The Increment That Pays

In month six, Maria tried to fill in the form.

Not to submit it. That part matters. It was a Tuesday evening, the plant was quiet, and she had the internal template open in one window because she had been idly curious whether the thing she had built could be described in the language the company uses to describe things it pays for. A private experiment. Nobody was going to read it.

The template is not unreasonable. Every large company has one and they are more alike than their owners imagine, and this is one of the better ones. Two pages. Problem statement. Proposed solution. Scope and systems affected. Resource requirement. And then, about halfway down, the field that stopped her:

Expected benefit (quantified).

She could do part of it. The tool saves her roughly a day a week, which she could defend, because she had been counting: the deviation reports used to eat Thursday afternoon and most of Friday, and now they eat an hour. Call it eight hours, call it twenty per cent of an engineer, and there is a number a controller would accept. She typed something like that and looked at it.

Then she thought about the seal defect, and stopped.

Because the seal defect is the whole point, and the seal defect is not in that number. In month four the tool flagged a failure mode appearing three times in a fortnight across two different suppliers, which the old process would not have connected for another five weeks, because the old process was her reading reports in supplier order on a Friday evening and the two suppliers were nowhere near each other alphabetically. Five weeks of a defect not reaching the line. Anyone can price that. It is the single most valuable thing the tool has ever done and it is worth more than the day a week by a margin that is almost embarrassing.

And she could not have written it in the box. Not because it is hard to quantify after the fact. Because in month zero, sitting at her kitchen table pasting an anonymised report into a chatbot, she was not trying to catch seal defects. She was trying to survive Friday. Catching the seal defect was not a benefit she anticipated and then delivered. It was something the tool turned out to do, and she found out it could do it by having it, and running it, for four months.

There is a further problem, and it is the one she thinks about more.

The version that caught the defect is not the version she built. Her first version clustered by supplier, because that is how the reports arrive, how the system sorts them, how everyone thinks about them. It was fine. It saved her an afternoon. Some time around week six a batch came out wrong — she had pasted in a set from the wrong fortnight — and in the mess she noticed that the same failure mode was showing up under two supplier names, and that she had never once looked at the data that way in fourteen years, because nothing she had ever used presented it that way. So she rebuilt it to cluster by failure mode across suppliers. That is the version that works. She got there by accident, in the sixth week, from a mistake.

You cannot put that in the box either. Expected benefit (quantified) wants a number at the point of asking, and everything that makes the tool worth anything was discovered after the point of asking, by running it.

Maria closed the template without saving. What she felt, sitting there, was slightly stupid — as though there were an obvious way to fill in the form and she was simply not clever enough to find it.

There isn’t. The form was asking her, in month zero, for something that did not exist until month four. That is not a failing of hers and it is not a failing of the form. It is the shape of the whole problem, and this chapter is about it.

The oldest number in this book

Let me put the other half of the argument on the table first, because so far Part II has only shown you the machinery that buries things. It has not shown you that the buried things are worth anything.

They are, and the evidence is nearly ninety years old.

In 1936, T. P. Wright published an analysis of costs in aircraft manufacturing and reported a regularity that has been reproduced ever since. As cumulative output of an airframe doubled, the labour hours required to build one fell by something in the region of twenty per cent. Not because anyone invested in new equipment. Not because a programme was launched. Simply because the plant had built more of them, and the people in it had got better, in a thousand unremarkable ways nobody wrote down.

Kenneth Arrow gave it theory in 1962, in a paper about learning by doing. His argument was that a substantial part of productivity growth comes from experience accumulated in the act of production itself, and that this makes experience an input in its own right — one you cannot buy, order, or accelerate by spending, because the only way to accumulate it is to run the process and pay attention while you do.

The learning curve is now among the most durable findings in operations. It has been observed across aircraft, shipbuilding, semiconductors, chemicals, solar cells and surgical procedures. By the standard we set in Chapter 2 — not old, but re-tested — it is about as battle-tested as anything in this book. Every large manufacturer in the world has a version of it embedded somewhere in its planning assumptions, and everybody senior has heard of it.

The clearest statement of it I know did not come from an economist.

In 2017 Andy Jassy, then running AWS and now Amazon’s chief executive, was asked by CNBC how he thought about competitors copying what his business had built. He said there is no compression algorithm for experience. The technology could be replicated well enough, was his point; the years of operating it at scale could not be, because the only way to acquire those years is to spend them.

I worked at AWS, and that line gets repeated inside the company the way company lines do, which is often and not always by people thinking hard about it. It is worth thinking hard about, because it is Arrow’s argument arriving from the opposite direction — an operator explaining a competitive moat rather than an economist explaining productivity growth, both landing on the same conclusion, which is that experience is an input you cannot buy and cannot rush.

And there is a small irony in where it comes from. Chapter 5 borrowed the one-way and two-way doors from the same company. So the firm that supplied manufacturers with their favourite vocabulary for being careful about irreversible bets also supplied the clearest argument for why the slow accumulating stuff is where the advantage actually sits. The first half has been imported enthusiastically. The second half has not.

Here is what I want you to notice about it.

The learning curve does not pay out in events. It pays out in repetitions. Its returns are a function of how many times you have done the thing, which means value arrives continuously, in quantities too small to see individually, and becomes visible only in aggregate and only in retrospect. It is the erosion from Chapter 5 expressed as an economic law: the grinding is where the shape comes from.

And the machinery we spent the last two chapters describing cannot evaluate anything of that shape. Hurdle rates want a return on a commitment. Business cases want a benefit at a date. A gate is an instrument for judging discrete proposals at discrete moments, and the compounding increment does not present itself as a discrete proposal at a discrete moment. It presents itself as a small thing that will be worth a lot if you run it for two years, which is not an argument the form has a field for.

So the most reliable engine of improvement in manufacturing pays in a currency the approval process does not accept.

Where the learning actually happens

The learning curve tells you that the improvement pays. It does not tell you where it comes from, and that turns out to matter more.

In 1997 Marcie Tyre and Eric von Hippel published a study in Organization Science of what actually happens when new production equipment is introduced into a factory — the messy early period when the machine is installed and nothing quite works. They followed the engineers doing the problem-solving, and what they found was that the process was situated. The clues needed to diagnose a problem were physically located in particular settings: at the machine, on the floor, in the hands of the operator running it. Engineers moved back and forth between the lab and the line, not because they were being thorough, but because the information required to solve the problem existed in more than one place and could not be assembled anywhere else.

The finding, stated plainly, is that adaptive learning happens where the work happens, and it cannot be relocated to a more convenient venue.

Now hold that against the geography of a decision. The approval sits in a meeting room, or a portal, or a council that convenes monthly. The learning sits at the line. The person who has the clues is standing in front of the machine and the person with the authority is two floors up looking at a template, and the thing being asked for — a description, in advance, of what the improvement will be worth — is precisely the thing that only becomes knowable in the setting the decision is not taken in.

Maria’s rebuild is a small instance of exactly this. Nothing in her fourteen years of experience told her to cluster by failure mode instead of by supplier. No workshop would have produced it. It came out of a mistake made in the sixth week of running the thing, in the presence of the actual data, and it is worth more than the original design.

The part that cannot be written down

There is a second reason the form defeats her, and it is older and better known than the first.

Ikujiro Nonaka’s 1994 account of organisational knowledge creation rests on a distinction between knowledge that has been made explicit — written, codified, transferable — and knowledge that remains tacit, held in a person, expressed in how they work rather than in any artefact. His argument is that organisations create knowledge by converting one into the other, and that the tacit-to-explicit conversion is both the most valuable step and by far the hardest.

Maria’s plant runs on tacit knowledge and it does not know it. She knows the seal supplier’s quality dips every August. She knows the third shift logs deviations differently to the first. She knows the downtime on line two has been logged against the wrong press since a system migration nobody ever fully unpicked. None of that is in a system. It is in her, and in a handful of colleagues, and it is the reason her tool works at all — she built the clustering rules around distinctions that only exist in her head.

A business case is an explicit artefact. It is, by construction, a request to convert tacit knowledge into codified claims before the conversion has happened. And the increments most worth having are the ones densest in tacit content, because that is what makes them local, and specific, and effective, and hard for anyone else to have thought of.

So the gate applies its heaviest documentation requirement precisely where documentation is least possible.

The paradox, and its name

Put the three findings together, because separately they are unremarkable and together they close a trap.

The learning curve says the value of an increment accrues through cumulative running. Situated learning says the design only becomes correct at the site of use, through iterations you cannot specify in advance. Tacit knowledge says the substance of the improvement resists codification. Three well-established results, from three literatures that rarely speak to each other, and all three say the same thing about timing.

The value of a good increment does not exist at the moment it must be justified. It is produced by the thing running, in the place it runs, over a period, and the largest part of it is usually something nobody predicted.

This book calls that the Subterranean ROI Paradox: the increments most worth having are the ones least able to justify themselves in advance. Not because they are small. Because their value has not happened yet, and the gate wants it now.

That name is a framing device, not a finding. Wright and Arrow did not write about approval processes. Tyre and von Hippel were studying equipment start-up, not governance. What each of them documented is real and I have cited it; the assembly is the book’s, and you should hold it as an argument rather than as a result.

But notice that it retroactively explains something we left unexplained in Chapter 1.

The bootlegging research found that hidden projects, on average, come out better than sanctioned ones, and the mechanism proposed for that was delayed assessment — that being hidden buys a project time before anyone judges it. That always sounded slightly like an excuse, a nice way of saying that unsupervised people enjoy themselves more. It isn’t. Delayed assessment works because early assessment is assessment of a thing that has not yet become what it is going to be. The submarine is not protecting the idea from criticism. It is protecting the idea from being measured before there is anything to measure, which is a different and much better defence.

What Toyota knew

I need to close off a reading of this chapter that would be convenient and wrong.

Nothing above is an argument against structure, or process, or formality. The easy version of the argument says bureaucracy kills increments and the answer is less bureaucracy, and there is a famous counterexample sitting in the middle of the industry this book is about.

In 1999 Paul Adler, Barbara Goldoftas and David Levine published a case study of model changeovers at NUMMI, the Toyota-run plant in California, in Organization Science. What they describe is a factory that was intensely bureaucratic — standardised work specified to a degree that visitors found startling, formal routines everywhere — and simultaneously one of the most flexible plants anyone had studied. It did both. The mechanisms they identify are unglamorous: metaroutines, routines for changing routines; partitioning, so that the work of improving is separated from the work of producing rather than crushed by it; switching between modes; enrichment of the production role itself. Underneath all of it, sustained investment in training and a level of trust that took years to build.

Toyota’s system is a machine for capturing increments. It is more formal than the system Maria works in, not less. So formality is not the variable.

What is different is when and where the judgement happens. In the Toyota case the change is made at the line, by the people at the line, and then evaluated by what it does — with the standardisation arriving afterwards, as the way a proven improvement is locked in and spread. The sequence runs try, observe, standardise. Maria’s plant runs justify, approve, try, and it does that because, as Chapter 5 showed, it learned its decision-making on objects where trying first is not an option. You do not pilot a body shop.

Both sequences are coherent. Only one of them can handle a thing whose value appears in month four.

The arithmetic nobody ran

Let me put a number next to this, with a warning attached: what follows is an illustration, not a study, and I am not going to dress it up as one.

Maria’s tool cost six hours and eighteen euros a month. It returns something like a day a week of a quality engineer, plus one early defect catch in eight months of the kind that keeps bad parts off a line. Run that for two years and the arithmetic is not close. It is not close by two orders of magnitude, and you could halve every figure I have given and it would still not be close.

The Quality Systems programme has eighteen months, two consultancies, a budget Maria found hard to repeat out loud, and a roadmap slide that is very good. It has shipped nothing to her plant.

The point of that comparison is not that the small thing beats the big thing. Sometimes it doesn’t, and Chapter 12 will make the case for big programmes properly, because they are the right instrument for some problems and this book is not an argument for a company made entirely of Marias.

The point is that the comparison was never available. To be compared, both sides have to produce a number in the same units at the same time, and only one of them can. The programme can generate a forecast on demand — a large, confident, thoroughly formatted forecast, produced by people whose profession is producing them. Maria’s tool can generate a real figure only after two years of running, and it will never run for two years, because to run at all it needed a forecast in month zero.

So a process that requires estimates in advance does not select for the most valuable improvements. It selects for the improvements whose value is estimable in advance, which is a different quality altogether, and one that correlates with size, precedent and distance from the actual work. There is no malice anywhere in this. Every individual step is defensible. The system just quietly optimises for forecastability and calls it rigour.

Where the evidence stops

Four things in this chapter run further than the sources do, and one of them could sink the argument.

The learning curve is common, not universal. It is not a law of physics and it is not automatic. It shows up where an organisation has the willingness and the ability to learn, and where somebody invests in the learning; where those are absent, cumulative output produces cumulative output and nothing else. It also has a boundary that matters here: the classic learning-curve effect describes improvement within a production method. It does not describe changing the method. If your improvement needs a different machine, a different line layout or a different process altogether, then it is a capital decision, the one-way-door caution from Chapter 5 is the correct instrument, and this chapter does not apply to you.

Some increments cannot be separated from the big bet at all. You cannot half-retool a line. There are improvements that are only available as part of a large indivisible commitment, and for those the gate is not an obstacle, it is the only mechanism that could possibly work. The claim I am making is narrower than “the gate is wrong”: it is that reversible, low-cost, locally-learned improvements are being routed through a process built for irreversible, high-cost, centrally-decided ones, and that the routing is the error.

The empirical base for the situated and tacit parts is two case studies. Tyre and von Hippel studied equipment start-up in a small number of plants. Adler and colleagues studied one factory, and it was an exceptional one — NUMMI is famous precisely because it was unusual, which is a limitation as well as a recommendation. These are the best manufacturing-floor anchors I have, and they are two case studies, and I would rather say so than let the word “research” do work it hasn’t earned.

And here is the study that would prove me wrong. Take a large organisation’s completed small process improvements, several hundred of them, and compare what each one’s pre-implementation forecast said it would deliver against what it actually delivered. If the forecasts turn out to be decent — if the benefit an improvement was predicted to produce broadly matches the benefit it produced — then the paradox in this chapter dissolves, the gate is doing exactly what it is supposed to do, and the increments that die at it deserved to. My expectation is that forecasts will systematically miss the largest benefits, because the largest benefits are the ones discovered in use. But I would like to see it done, and as far as I can find, at scale, it hasn’t been.

The trap, complete

Part II is finished, and it is worth saying out loud what it has assembled, because the pieces only become alarming in combination.

Chapter 4 showed that an increment can be killed without ever being judged on merit, because it only has to land on somebody else’s territory. Chapter 5 showed that it is rarely even judged, because capital, attention and the calendar all bend toward mass, and a six-hour tool has no mass to bend them with. And this chapter has shown that the thing being buried is not sentimental. Small compounding improvement is the best-documented engine of productivity growth in manufacturing, it has been for ninety years, and the reason it cannot get through the door is not that it is worth too little. It is that it cannot prove what it is worth until it has been allowed through the door.

The trap is now fully visible from both sides. The organisation cannot fund what it cannot forecast. The increment cannot forecast what it has not yet done. And the person holding it — who can see all of this perfectly well without any of the vocabulary in this book — draws the obvious conclusion, which is to stop asking and carry on quietly.

Which leads to an answer that will feel, right now, completely satisfying. If the increment does better underwater, and the gate can only damage it, then leave it there. Protect the submarine. Let the fleet sail.

Hold that conclusion for as long as you can, because Part III is going to take it apart.

Next: Part III. Why it’s never been fixed — starting not with a mechanism, but with a graveyard.