An endurance race · Episode one
AI adoption is not a sprint and it is not a marathon. It is an endurance race — and almost nobody in the paddock has said out loud what winning would look like.
Le Mans runs for twenty-four hours. Hour eighteen lands the next morning — the night survived, nobody who has slept, the sun low in everyone's eyes, and fifteen hours of watching the same car pass the same board.
It is where the race is decided — and it is decided in both directions. The cars that stop, stop around here. So do the ones that go on to win.
Map it onto a four-year technology adoption and hour eighteen is year three — which is, not coincidentally, roughly when the consequences of the first year's decisions become measurable.
The clock recurs through this series. Each episode marks where on the lap it sits.
It is mid-morning and the sun is low enough to be a problem. Eighteen hours in.
Nobody in the garage has slept. Two of them lay down around four and got up worse. The night was the frightening part — the cold, the traffic, the hour when things break — and the night is over.
What is left is a car going round. It has been going round since yesterday afternoon and it will go round for another six hours, and there is a particular numbness that sets in somewhere past the fifteenth hour: not drama, not crisis, just the same car passing the same board at the same interval until watching it becomes something your eyes do rather than something you are doing. That is the first thing to understand, because it is the reason nobody is looking.
Nothing has gone wrong. The lap times are good, a shade off what they were at dusk. The driver has not complained. From the grandstand the car looks exactly like a car that is winning.
On the pit wall an engineer has had one channel open for two hours without really reading it. Oil temperature, four degrees above where it sat overnight, climbing slowly enough that no alarm has triggered. The screen is washed out in the glare; to see the trace at all you have to cup a hand over it and lean in. Nobody has, because there has been no reason to.
It was visible at hour eleven. At hour eleven the team was leading, and the conversation in the garage was about how much of the lead to protect.
The car will stop on the far side of the circuit in about forty minutes, and the debrief will be unanimous and correct: it should have come in at hour twelve, for a stop that would have cost ninety seconds.
Every organization now putting AI into software delivery is somewhere on this lap. Most of them are at hour six. It feels wonderful at hour six. That is the problem with hour six.
What makes the ending of that story inevitable is not the temperature. It is that nobody in the garage had written down, in advance, the number at which they would stop. Without it there was no moment at which anyone was obliged to act — only a slowly worsening reading that everyone could see and no one owned.
Which is the same failure, exactly, that sits underneath most enterprise AI programs. An initiative whose stated goal is to adopt AI succeeds by adopting AI. There is no state of the world in which it has failed, so there is no reading that obliges anyone to stop, and no moment at which it can be honestly evaluated.
So before any argument about constraints or bottlenecks or measurement — and this series will spend a great deal of time on all three — there is a prior question, and it is almost never answered in writing.
What, precisely, would winning look like? And would you recognize losing in time to do anything about it?
The metaphors in circulation are a sprint or a marathon. Both are wrong, and the way they are wrong matters, because each one licenses a different bad decision.
A sprint says: move now, move fast, the window is closing. It licenses buying tools before understanding the pipeline they are entering.
A marathon is the more sophisticated-sounding error, and the more common. It sounds like patience. But a marathon is a known distance, over a fixed course, run alone, in a single discipline, and then it is over. None of those five things is true here.
| Marathon | Endurance race | |
|---|---|---|
| Bounded by | A fixed distance you can train for | Time. There is no finish line |
| Competitor | One person | A team, with driver changes |
| Machine | None. Your body is the machine | Degrades under load, and you manage the wear |
| Strategy | Even pacing | Pit stops — stopping is how you go faster |
| Conditions | Broadly constant | Weather, darkness, traffic, safety cars |
| Winning | Being fastest | Still running at the end |
That final row does most of the work. At Le Mans a substantial share of the field does not finish. Attrition decides the result at least as much as pace, and the team that leads at hour six is frequently absent at hour twenty-four — because the same driving that produced the early lead is what put the car into the barriers or the engine into the gravel.
The pit-stop row is the one that reframes the technology argument. In endurance racing, deliberately stopping is how you win. A team that refuses to pit because it is currently leading does not stay leading. Translated: funding review capacity, specification capacity and operational headroom looks like slowing down, and it is the only thing that converts a faster build phase into a faster business.
Enterprise AI adoption does not fail in the build phase. It fails at hour eighteen — in the functions nobody upgraded.
Stated so it can be wrongThroughput gains from AI-accelerated build are absorbed by adjacent-stage constraints within roughly eighteen months of scaled adoption, such that end-to-end lead time improves materially less than build-stage time, while operational load rises approximately in proportion to change volume.
The second version is the one that matters. It can be measured, and it can come back false — in which case this series is wrong and you should disregard it. The first version is just the one you can remember at a board meeting.
Notice what the claim does not say. It does not say AI tooling fails to make developers faster. That part appears to be real and is not seriously in dispute. It says the gain is conditional on work in functions that did not buy the tool, were not consulted, and do not report to whoever declared the win.
Le Mans is not one race. It is several, run simultaneously on the same track, in classes. Most of the field is not competing for overall victory and knows it — they are racing their class, against comparable machinery, for a result that is real and achievable.
A team that does not know its class makes one of two mistakes. It burns itself out chasing a win structurally unavailable to it. Or it wins its class and cannot tell, because it was measuring against the wrong field.
"Become an AI-led organization" is entering the top class against competitors with more money, better data and fewer constraints. It is also unfalsifiable, which is why it survives contact with a strategy review.
"Reduce the cost of maintaining our legacy estate by 15% over three years" is a class entry. Unglamorous, specific, worth real money, and — this is the part that matters — capable of failing. You would know by year two whether it was happening.
A useful test for any AI objective: can it be falsified by an observable outcome within the planning horizon?
If not, it is not a goal. It is a direction of travel with a budget attached, and it will absorb whatever money is pointed at it without ever registering a result.
Watch any industry adopt anything and the striking thing is the convergence. The same tools, the same vendors, the same pilot-then-scale narrative, on roughly the same timeline. It looks like a failure of imagination. It is not.
Every paddock has a red car — the machine everyone wants to be seen in, chased for reasons that have rather less to do with what is under the bonnet than with who else is chasing it. The interesting question is not why it is attractive. It is why being in it feels so much safer than being right.
Consider a room in which everyone agrees the sun rises in the north. It rises in the east. Nobody in that room is right — and, critically, no one is wrong either. The error is real, and it is unattributable.
That asymmetry is the entire mechanism. Being wrong alone ends a career. Being wrong alongside every peer institution is a market condition — regrettable, widely shared, nobody's fault. Keynes put it exactly:
It is better for reputation to fail conventionally than to succeed unconventionally.
This is worth separating from the textbook account of groupthink, which describes a group genuinely persuading itself. What happens in most enterprises is colder and more rational. Half the room may privately know it rises in the east. They do not need to be convinced of anything. They need only decline to be the one who says so — and the outcome is identical to sincere consensus, at a fraction of the cognitive effort.
And the clearest evidence for all of this is the timing. Almost every large organization entered this race inside the same eighteen months, running comparable strategies, arriving at comparable positions. If adoption were driven by organization-specific reasoning you would expect a scatter — some three years in, some yet to start, some having looked and declined. Instead there is a peloton. The synchronization is not a coincidence to be explained away; it is the phenomenon, visible from orbit.
Which tells you something useful about the remedy: more information will not fix it. Everybody already has the information. The thing missing is not insight. It is attribution.
If attrition decides endurance races, there is an obvious strategy: let the fast starters break, run your own pace, inherit the lead. It is a real strategy and it wins real races.
Applied here, the argument is strong. Tooling depreciates quickly — wait eighteen months and the capability is better, the price is lower, and someone else has publicly documented the failure modes. The second mover buys a more mature product with a shorter list of unknowns.
And it is almost never available, because no executive can say "we are waiting" out loud and retain credibility, even when waiting is correct. The strategy is ruled out by the reporting structure rather than by the analysis — which is worth naming plainly, because it is the same mechanism as the red car, one level up.
But there is a genuine objection, and it is not the political one. Watching is only a strategy if you are actually watching.
A team in the pits — measuring, tuning, building the capacity it will need — is running a race. A team in the paddock is not competing. In practice "fast follower" is very often a euphemism for doing nothing and then panicking in year three, which delivers all the costs of late adoption and none of the benefits of deliberate patience.
The distinction is concrete and testable: are you building absorptive capacity while you wait? Review capacity, specification capacity, operational headroom, and the instrumentation to know when they are binding. If yes, waiting is strategy. If no, waiting is just being late with extra steps.
If collective error persists because it is unattributable, then the fix is not better analysis. It is putting a name on a position before the outcome is known.
Concretely: write down, in advance and with a signature against it, the result that would make you stop or scale back. Not a target — targets are easy and get quietly revised. A failure condition. "We stop if rework rises 25%." "We reassess if incident volume doubles." "We were wrong if lead time has not moved by Q4 of year two."
Three things happen when that document exists. The goal becomes falsifiable, which makes it a goal. The evaluation cannot be retrofitted to whatever the data happened to do. And the decision acquires an owner — so being wrong together stops being free, which is the only pressure that reliably changes the behavior.
It is an unusually cheap intervention. It costs one meeting and a page, and it has to happen before the first tool is bought, because its whole function is to be written while the answer is still unknown.
Eight episodes. The frame is set; from here each one takes a single stage of the pipeline, in the order the pressure reaches it — and each one closes with the line through that corner.
The first half is diagnosis, because you cannot fix a constraint you have not located. From Episode 5 the series turns: pit strategy, managing the wear, who is in the seat, and what to put on the pit wall. If you want the short version of where it lands — this is winnable, and the methods are known.
Everything above describes how this goes wrong, and it would be a poor argument if going wrong were the only available outcome. It is not. People win this race, repeatedly, and the way they do it is neither secret nor accidental.
Audi won Le Mans thirteen times between 2000 and 2014. Not by building a car that did not break — at racing load for twenty-four hours, everything breaks somewhere. They built one that could be repaired faster than anyone else's. The R8's entire rear end, gearbox and suspension together, was engineered to unbolt and be swapped as a single unit in minutes, where rivals needed hours for the same job.
So when something failed, and it did, Audi lost a pit stop. Their competitors lost the race.
They did not assume the machine would hold. They decided in advance where they could afford it to fail, and engineered so that failing cost ninety seconds instead of the weekend.
That is the entire discipline, in one design decision. It is also, almost exactly, the thing large organizations do not do with technology adoption — where the plan is generally that nothing will go wrong, and the contingency is that someone will notice in time.
The useful question was never will this work. It is: when the constraint moves — and it will — how fast can we move with it? That is answerable, it is answerable in advance, and it is what the rest of this series is about.
What the teams that get this stage right do differently:
None of these require knowing whether AI works. They require knowing what you are racing for — which is available to you today, at no cost, and is the one thing that cannot be retrofitted later.
None of this is an argument against adopting AI. The build phase really is getting faster, and an organization that ignores that will be beaten by one that does not.
It is an argument that the gain is conditional — on work that has not been budgeted, in functions that did not ask for it, over a timescale longer than the one the decision is being judged against. And that the first condition, before any of the others, is being able to say what you are racing for.
The sun still rises in the east. Agreement in the room changes nothing about that — only who has to answer for it.
At hour twelve there is a stop that costs ninety seconds. It is available to every organization reading this, right now, at a price that will never again be as low. The difficulty was never affording it.
Next episode: The car is a system — why accelerating one stage of a constrained pipeline relocates the constraint rather than removing it.
Draws on Theory of Constraints (Goldratt), queueing theory for product development (Reinertsen), the Abilene Paradox (Harvey), and Keynes on conventional failure.
© 2026 Naveen Dewan · Left Coast Services. All rights reserved.