The Graveyard of Good Programs

In month seven, Maria went looking for a legitimate route.

She had not decided to take one. That distinction keeps mattering. What she had was an arithmetic problem left over from the previous chapter: her tool is worth what it is worth only if it runs for years, because a day a week is a rate and the seal defect was a single event in eight months, and neither of those means anything over a quarter. Two years is the minimum honest window. Three would be better. So the question she was actually asking was not whether the plant would let her do this. It was whether whatever let her would still exist when the arithmetic came in.

She spent about forty minutes on the intranet. This is what she found.

The suggestion scheme is alive. Actually alive, in the sense that a form exists, a committee reviews submissions quarterly, and there is a payout with a cap that was set in a currency and a decade that no longer apply. She knows two people who have used it. One of them got a hundred and fifty euros for a fixture change that saved a shift’s worth of rework a month, which he still mentions, and which she cannot decide is a good story or a bad one. The form has a mandatory category field. The categories are Safety, Quality, Cost, Ergonomics and Environment. Below the category field there is a field labelled attach drawing or photograph.

There is no category for having built something.

The idea portal is two and a half years old. It launched with an app, and a video from somebody at head office standing in front of a wall of screens, and a competition. Maria logged in — the login still works, which surprised her — and the news feed on the landing page has one item on it, dated nineteen months ago, announcing the winners of the competition. Nobody has removed the portal. Removing the portal would be somebody’s job, and it is not anybody’s job.

And then there was Improvement Time, which is the one she actually remembered, and the one worth the paragraphs.

Improvement Time was two protected hours every second Friday, with no reporting requirement attached, which was the part that made it unusual and the part that made it good. The plant manager at the time launched it himself and came to the first session. Maria went to three. The first one was better than she expected — eight people, a flipchart, and an hour in which somebody from logistics said something about pallet labelling that she is still using. The second had four people, because the first shift had run long. The third was cancelled on the morning, because a supplier escalation came in overnight and everyone with a quality badge was needed in a room on the other side of the plant.

Which was correct. That is the thing to hold on to. Cancelling was the right call that morning, and it was the right call the next time, and the time after that, and nobody ever made a wrong decision anywhere in this story.

There was no fourth session. There was no announcement that there would be no fourth session. At some point in the following months Improvement Time stopped being a thing that was happening and became a thing that used to happen, and Maria could not tell you the month it changed, because nothing changed in any month. The plant manager moved to another site about a year later. She does not think those two facts were connected at the time, and she has since revised that view.

The poster is still up. It is in the corridor by the coffee machine, on the way to the room where Tuesday’s meeting will happen, laminated, with one corner lifting away from the wall. It has a QR code on it. The QR code still resolves, to a login page for a system she does not have access to.

Maria was not angry about any of this, and sitting there she noticed that she had never been angry about it, not once, in four years of walking past that poster. It is difficult to be angry at something that stopped. There is no decision to object to, nobody did anything to her, and the only evidence that anything was ever taken away is a piece of laminated card that nobody has got round to removing.

What she registered instead was the shape of the answer. She had gone looking for a route that would last three years. The longest-lived object she found in forty minutes of searching was the poster.

Part II ended by telling you to protect the submarine. Protection is a programme. This chapter is about what happens to programmes.

The headstones are good

I want to be clear about what is in the graveyard, because there is a lazy reading of this chapter in which the fixes died of being bad.

They were not bad. Several of them were better than anything that has been written about the problem since.

3M has had a version of its fifteen per cent since the late 1940s. Google offered twenty per cent from around 2000. Kaizen and the continuous-improvement family have been deployed in tens of thousands of plants for forty years, with training budgets, certified practitioners and a literature. Suggestion schemes are older than all of it. And the single most interesting intervention in this whole book, which we will come to properly in Chapter 10, is Bernstein’s curtain — the one design that produced a measured, sustained improvement in a real factory, and it did that by giving managers less visibility rather than more.

Every one of these was designed by competent people who had correctly identified the problem. That is what makes the graveyard interesting. Chapter 2 registered the instinct — companies noticed that unofficial work produced value and tried to authorise it — and promised you we would watch it fail repeatedly. This is the chapter where I make good on that, and the failure mode is not what most people assume. The programmes did not get rejected. They got launched, and then they thinned out, and then at some point they were being described in the past tense by people who could not name the day.

The most famous fix

The Google case is the one everybody knows, and it is worth going through carefully, with a warning attached: what follows is journalism and memoir. It is a case narrative, not a study, and I will flag where it thins.

Twenty per cent time ran from around the turn of the century. Engineers could spend a day a week on work of their own choosing, without seeking approval for the choice.

Marissa Mayer, then a vice-president there, used to tell a story about where Google News came from. After September 11 a researcher named Krishna Bharat was going to a dozen or more news sites a day trying to follow the story, got tired of it, and wrote a crawler to cluster the articles for him. He emailed it around. Somebody who received it looked at it and thought it was not a private tool but a product.

Read that again with the previous six chapters in mind. A person automating their own Friday afternoon, for themselves, with no product intention anywhere in the account, producing something the company had not asked for and could not have specified. That is Maria with a different job title, and the only reason it became a product instead of a private habit is that the environment it happened in did not require it to be justified in advance.

Mayer also supplied the number that made twenty per cent time famous. In a talk at Stanford she said she had mapped the product and feature launches of the second half of 2005 and worked out which came from twenty per cent time and which came through the normal process, and the answer was half.

Take that figure the way it deserves to be taken. It comes from a talk, given by an officer of the company, about her own employer’s most celebrated practice, and it is her own assessment rather than an audited one. It has been repeated ever since in books and slide decks by people who did not check. By the standard this book set in Chapter 2, it is a claim, and I am reporting it as a claim.

But now notice what happened next, because the claim’s accuracy turns out not to matter much.

By 2013 the programme was being written about in the past tense. The reported causes are unglamorous and they are all recognisable from Part II. Post-IPO quarterly pressure, and the reporting rhythm that comes with it. No formal structure and no tracking, so nothing to point at. Uneven adoption — the figure that gets repeated is something like one engineer in ten using it consistently, which is also a practitioner number and also unaudited. And a steady re-centralisation of innovation into elite labs, which is Chapter 5’s big factory reasserting itself in a company that was supposed to be immune to it. Laszlo Bock, who ran people operations there, described the thing as having waxed and waned with the level of leadership support.

There was no cancellation. That is the part I find hardest to get past. Twenty per cent time was not ended by a decision anybody can date. It became harder to take, then it required a manager’s agreement, then engineers started using the phrase “a hundred and twenty per cent time” about it, which is a programme’s obituary written by the people inside it.

And the death is contested even now. In 2015 Mayer said the twenty per cent had never really been twenty per cent — that it was a full workload plus whatever else you wanted to do. A Google engineer publicly disagreed, pointing at his own long-running twenty per cent project and saying he was not working a hundred and twenty per cent of anything. Both of them were there. Neither of them is lying.

So here is the summary of the most famous fix in the history of corporate innovation. It ran for over a decade, it was credited by a company officer with half of a year’s launches, and it declined without an announcement, in a way that leaves the people who were present disagreeing about whether it existed as described and about when it stopped.

A programme with that record could not defend its own budget line. I think that is the single most important fact in Part III, and the reason for it is sitting in the chapter you just read.

Ask what twenty per cent time delivered last year and you get a story about 2001. Ask what it will deliver next year and nobody can answer, because the whole design principle is that nobody is specifying the work in advance. The programme’s returns arrive late, arrive as something else, and get attributed to whoever shipped the product rather than to the slack that produced it. Which means that when the programme comes up for review — and everything comes up for review — it has to justify itself with exactly the forecast it exists to make unnecessary.

The Subterranean ROI Paradox does not only apply to increments. It applies to the mechanisms built to protect increments. The fix is subject to the gate.

The manufacturing version, which is better documented and less famous

The Google case is vivid, and vividness is not evidence. The stronger material is in the continuous-improvement literature, where the same thing has been studied for longer by people with no stake in the outcome.

Nicola Bateman’s 2005 paper in the International Journal of Operations & Production Management is the one whose title says it. She called sustainability the elusive element of process improvement — not achieving the improvement, sustaining it. Analysing forty improvement activities, she found five conditions associated with activities that actually landed, and five more associated with the smaller number that went on to become continuous. In 2014 Rachel McLean and Jiju Antony published a systematic review in the International Journal of Productivity and Performance Management on why continuous-improvement initiatives fail in manufacturing environments, which is a topic with enough failures behind it to warrant a systematic review.

The mechanism these converge on is simple enough to state in a sentence, and I have not found a better sentence for what happens to fixes in general.

The conditions that enable an improvement are temporary. The conditions that generate the problem are permanent.

Everything on the enabling side has an end date. The consultant leaves. The workshop week finishes. The sponsor is promoted or transferred. The training budget covers this year. The attention that made the whole thing feel important lasts about as long as attention lasts, which Chapter 5 already told us is not long and is not yours to allocate.

Nothing on the generating side has an end date. The volume targets are permanent. The escalations are permanent. The headcount is what it is, the schedule is the schedule, and the thing that made the process bad in the first place — whatever combination of history, cost pressure and organisational geometry produced it — is all still there on the Monday after the programme finishes.

So the permanent conditions outlast the improvement rather than defeating it, which is cheaper and requires nobody anywhere to oppose anything.

Session three of Improvement Time was cancelled because of a supplier escalation. That is the whole mechanism in one Friday morning. The escalation was permanent in the sense that matters here — not that one escalation recurs, but that there is always an escalation, and the two hours were always going to be the thing that yielded.

There is a set of figures that circulates about how fast this decay happens. You will have seen them: some large percentage of improvements gone within twelve or eighteen months. I am not going to repeat them, for the same reason Chapter 2 refused the famous innovation-failure-rate numbers. They are practitioner and consultant figures, they get cited to each other in a loop, and when you follow them back there is nothing at the end. The direction is documented. The magnitude is folklore, and this book does not launder folklore into fact even when the folklore is pointing the way I want to go.

What a programme actually is, on a cost sheet

Underneath all of these fixes, from 3M’s fifteen per cent to two protected hours in a plant nobody has heard of, there is one resource being spent, and it has a name in the literature.

Cyert and March called it organisational slack in 1963: resources beyond what current operations strictly require. Time not committed, attention not booked, budget not allocated to a named deliverable. Slack is what the underground has always run on, and every programme in this chapter is an attempt to create some deliberately and hand it to people.

Nohria and Gulati studied the relationship between slack and innovation and found an inverted U, which is the honest shape. Too little starves experimentation and there is nothing to divert. Too much produces undisciplined projects that should have been stopped. There is a middle, and the middle is where you want to be, and I have never once seen a large manufacturer worry that it might be to the right of it.

Here is the problem with slack, and I think it is the mechanism that explains the graveyard better than anything else I have.

The cost of a protected-time programme is legible. It is hours, it converts to headcount equivalent, it lands under a budget code, and a controller can produce the figure in an afternoon. Two hours a fortnight across four hundred people is a number, and it is not a small number, and it will be on a slide at some point.

The benefit is not legible. It arrives late, it arrives as something else, it gets credited to the person who shipped the thing rather than to the arrangement that let them build it, and a great deal of it never surfaces at all because it stays as somebody’s private improvement to their own Friday. There is no line for it anywhere.

So the fix appears on the cost side of the ledger with complete precision and on the benefit side not at all. In a good year nobody looks. In a bad quarter somebody looks at everything, and this is the cheapest thing in the building to stop, because stopping it requires no negotiation with a customer, no contract to unwind, no capital to write off, and no one person whose job disappears. You stop it by not scheduling it.

And there is a real conflict here rather than a stupidity, which is worth saying because this book keeps insisting that every step is defensible. Forty years of improvement doctrine in manufacturing has been about the removal of slack. That doctrine was right about a great many things and it built the modern plant. But it cannot distinguish between slack that is waste and slack that is the input to the learning curve, because the two look identical at the moment you measure them. They are only distinguishable afterwards, and Chapter 6 has already explained why afterwards is too late to be useful.

Which gives the programme its half-life. A fix of this kind survives as long as somebody with standing is prepared to defend a cost line with no matching benefit line, on the basis of a belief they cannot evidence. That is a personal act, not an institutional one. So the programme’s lifespan is roughly the tenure of the person who believes in it, and Maria’s plant manager moved sites about a year after Improvement Time stopped happening, and I do not think she is wrong to have revised her view about the connection.

The fix that eats itself

There is a second way these programmes die, and it is the one that catches the organisations that take the problem seriously.

Recall the finding from Chapter 1 that Chapter 6 finally explained. Criscuolo, Salter and Ter Wal found that bootleg efforts were associated with higher individual innovative performance, and the mechanism was delayed assessment — the hidden project gets to become what it is going to be before anybody judges it.

That is not the only thing they found. In the same study, during periods of R&D process formalisation, more bootlegging was associated with less innovation. The behaviour did not change. What changed was the surroundings, and the same behaviour in formalised surroundings stopped producing the same result.

Now think about what happens to a well-run protected-time programme over its first two years, in an organisation with competent management, which is to say an organisation where things get managed.

Somebody sensibly asks how many people are using it, so there is a sign-up. Somebody sensibly asks what has come out of it, so there is a showcase each quarter. Somebody sensibly asks how we know the time is being used well, so there is a field on a timesheet. Somebody sensibly asks whether we should be steering it toward the strategic priorities, and now there is a list of suggested themes, and the list is not compulsory, and everybody understands the list.

Each of those is a reasonable request from a reasonable person. Together they have rebuilt the gate inside the sanctuary. A protected day with a sign-up, a tracking field, a quarterly showcase and a themes list is an approval process with better manners, and the thing that made the underground work — that nobody is assessing the embryo — has been removed by people trying to look after it.

So the programme has two available deaths and it must choose one. Audit it, and it stops being a nursery, because assessment is the thing it existed to defer. Leave it unaudited, and it has no evidence of its own value when the bad quarter arrives, which is the death described in the previous section.

And notice what actually did the killing, because nobody judged the programme and found it wanting. The looking was the whole event, and what the programme became afterwards was set by the looking rather than revealed by it.

I have not found a version of this trap that anybody has solved in the general case. Bernstein’s curtain is the closest thing to an escape I know of, and it works precisely because it does the opposite of what a manager’s instinct demands: it reduces what can be seen, on purpose, and then measures the line rather than the people. That is Chapter 10, and it is the reason that study matters more than anything else in this book.

Where the evidence stops

Four limits, and the first one is about the chapter’s most quotable material.

The Google case is journalism, and its death is disputed. Everything in that section comes from press coverage, from talks given by executives about their own company, and from memoir. There is no study. The adoption figure and the launch-share figure are both unaudited claims that have been repeated into furniture. And the participants do not agree on the basic facts: one senior figure has said the twenty per cent was always effectively a hundred and twenty, and engineers who were running twenty per cent projects at the time have said publicly that this is not what it was like for them. I am using it anyway, because it is the most legible instance of a pattern the manufacturing literature documents far more carefully and far less vividly. But the careful documentation is Bateman and McLean, not Google, and if you take one citation out of this chapter take one of those.

Programmes do survive, and I cannot fully explain the ones that do. 3M’s fifteen per cent has lasted more than seventy years, through every pressure described above and several recessions. So the claim here is a tendency, not a law, and any honest version has to account for the exception. My guess — and it is a guess — is that it survived because it predates the reporting culture that would have made it justify itself, and because it long ago stopped being a programme and became something the company understands itself to be, which is a much harder thing to remove in a bad quarter than a line item. I have no test for that and I am not going to pretend otherwise.

The decay figures are folklore. Stated above; restated here because it is the kind of number that gets reinserted by a helpful reader. The direction of travel is documented. The percentages are not.

And here is the study that would settle it. Take protected-time and continuous-improvement programmes across a few hundred firms, with dated launches, and measure how long each survived — against two events specifically: a change of sponsor, and two consecutive bad quarters. Then compare that survival against matched organisational initiatives of similar annual cost whose benefits are attributable to them. If the two groups die at the same rate, then this chapter has discovered only that initiatives end, which is true, well known and not worth a chapter. My expectation is that the unattributable-benefit programmes die substantially faster, and that sponsor tenure predicts a programme’s lifespan better than any measure of what the programme actually produced. As far as I can find, nobody has assembled that dataset, and it would not be hard to assemble.

What this does to the answer you were just given

Part II ended on a conclusion that felt complete. If the increment does better underwater and the gate can only damage it, then leave it there, protect it, let it run. I asked you to hold that as long as you could.

Here is the first thing wrong with it.

Protection is a programme. Whatever form it takes — protected hours, a sanctioned sandbox, an exemption, a tolerant manager who has decided not to ask questions — it is an arrangement that costs something legible and returns something illegible, and it therefore has a half-life. The graveyard is full of the good ones. And the payback period of the thing being protected is longer than the half-life of the protection, which is the whole problem in one sentence: Maria’s tool needs three years to prove what it is worth, and nothing in that building has lasted three years except a laminated poster.

There is a second thing wrong with it, and it is worse, and this chapter cannot pay it off.

Look again at how these programmes ended. Nobody cancelled twenty per cent time. Nobody cancelled Improvement Time. There was no meeting, no decision, no record, and no one to be angry at — the thing simply thinned out until it was in the past tense. Hold that shape in your mind, because it is going to come back in Chapter 9 wearing different clothes, and when it does you will find that a protection which is quietly withdrawn does more damage than a protection that was never offered. Everyone who used it learns something from the withdrawal. The learning is the expensive part, and it is not on any ledger.

So Part III opens with the fixes already existing and already failing, and the obvious next question is why. The proximate mechanisms are above and I think they are real. But there is a harder version of the question underneath them, which is this: these programmes get launched by senior people who understood the problem well enough to launch them, and they get allowed to lapse by senior people too, often the same ones, often without anybody noticing they have changed their mind.

The answer is not in the design of the programmes. It is in who is doing the deciding, and how those people got to be the ones deciding.

Next: Chapter 8 — why the fixes die, which turns out to be a question about who gets promoted.