A verdict on someone else’s system gets more confident as the context behind it gets thinner. That is not a paradox, it is the mechanism: the less an observer can see of the constraints a system was built against, the fewer competing explanations they have to weigh, and the cleaner the judgement feels. Executives are surrounded by people who deliver confident verdicts on their technology from a position that structurally hides the reasons. The cost of that arrangement is not the noise. It is the day a leader acts on a verdict that was cheap to produce, and pays to undo something that was working.
Confidence in a verdict is inversely related to the context behind it
The strength of an opinion about a system tends to rise as knowledge of that system falls. Anyone who has sat in a review meeting has watched it happen. The person with three weeks of tenure has the sharpest opinion in the room. The person who has been running the thing for nine years hesitates, qualifies, says “it depends on what you mean by”. The newcomer sounds competent. The veteran sounds evasive.
That reads backwards, and it is not backwards at all. The veteran hesitates because they hold five constraints in their head at once and none of them fits in a sentence. The newcomer is decisive because they hold one variable: what a modern setup usually looks like. A single variable produces a clean answer. Five produce a heavy one.
So the market for opinions is rigged in a specific direction. The verdict that carries the least information is the one that travels best. It is short, it is portable, it is memorable, and it makes the person delivering it look sharp. The explanation of why the system is the way it is has none of those properties. It is long, it is local, it is full of people who have left the company, and it requires the listener to sit through a story.
Leaders do not choose between a good verdict and a bad one. They choose between a fast verdict and a slow explanation, and the market has already decided which one gets to them first.
An observer never experiences the missing information as missing
The hardest part of this is not that outsiders lack context. It is that lacking context feels exactly like having it.
Think about the last time you sat behind a car that did not move on a green light. The gap between what you could see and what the driver could see was total, and at no point did you experience a gap. You experienced a complete picture of a road, a car, and a driver doing nothing. There was no flicker, no warning, no sensation of a hole in the data. The pedestrian, the ambulance, the stalled vehicle: none of them registered as absent, because absence does not register.
You then judged the complete picture, with the confidence a complete picture deserves.
That is the whole problem, transposed to organisations. When a consultant walks into a company and sees a spreadsheet holding a critical process together, they do not see a spreadsheet plus an unknown number of reasons. They see a spreadsheet. The contract clause that constrains where the data lives, the regulator that has to be able to reconstruct a calculation, the two customers whose file formats have not changed since 2011, the failed migration of 2019 that cost a quarter: none of that is in the frame, and none of it announces itself as being out of frame.
Abraham Wald’s work on aircraft armour during the Second World War is the cleanest illustration of the shape of this error. The Statistical Research Group at Columbia was asked where to add armour on bombers, and the available data was damage patterns on the planes that came back. The instinct is to reinforce where the holes are. Wald’s contribution was to reason about the planes that were not in the sample at all, and to conclude that the undamaged areas on returning aircraft were the ones worth protecting, because hits there were the hits nobody survived to report.
The lesson usually gets filed under statistics. It belongs in the executive’s toolkit as something simpler: the data that reaches you has been filtered by a process you did not design, and the filter is invisible from where you sit.
Chesterton wrote the rule down in 1929, and almost nobody applies it to systems
There is a well known principle for exactly this situation, and it predates every technology it now applies to.
G. K. Chesterton, in a 1929 collection, put it as a reformer arriving at a fence built across a road. The reformer says the fence is useless and proposes to clear it away. The reply is that the fence should not be touched until the reason for building it has been established, and that the person who cannot see any use for it is the person least qualified to remove it.
The principle has been passed around long enough to become a slogan, which is usually where a good idea goes to die. The slogan version is “do not remove a fence until you know why it is there”. The operative version is harder and much less comfortable: the inability to see a reason is evidence about the observer, not about the fence.
Applied to technology, that reverses the burden of proof that most organisations run on. The default today is that anything unusual is guilty until justified. Legacy is a pejorative. On-premise infrastructure is presumed to be inertia. A spreadsheet in a critical path is presumed to be negligence. In each case the organisation carrying the thing is expected to defend it, and the visitor rendering the verdict is expected to do nothing at all.
Flip the burden and the conversation changes. The person proposing the change owes an account of what the current arrangement was solving. Not a guess, an account. If they cannot produce one, they have not finished their work, and the leader has no business acting on their recommendation yet.
This is not conservatism. A reason that is found and then judged insufficient is a reason to change, and the change is now safe, because everyone knows what they are giving up. What the principle forbids is not change. It is change decided by someone who never established what the thing was doing.
Attribution research explains why the reflex is not laziness
The reflex that produces the confident verdict has been studied for decades, and it is a feature of normal cognition rather than a defect of arrogant people.
Jones and Harris ran an experiment in 1967 in which participants read essays defending or attacking Fidel Castro. Participants were told the writers had been assigned their position by a coin toss. They still concluded that a writer defending Castro must personally favour Castro. Knowing the constraint explicitly was not enough to stop them reading the behaviour as character. Lee Ross named the pattern the fundamental attribution error a decade later: observers overweight disposition and underweight situation when explaining what other people do.
There is a detail in that research that matters more to a CEO than the headline. The bias is asymmetric. People explain their own behaviour by circumstances and other people’s behaviour by character. You know why you kept the old system: the budget, the timing, the customer who threatened to leave. You look at another company’s old system and see a leadership that never got around to it.
Every outsider who walks into your organisation is running that asymmetry against you. Not maliciously. Structurally. They have access to your artefacts and no access to your circumstances, which is exactly the condition under which the bias is strongest.
Which means that the confident tone in the room is not a signal of quality. It is a signal of distance.
A case: the system everyone agreed was backward
Here is a composite case, anonymised and assembled from patterns seen across several organisations rather than from any single engagement. Nothing in it identifies a client.
A mid-sized industrial company, family owned, a few hundred people, three production sites. The core of the operation was an order and scheduling system that had been built in house over a long period and extended by three generations of internal staff. It ran on an operating system version well past its fashionable life, on hardware in a room at the main site. Around it sat a constellation of spreadsheets, each of which did one specific thing, each of which had a name that made no sense to anyone hired in the previous five years.
Every visitor to that company delivered the same verdict within an hour. Backward. Risky. Unmaintainable. A liability at exit. The bank’s advisors said it. Two consulting firms said it, one of them in a formal document. A new commercial director said it in his first month, loudly, and was widely agreed with.
What none of them did was ask why any individual piece was the way it was.
The scheduling logic, when someone finally sat down with the two people who understood it, encoded about fifteen years of accumulated rules about what could be run in which order on which line, given tooling changeover times, given a handful of customers with contractual delivery windows, and given a specific set of quality constraints that applied only to certain material batches. None of it was documented, because each rule had been obvious to the person who added it. The set of them, taken together, was the actual competitive advantage of the company: it was why they could quote delivery dates their competitors could not, and hit them.
The spreadsheets were not decoration either. Two of them existed because a large customer sent data in a format that no packaged system on the market would ingest without a bespoke connector, and the spreadsheet was the bespoke connector, written by someone who had left in 2016. One existed because the finance team needed a specific reconciliation that the accounting package produced in a form the auditors had objected to years earlier. One existed for no reason anybody could reconstruct, and it turned out to genuinely be dead weight.
The on-site hardware had a reason too, and it was the least interesting one: it had been bought, it had been paid for, it was fully depreciated, and it worked. Moving to a subscription model would have converted a settled asset into a permanent monthly line at a moment when the company was funding a production investment. That is a treasury argument, not a technology argument, and it had been made deliberately by a finance director who understood it perfectly.
Under pressure from the accumulated verdicts, the company acted. A replacement programme was launched, with a packaged solution and an integrator, on a timeline of nine months. The business case rested on modernisation and risk reduction. It did not rest on a specific operational problem, because there was not one that anybody had named. The trigger was reputational: leadership had grown tired of defending the setup to every visitor.
The programme did what those programmes do when the reasons were never collected. The scheduling rules were discovered one at a time, in production, by failing. Each discovery produced a specification change. The two people who held the knowledge became a bottleneck and then became exhausted, and one of them left during the second year, which converted a bottleneck into a permanent loss. Delivery date commitments to customers degraded during the transition, quietly at first, then visibly. The integrator was not at fault in any way you could put in a letter: they built what was specified, and the specification was incomplete because the reasons had never been written down anywhere.
The programme finished, late and well over its original envelope. The company ended up with a modern system that could not yet do what the old one did on the specific thing that made them money, and a plan to rebuild that capability over the following year.
Count the cost properly and it has four parts. There is the direct overrun, which is the only one that appears in a board pack. There is the operational damage during transition, which shows up as customer irritation and later as pricing pressure. There is the loss of the people who held the undocumented logic, which is permanent. And there is the opportunity cost of two years of management attention spent on a project that was not solving a stated problem.
The uncomfortable part is the counterfactual. There was a real version of this project. A serious reading of that system would have found the dead spreadsheet, would have documented the fifteen years of scheduling rules while both holders were still there, would have isolated the genuine risks, and would then have replaced the pieces that deserved replacing in an order that protected the core. It would have cost a fraction. It was never proposed, because proposing it requires spending weeks establishing why things are the way they are, and nobody in the chain was paid to do that.
Everyone in that story behaved reasonably. The visitors gave the verdict their position allowed them to give. Leadership acted on an accumulation of expert opinion, which is what leadership is supposed to do. The failure was structural: not one of the confident opinions had been produced by someone who had looked at the problem from in front of the car.
The cost of a surface verdict is asymmetric and lands outside the accounts
The economics here are worse than they look, because the two errors are not symmetric.
Keeping something too long produces a cost that is visible, gradual, and attributable. The maintenance is expensive, the risk is discussed, everyone knows what is happening. It is a bad situation that the organisation can see.
Acting on a surface verdict produces a cost that is mostly invisible in the accounts. The overrun is visible. The rest is not. The knowledge that walked out the door does not have a line item. The delivery reliability that degraded does not get attributed to the programme. The management attention has no price. And the counterfactual, the cheaper project that was never proposed, is invisible by construction: nobody misses a proposal that was never made.
That asymmetry is exactly why the confident verdict wins. Its failures do not show up where decisions are reviewed.
For a leader, the practical consequence is narrow and usable. Before authorising work that removes something which is currently functioning, the question that matters is not whether the thing is good. It is whether anybody in the room can state what it was solving.
Four questions surface the reasons before the verdict is worth anything
The discipline required here is small, and it fits in a meeting.
Ask what breaks if this disappears tomorrow morning. Not in theory: name the customer, name the process, name the person who calls. A system nobody can describe the failure of is either dead weight or badly understood, and the answer tells you which.
Ask who decided this, and what they were afraid of. Every unusual arrangement was once a decision made by someone with a fear. The fear is usually still documented in the shape of the thing even when it is documented nowhere else.
Ask who holds the knowledge, and how long they have been here. If the answer is one person with twenty years of tenure, the risk in the room is not the technology. The risk is that the replacement programme is also the event that makes that person leave.
Ask the person delivering the verdict what constraints they checked. Not to embarrass them: to calibrate. An adviser who answers “the contracts, the volumes, the two regulatory windows, and I spoke to the two people who built it” is giving you an opinion worth acting on. An adviser who cannot answer has given you an impression, and impressions are free.
None of this prevents change. It changes who carries the burden of proof, and it moves the burden to the only place where it belongs: onto the person proposing to remove the fence.
The verdict that cost nothing to produce is the one to distrust
A judgement about a system is only as good as the context behind it, and the tone in which it is delivered carries no information about that context at all. That is the trap. Certainty is cheap to perform and expensive to earn, and the two are indistinguishable at the moment a leader has to decide.
There is one asymmetry to lean on. It is always possible to ask what an opinion cost to produce. Who was interviewed, what was read, which constraints were checked, what was found that changed the initial impression. The answer arrives fast, and it sorts the room.
The organisations that pay the most for their technology are rarely the ones that moved too slowly. They are the ones that moved confidently, on the advice of people who were standing behind the car.
Sources
- G. K. Chesterton, The Thing: Why I Am a Catholic (1929), chapter “The Drift from Domesticity”, where the fence principle is stated: https://en.wikipedia.org/wiki/G._K._Chesterton
- Fundamental attribution error: Jones and Harris (1967), phrase coined by Lee Ross (1977), including the actor-observer asymmetry: https://en.wikipedia.org/wiki/Fundamental_attribution_error
- Abraham Wald and the Statistical Research Group memoranda (1943) on estimating aircraft vulnerability from returning planes: https://en.wikipedia.org/wiki/Survivorship_bias
- 37signals, “Leaving the Cloud”, a documented case of an unfashionable infrastructure choice made on explicit economics rather than defaults: https://basecamp.com/cloud-exit
FAQ
What is Chesterton's fence, and how does it apply to technology decisions?
G.K. Chesterton’s 1929 principle says a fence found blocking a road should not be removed until you know why it was built — and that the person who can’t see a reason for it is the person least qualified to remove it. Applied to technology, this reverses the usual burden of proof: instead of an organization defending a legacy system to every visitor, the person proposing to remove it should have to produce an account of what the current arrangement was actually solving. Change is still allowed — a reason that’s found and judged insufficient is a legitimate reason to change — but a decision made by someone who never established what the thing was doing isn’t safe change, it’s a guess.
Why do outside advisors and new hires often deliver more confident verdicts on a system than the people who built it?
Because confidence in a verdict rises as knowledge of the system falls — the person with one variable (what a modern setup usually looks like) produces a clean answer, while the person holding five real constraints in their head produces a hedged one. This is the fundamental attribution error studied by Jones and Harris in 1967: observers explain their own choices by circumstances but explain other people’s choices by character, and outsiders run that asymmetry against an organization structurally, with access to its artifacts but none of its context.
