry and knowing how coatings chemistry behaves in this plant, with these customers, under this history of things that went wrong. A chemist can be hired with the first. The second is what the company is worried about losing when the director retires, and what it is quietly relieved its competitors do not have. Treating the veto as an examinable decision gives the reader a reason to keep going: what evidence would let someone else arrive at a good enough version of that same call, and how much of that evidence already exists?

What the mistakes left behind

The turn Henry's position needs is this: the difficulty of explaining a judgment does not establish that it is impossible to learn to reproduce part of it. Henry may be unable to state a complete rule and still have left, without intending to, a trail of corrected decisions, amended specifications and test results. The question moves from can he explain it to what do the traces allow someone to reconstruct, and that is a question with evidence attached.

Three studies make the point uncomfortable for him, and they are useful precisely because they show three different mechanisms with three different limits. The first is the one closest to being wrong in expensive ways. In 2016, Raccuglia and colleagues took the archived notebooks of a chemistry laboratory, including the syntheses that had failed, added physicochemical descriptors, and trained a model on the crystallisation of vanadium selenites. The model predicted conditions that produced new compounds with an 89 percent success rate in the trials they report, better than the chemists it was compared against. The scope was narrow. What matters is the input: decades of experiments that had been considered dead, sitting in notebooks, turned out to carry enough structure to guide new decisions. Whoever ran those failed reactions paid for an education. The model inherited part of it at no charge.

The second study reproduces a function of expertise by a route the expert never took. In 2021, a team from Princeton and Bristol Myers Squibb compared Bayesian optimisation against chemists and engineers on reaction optimisation problems, and the algorithm found good conditions with fewer experiments and more consistently within the tasks studied. Expert work remained in framing the problem and running the reactions. But choosing the next experiment well, which is a large part of what a senior chemist's judgment is worth economically, did not require reconstructing how that chemist came to have good judgment. It was done another way. This is the part that a sentimental defence of Henry misses: nobody needs to copy his map if they can reach his decisions from a different direction.

The third brings the mechanism to the generative models everyone is now arguing about. A 2025 field experiment reported in The Cybernetic Teammate put 776 professionals at Procter & Gamble to work on product development tasks. Individuals working with an AI assistant produced proposals judged comparable in quality to two-person teams working without one, and their proposals integrated commercial and technical perspectives better than the solo work of specialists on either side. These were one-day workshops with expert evaluation of the proposals. Nothing was manufactured. The result still bears on the veto in the meeting: some of what appeared to require another specialist in the room could be carried into a proposal by assistance.

A well-evaluated proposal and a product that works are separated by everything that happens after the meeting. Two of these three mechanisms predate the chatbots by years, which makes it harder to file the argument under enthusiasm for the latest interface. None of the three shows a machine reproducing an industrial process end to end. What they show is that specific, costly, apparently tacit judgment has left traces that could be learned from, in cases where someone bothered to look.

The map of competitors

Now back to the thing I said was the whole game, and whether any of it is reconstructible. Precision matters here, because this is where the argument can slide into either a defence of Henry or an advertisement. Parts of the map are candidates for reconstruction. Other parts depend on evidence that may never have been recorded. I am not claiming that a model can rebuild Henry's particular map today.

Take the two halves of what he knows about a competitor. The first is declared capability: what they say they can do. For this the sources are patents, technical documentation and the products themselves. A patent has to disclose enough for a skilled person to reproduce the invention, and its claims mark out what is protected. That is real information, and a model can read all of it in an afternoon. It is also, as anyone who has read a competitor's patent knows, evidence of what they wanted to protect and only weak evidence of what they can actually ship at scale. The second half is observed capability: what they turned out to be able to do when it mattered. For this the sources are internal and unglamorous. Lost tenders, and why. Rejected samples. Comparative tests run against their product. Complaints that came back from a customer who had switched and then switched again. Those records exist in most companies, badly indexed, and before they could support any prediction someone would have to check how much of the history they cover and how honestly they were written at the time.

The most valuable part of Henry's map sits above both halves. It is the step from that company can do this to I know how that person will react when he discovers he cannot. To learn that distinction you need episodes: identifiable people, under identifiable pressure, doing identifiable things. A company profile does not contain them. A CRM contains a caricature of them. And the gaps are structural. If Henry never hired a certain supplier, his history says nothing about what would have happened if he had. If a post-mortem was written after the outcome was known, it cannot be treated as a rule that was available beforehand; it reads like foresight because hindsight wrote it.

There is one more separation worth keeping, because it explains why knowledge travels badly even when the record is complete. Relational knowledge and the capacity to act on it are different things. Suppose a customer gives Henry a second chance after a bad batch because of who Henry is to him. A successor can be told about that relationship. The customer is under no obligation to extend the same courtesy to the successor. What travelled was the information. What did not travel was the situation, because the person acting in it changed. Some of what Henry cannot delegate is missing data. Some of it is that the world responds to him and not to the file.

Reproduction cost

The competitive threat that follows from all this does not need a leak. That is what makes it more serious than the version most companies are worried about, where a vendor absorbs their context and hands it to a rival. Assume no provider leaks anything and every contract holds. Imagine a competitor trying to build an equivalent capability. It has its own experiments, the public documentation, a laboratory and access to testing. The only question is how much work it needs to do to reach an acceptable result.

The three studies above justify examining whether certain experimental searches can be shortened. They do not establish that any industrial process can be reproduced on demand, and that distinction belongs in the body of the argument rather than in a defensive note at the end. But a search that once took a competitor five years of trial and error taking two changes the value of having done it first, even when the first company's formulations stay locked in a safe.

Part of an advantage can deteriorate because the work needed to rebuild it has fallen, while every piece of confidential information behind it remains protected.

This concerns people who think about patents and trade secrets, but it concerns more directly the people who believe their competitor still has to walk the same years of failure they walked. Some of them are right. Access to installations, customer trials and new results can keep limiting what an outsider can learn, and a plant is not a notebook. The honest version of the question is a distribution of costs, task by task: which parts of what took Henry twenty years now take a well-equipped newcomer two, which still take twenty, and which cannot be had at any price without the customer's cooperation.

The same question drawn task by task. Published knowledge and experimental search sit where the record is rich and the test is cheap, which is where the three studies live. Customer trust and behaviour under pressure leave almost nothing behind and can only be tested by living through them.

Once you have looked at reproduction from outside, transmission inside the company looks different. The company has every reason to make it easy for the successor what it wants to keep hard for the competitor. It can give the successor the internal record, time watching the director work, and exposure to real decisions while the director is still there to be wrong in front of him. Henry's problem now has a precise shape: which part of his experience can become someone else's capability, and what conditions does that person need in order to acquire it.

The usual approach is a series of interviews before retirement, transcribed and filed. I would avoid treating knowledge capture as an exit interview, because what comes out of an exit interview is explanation, and explanation is the least reliable artefact an expert produces. For the raw-material veto, the more useful exercise is reconstructing what the director had in front of him when he intervened, what signal made him stop the approval, and what result would have changed his mind. The third question is the one nobody asks, and it is the one that turns a veto into something testable.

This is where the idea of validity conditions earns its place. A record of a failure becomes useful when it lets you see under what conditions the thing failed and what has changed since. The 1998 batch that went wrong on the marine range failed with a resin that has since been reformulated twice. The director's caution may still be warranted, or it may be a constraint whose reason disappeared a decade ago and whose authority did not. Both are possible, and the same reconstruction that would let a successor learn from the veto would also let the company find out which one it is. That cuts both ways for Henry, and I think it should.

Models can help here as instruments: recovering cases from a badly indexed archive, contrasting the director's account with the test results, organising the questions a successor should ask before approving something. The test of the handover is behavioural. The successor takes better decisions on cases the director never saw, including the decision to go and ask someone. An archive full of convincing explanations proves nothing about that.

There is a human question inside this that deserves its own paragraph. For the expert, transmitting judgment means relief and loss of standing at the same time. A company that frames the project as knowledge extraction has to ask what it is offering the person who must teach, because a technical director who has spent thirty years being the only one who could distinguish an important exception from an ordinary incident is being asked to make himself less necessary. Some of the resistance to these projects is that, and it is a reasonable position. Filing it under fear of technology gets the incentive wrong and the project fails politely.

The essay could stop here, at an exercise that has not yet been done. Take one real decision of Henry's and rebuild what could have been known before the outcome. Compare what the existing archive supports with what changes once his own observations are added. Then check whether what was learned helps with subsequent decisions, under supervision, against criteria decided in advance. An answer that resembles Henry's explanation is a weak signal. What matters is whether it helps someone catch a constraint, avoid repeating a test, or ask for the evidence that is missing.

What I would want, to take this further, is the case itself: a veto, a rejected bid, or an abandoned experiment whose reasons and outcome can still be recovered. One episode reconstructed properly could carry the whole argument, with the external studies there to question what the narrator thinks he is seeing.

Henry, if he reads this, will point out that I still have not said what his map contains. That is true. I have said where the edges of it might be, and which parts are worth trying to trace before someone else does. He would say the parts that matter are the ones nobody wrote down. He may be right. It is also possible that he has not looked at what was written down in a long time, and neither have I.