Companies do not accidentally lose the people who would have spotted the failure in their information system. They filter them out on purpose, using recruitment and performance processes they are proud of. The traits that make someone hard to sit in a meeting with, refusing to close a subject, following a decision down to its fourth consequence, hearing what nobody said out loud, are the same traits that detect low probability, high cost failure modes years before a dashboard does. An executive who wants fewer surprises in his systems has to look at who his organisation finds exhausting, and why that person stopped speaking.
Organisations do not select for insight, they select for low friction
Every hiring process optimises for something. Very few optimise for what they claim.
Ask a management team what it recruits for and you will hear competence, ownership, culture. Watch the same team decide, and the actual criterion surfaces fast: how easy will this person be to work with, week after week, in a room where twelve subjects have to be closed before lunch.
That criterion is not stupid. An organisation that cannot close subjects does not ship. Friction has a real cost, paid daily, visible to everyone. The cost of the opposite failure, closing a subject that was not resolved, is paid once, later, by someone else, often by a successor.
Lauren Rivera documented the mechanism in elite professional service firms. Evaluators consistently favoured candidates who felt culturally familiar, and constructed their definition of merit around that familiarity. They were not lying about wanting the best. They genuinely could not separate “best” from “reads like us”.
The output is an organisation with a narrow band of cognitive style. Everyone in the room processes information roughly the same way, at roughly the same speed, with roughly the same tolerance for open questions. Meetings get shorter. Alignment gets easier. Risk detection gets worse, because risk detection is a minority function. It requires someone who is willing to be the only person still talking about a subject the room has already agreed to leave.
Cass Sunstein and Reid Hastie described what happens to groups when that person is absent or silent. Errors get amplified instead of corrected. Early speakers set the direction. Information held by a single member never reaches the table. A group of intelligent people reliably produces a decision worse than its best member would have produced alone.
None of this is a diversity argument in the moral sense. It is a detection argument. A system that filters for low friction is a system that has removed its own alarm.
The traits an organisation corrects are the traits that read a system
Look at what actually goes into performance reviews.
Goes off on tangents. Struggles to prioritise. Gets stuck on details. Overthinks. Too direct. Reads too much into things. Does not switch off.
Now look at what an information system failure actually requires, in order to be seen before it costs money.
Going off on tangents means walking the consequence tree of a decision past the point where the vendor’s demo stops. Getting stuck on details means refusing to accept “it restarted, it is fine” as an explanation of why a service stopped. Reading too much into things means noticing that the project manager has stopped escalating, which has never meant everything is fine. Not switching off means covering ground on a Sunday that nobody covered in three years of steering committees.
The overlap is not partial. It is close to total. Every trait a company spends money correcting is a trait that, pointed at a system, produces early detection.
This is where most writing on the subject goes wrong. It turns into a celebration, a claim that unusual minds are better, a request for recognition. That framing is useless to an executive, and it is not the point. The point is narrower and colder: a company that trains this behaviour out of its people has traded a permanent detection capability for a modest gain in meeting comfort, without ever putting the trade on a slide.
Robert Austin and Gary Pisano made a version of the same argument in Harvard Business Review, and it is worth reading for the operational detail. Their case is that standard hiring processes screen out capabilities the firm actually needs, then the firm pays consultants for those capabilities later. The screening is not malicious. It is procedural, and it is invisible from the inside.
The filter operates in four moments, and none of them looks like a filter
Executives who accept the argument in principle usually reject it in practice, because they cannot see where the filtering happens. Nobody in their company has ever written “reject candidates who ask too many questions” in a hiring guideline. The filter is real anyway. It operates in four moments, and each one is defensible on its own.
The first is the screening call. A recruiter with fifteen candidates and forty minutes each is measuring one thing above all others: is this conversation smooth. A candidate who answers a question with a better question, or who needs to understand the business before answering a technical one, is producing friction at the exact moment friction is most expensive to the recruiter. That candidate does not fail. He is simply not moved forward, which is a different thing and leaves no trace.
The second is the probation period. New arrivals who ask about things outside their remit in their first three months are read as either promising or presumptuous, and the reading depends almost entirely on how confident their manager is. An insecure manager experiences the same behaviour as a challenge to authority. Nothing in the process distinguishes the two interpretations, and the outcome is recorded as a fit issue.
The third is the annual review, which is where most of the damage happens, precisely because it is the moment everybody trusts. The vocabulary is developmental. Learn to prioritise. Work on your executive presence. Pick your battles. Every one of those sentences is well intentioned, most are individually reasonable, and their aggregate effect over three years is a person who has learned to stop mentioning things.
The fourth is promotion. Organisations promote people who make the level above comfortable. It is not a conspiracy, it is a preference expressed under time pressure by humans who will have to work with the result. The consequence is that the further you go up the structure, the fewer people you will meet whose instinct is to reopen a closed subject, which is another way of saying that your detection capability decreases with seniority exactly where your decision stakes increase.
Four moments, four defensible decisions, one outcome. Your organisation has removed its ability to see what you cannot see, and it did so using processes you would defend in front of a board.
The only way to notice from the top floor is by absence. Nothing gets flagged. Nothing appears on a risk register. You simply stop receiving a category of information, and you cannot miss what never arrives.
The Post Office spent two decades proving that the system was never questioned by anyone the organisation could hear
The clearest case I know of an organisation destroying its own detection capability is the Horizon scandal in the United Kingdom.
Horizon is the accounting and point of sale system supplied by Fujitsu and deployed across Post Office branches from 1999. Sub-postmasters, the people who actually run those branches, began reporting shortfalls they could not explain. Money that the system said was missing, that they had not taken, that appeared and reappeared across branches with no common cause.
The Post Office response, sustained over roughly two decades, was that the system was reliable and the sub-postmasters were not. More than 900 people were prosecuted for theft, fraud or false accounting on the strength of Horizon data. People went to prison. People lost homes. Some died before their convictions were overturned.
Read the structure of that failure rather than its scale, because the structure is what repeats.
The organisation had a large population of front line users reporting a consistent, reproducible anomaly. That is, in engineering terms, an excellent signal. Hundreds of independent observers, same symptom, different geographies, no coordination between them. Any system owner receiving that pattern should treat it as the highest quality bug report available.
The organisation reclassified it. The signal became a personnel problem: dishonest or incompetent postmasters. Once that classification was made, every further report confirmed it. More reports meant more dishonest postmasters, never a more probable defect. The reclassification was self sealing.
Inside the organisation, the people who could have broken the seal were exactly the people whose input was structurally hard to hear. Raising it required going off topic, questioning a system the board had already declared sound, and doing so repeatedly, against colleagues who had a professional interest in the subject staying closed. It required, in short, every behaviour a well run corporate performance process discourages.
The English High Court found in 2019 that Horizon contained bugs, errors and defects, and that remote access to branch accounts had been possible, contrary to what had been asserted for years. A statutory public inquiry chaired by Sir Wyn Williams has since spent years reconstructing how an organisation reached that state and stayed there.
What matters for an executive reading this is not the legal outcome. It is the timeline. The information required to prevent the entire disaster existed inside the organisation from the early years. It was not missing. It was present, repeated, and structurally unhearable, because hearing it meant reopening a subject that senior people had closed, and because the individuals inclined to keep reopening it had either been managed out, moved on, or learned what happens to people who do that.
Nothing about Horizon required exceptional insight. It required someone with the standing to be exhausting, and a room willing to sit through it.
Every company I look at has a smaller version of this. Not with the same human cost, and rarely with the same duration. Same shape. A recurring anomaly that has been explained away often enough that the explanation has become the official account. A vendor whose reassurances are accepted because reopening the question implies the original decision was wrong. A team whose warnings arrive in a format the executive committee does not process, and stop arriving after a while.
The organisation is not lying to its CEO. It has simply optimised itself into a state where the truth about its systems has no route to the top floor.
Cultural fit is a technical risk, not an HR preference
Executives treat hiring style as a people topic, delegated to HR, reviewed once a year. It belongs on the risk register.
When you standardise the cognitive profile of everyone who touches your critical systems, you have made an architecture decision. You have decided that your detection layer will be composed entirely of people who process the world the same way, which means they will miss the same things, in the same direction, at the same time. Correlated sensors. Any engineer would flag that design.
Katherine Phillips summarised the research on the mechanism in Scientific American: groups that contain genuine difference prepare more carefully, examine facts harder, and reach better conclusions, while reporting that the experience felt worse. Homogeneous groups report smoother meetings and produce weaker decisions. The subjective signal is inverted relative to the actual quality of the output.
That inversion is the whole trap. Your comfort in the room is anti correlated with the quality of the scrutiny in the room. If your steering committee feels easy, that is information, and it is not good news.
Charles Perrow’s work on complex systems, and Diane Vaughan’s reconstruction of the Challenger launch decision, both land on a related point from a different direction. Catastrophic outcomes in tightly coupled systems rarely come from a single dramatic error. They come from an accumulation of small deviations, each individually justified, each accepted because the last one was accepted. Vaughan called it the normalisation of deviance. Stopping that accumulation requires someone whose reaction to “we have always done it that way and it has been fine” is to ask why, again, for the fifth time.
That person is expensive to manage and cheap to ignore. Ignoring them costs nothing today.
What this actually costs, in money you already recognise
Abstractions do not move executives. Invoices do.
The unexplained infrastructure budget increase, where nobody can name the decision that caused it, is a consequence tree that was never walked. Somebody chose a product, the demo covered the first branch, and the cost of the sixth branch arrives eighteen months later as a line item with no owner.
The migration that slipped by two quarters was visible in team behaviour before it was visible in the plan. Somebody heard it. That person either did not have a route to you, or had learned that using it produced friction and no result.
The vendor you cannot leave was a reversibility clause that nobody read closely, because reading it closely meant slowing down a deal the whole room wanted to close.
The outage on a component nobody knew was critical was documented, once, by someone who was told the subject was resolved.
Each of these has the same root. Not incompetence. Not bad luck. A detection function that was removed because it was uncomfortable, and an executive who never saw the removal happen, because it happened in performance reviews and hiring loops, one polite conversation at a time.
What a CEO can change without hiring anybody
The useful part of this argument is that it does not require a recruitment programme.
Start by finding out who in your organisation used to raise things and stopped. Not who complains. Who used to, and no longer does. That transition is recorded in the memory of every team, and it is not in any of your systems. The people who went quiet are your decommissioned sensors.
Then look at your last three significant technical decisions and ask a single question about each: who in the room was allowed to be tiresome, and for how long. If the honest answer is nobody, or five minutes, you did not make a decision. You ratified a proposal.
Then check what happens to a disagreement in your organisation when it survives past the first meeting. In healthy structures it escalates and gets arbitrated. In most structures it dies quietly, and the person who raised it learns something about the cost of raising things.
Amy Edmondson’s work on psychological safety is the academic version of this, and it is worth noting that her finding is not about being nice. It is about whether interpersonal risk taking is survivable, because learning behaviour depends on it. A team that cannot survive its own dissent cannot detect its own failures.
None of this requires a budget line. It requires an executive willing to be told things that make the week longer.
The uncomfortable person is the control you removed
There is a version of this article that ends as an appeal for tolerance. That version is worthless to you.
Here is the useful ending instead.
Your information system will fail in a way you did not anticipate. Not maybe. The only variables are when, how expensive, and whether anyone saw it coming and could not get through.
The probability that someone saw it coming is high. Organisations rarely fail from a total absence of information. They fail from information that had no route.
The traits that produce that information, the tangents, the stubbornness, the register switching, the reading of what was not said, the mind that does not stop at 18h, are traits your processes are actively grinding down right now, in good faith, with the language of professional development. Every one of those conversations is a small, deliberate, well intentioned reduction of your own visibility.
You do not have to like the person. You do not have to promote them. You have to know where they are, and make sure that when they say the subject is not closed, the sentence reaches you before the invoice does.
Sources
- Lauren A. Rivera, Hiring as Cultural Matching: The Case of Elite Professional Service Firms, American Sociological Review 77(6), 2012. https://journals.sagepub.com/doi/10.1177/0003122412463213
- Robert D. Austin and Gary P. Pisano, Neurodiversity as a Competitive Advantage, Harvard Business Review, May 2017. https://hbr.org/2017/05/neurodiversity-as-a-competitive-advantage
- Cass R. Sunstein and Reid Hastie, Making Dumb Groups Smarter, Harvard Business Review, December 2014. https://hbr.org/2014/12/making-dumb-groups-smarter
- Katherine W. Phillips, How Diversity Makes Us Smarter, Scientific American, October 2014. https://www.scientificamerican.com/article/how-diversity-makes-us-smarter/
- Diane Vaughan, The Challenger Launch Decision: Risky Technology, Culture, and Deviance at NASA, University of Chicago Press, enlarged edition 2016. https://press.uchicago.edu/ucp/books/book/chicago/C/bo22781921.html
- Post Office Horizon IT Inquiry, chaired by Sir Wyn Williams, official inquiry site. https://www.postofficehorizoninquiry.org.uk/
FAQ
How does a normal hiring and performance review process end up removing a company's ability to detect risk?
It operates in four moments that each look defensible on their own: the screening call filters out candidates who create friction by asking hard questions; the probation period reads outside-remit questions as presumptuous; the annual review uses developmental language (‘learn to prioritise,’ ‘pick your battles’) that teaches people to stop mentioning things; and promotion favors people who make the level above comfortable. None of these is written down as a filter, but the aggregate result is an organization that has quietly removed the people most likely to catch a failure before it becomes expensive.
What did the UK Post Office Horizon scandal reveal about organizational blind spots?
Hundreds of sub-postmasters independently reported the same accounting anomaly over two decades — an excellent, reproducible signal by any engineering standard. The organization reclassified that signal as a personnel problem (dishonest postmasters) instead of a system defect, and every further report then confirmed the reclassification rather than questioning it. The people who could have broken that self-sealing loop were exactly the ones whose insistence on reopening a closed subject a well-run performance process is designed to discourage.
