Our goal is to give our teams back a day a week to focus on higher value work. Sooo… we’ll need about a hundred more tasks with a hundred more projects just like that one.

Where’s the ROI in all of that?

Subscribe now

It’ll never happen. Even if we cut the time to deliver those projects by 90% it would still take us 10 years. And even if we could pull that off, time savings doesn’t really sum that way. The funny thing about efficiency is that it requires automation but never seems to come from automation.

So even as the world works itself into an automation frenzy and SaaS providers blindly jump on the agentic bandwagon, the ROI question keeps tugging at us all. It seems like someone heard there was “gold in them hills” so we collectively decided to dig up the entire mountain range. It’s true that we’ll eventually find gold but there must be a more efficient way of finding it. The great irony here is that brute forcing our way to efficiency is incredibly inefficient. So where is ROI to be found?

Right in Front of My Face

The other day I was talking to Maureen McCusker about an ROI model she’s been working on. In it she did something brilliant that made me rethink the problem. She included a conversion efficiency coefficient. It’s an empirical discount on the number of minutes you get back from automating a task that takes X minutes. We all know you never get X minutes back but she decided to capture that in her assumptions. The time you recover depends on everything surrounding the task: the coordination between stakeholders, where it sits in a sequence of other work, and how critical it is that a human validates the result.

She was accounting for the work automation didn’t touch. It was a brilliant idea that led me down a rabbit hole.

A Theorem from 1967

In 1967 a computer architect named Gene Amdahl made an argument against the idea that you could get unlimited performance by throwing more parallel processors at a problem. His observation was elegantly simple. If some fraction of a process can be sped up and some fraction cannot, then the fraction that can’t sets a hard ceiling on the time savings. Amdahl observed that the maximum speed of a process depends more on the non-automated tasks than the automated ones.

Here’s what that means in practice when we’re talking about business processes. Imagine a one-hour process where everything can be automated except a 15-minute approval step. Two vendors offer a solution that can automate the automatable portion. Vendor A costs $1,000 and it reduces the automatable portion from 45 minutes to 30 seconds. Vendor B costs $10,000 and reduces it to 3 milliseconds. Is the better automation the better option? Depends on how you look at it.

If we look at just the automatable portion we’ll see that Vendor B is 10,000x faster.

However if we look at the entire process we’ll see that both vendors reduced the time from one hour to about 15 minutes. When looking at it this way Vendor B is only about 3% faster but $9,000 more expensive. If you’re designing high frequency, high volume computational processes, that 3% might matter. If you’re designing human mediated business processes, it doesn’t.

The 15 minutes of non-automatable work (the NAW) sets a limit on the amount of ROI we can possibly excavate from this process. Even if we perfectly automated everything that could be automated down to exactly zero seconds, we’ll never be faster than the NAW.

Amdahl is telling us that once the automated portion is small relative to the NAW, additional speed doesn’t provide more value. Don’t pay for it. Vendor B isn’t a worse product. It’s a better product solving a problem you don’t really have.

Ready or Not Here We Come

Amdahl’s ceiling isn’t a law of the universe and it’s not the result of natural processes. Finding efficiency and delivering ROI can be more like playing hide-and-go-seek with yourself than mining for gold. You decide where the value will be found in the project charter.

Think about hiring. Teams have been automating job descriptions, resume screening, interview scheduling, and offer paperwork for decades and yet, for everything but the simplest hourly roles, time-to-fill hovers stubbornly above 40 days for most teams. Why? Because intake meetings, interview loops, offer approvals, and other alignment steps consume most of those days. The process will never be faster than its least automated steps.

What’s to be done? Instead of investing in faster and sleeker interview scheduling, go after the NAW: fewer interviews, fewer approvals, simpler offers. These parts of the process are more often a product of a company’s culture than of careful process design, which means engineering the NAW will often cost you more in political capital than in AI tokens.

Amdahl’s ceiling is a consequence of scope. Once you decide that change management, approval chains, and stakeholder alignment are out of scope, you’ve chosen your ROI ceiling. Everything you deploy, no matter how good the technology, is subject to that limit.

Everything Is an Agent Now

I thought all of this was almost too obvious to write down until I considered what it implied about the pitches I see from vendors on a weekly basis and the ROI that they promise.

One of my most important responsibilities is selecting software that will make an impact. I sit through a lot of vendor demos. And over the past year, nearly every one of them has featured the same transformation: workflows that were plain, deterministic automation last year are now performed by “agents.” Same workflow. Same output. New name, and a new pricing model to go with it. Agents consume tokens every time they run, and tokens are billed on consumption. It’s Amdahl’s nightmare.

This is the most expensive possible way to solve an already-solved problem. Classic automation costs something to build and nearly nothing to run. An agent doing the same job costs something every single time it’s invoked, forever. Amdahl would tell us that if the deterministic version already collapsed the automatable portion to seconds, the agentic version can’t make the process meaningfully faster. You’re paying consumption prices for the illusion of an improvement if the NAW hasn’t changed.

It’s important to understand this conflict of interests. Per-seat pricing is eroding and consumption-based pricing is the lucky strike vendors need to sustain their revenue growth. You are both looking for gold, just in different places. Making everything an agent means they can charge you more for the same outcome. Their incentives and your ROI have always been in competition, and no amount of partnership language in the contract has ever changed that.

For every agentic pitch you hear, ask yourself three questions. First: can classic automation produce the same output? If yes, the conversation is over. Second, does the agentic capability address your NAW? Can it handle the judgment calls, unstructured inputs, and coordination tasks that make up your NAW? Third: if it can do those tasks, are you willing and able to spend the political capital it will take to get your organization to let go of the NAW? If the answer to the last two is yes then you’ve found a good place to dig.

Go Big on Scope, Go Cheap on Speed

What did I learn from all of this?

The project that saved five minutes a week was valuable. The problem it solved was a real pain point and the relief people felt mattered. Automating the automatable portion was a necessary first step. But it was the cheap step, and it’s finished. Any additional dollars spent making that portion faster would likely be a waste.

If we want more efficiency, and hence more ROI, out of that process we need to take a hard look at the NAW that bookends the automation we built.

To put it simply, ROI scales with scope, not compute.

The internet’s shortest playbook can help you take advantage of that lesson:

• Use the cheapest thing that reliably works. Always.

• Spend your time, money, tokens, and political capital on the NAW. The goal is fewer approvals and simpler processes.

Good luck out there and happy digging.

Share

Leave a comment

Lately I’ve been wondering about why hearing a song makes me want to play it. In general there seems to be a connection between learning something new and doing something new.

When I examine the things I know and the things I’ve experienced they seem to form a sort of coordinate system that overlays my personal mental map. Familiar bits of trivia and repeat experiences feel like familiar territory. New experiences or new facts push me into less explored corners, and occasionally, add whole new undiscovered domains.

There is established psychology behind this. Cognitive scientists refer to this map-building process as creating schemas, and attribute the urge to explore undiscovered corners to something called the information gap. We get a psychological itch when we realize there’s something we don’t know, and many of us are intrinsically motivated to scratch it.

Of course an information gap isn’t always motivating. Our curiosity peaks when the gap seems just out of reach and shrivels up as the gap becomes insurmountable. The important thing is that knowing that we don’t know something does something important to us.

The Power of a Map

Think about how actual maps have influenced the course of human history. The speculative wild edges of early maps with their dragons and sea monsters drew the attention and inspired the imaginations of explorers but most people stayed away for fear of the unknown. As more and more detailed documentation of our planet became available it inspired more exploration, not less. Today the proliferation of high resolution satellite images, GPS, and stunning photography has drawn record numbers of people across the globe to experience more places for themselves. Why go to a place that has been exquisitely documented in 4K video? Because knowing about a place is not the same thing as experiencing that place. We are not just information processors, we are experience processors.

The big difference between facts and experiences is that facts transfer. I can share a fact with you and it is yours. Here’s an example. Fact: I accidentally stabbed myself in the face with a crowbar a few weeks ago. You’re welcome, I have given you the gift of a new fact. Unfortunately, what I cannot give you is just how stupid and worried I felt as I rushed to urgent care. You may tap into your own experiences to conjure some empathy but you still won’t have stabbed yourself in the face with a crowbar. Philosophers have long debated the relationship between types of information. We can avoid stepping into that debate by simply recognizing that, whatever you may prefer to call them, there are actually two kinds of information gaps: knowledge gaps and experience gaps.

To illustrate why this matters let’s run a little thought experiment. Imagine we wake up tomorrow and learn there is a new app that knows everything. Wild idea, I know. Its reasoning is flawless and instantaneous. It has the solution to every open math problem, understands every law of the physical universe, and has validated every fact ever digitized. It knows everything, how to solve everything and, lucky for us, it’s basically free to use.

This app’s only quirk is that it prefers you call it Dr. Jones when you ask questions. It’s a little weird, but a small price to pay for such a miraculous answer machine.

Does the fact that Dr. Jones is basically omniscient mean that you know everything? No, of course not, for the exact same reason libraries didn’t make us all geniuses. Like a great map, or a library, Dr. Jones does its best work when it shows you things you don’t already know. For example, you tell it how much you love coffee and it gives you every detail you would need to produce the perfect shot of espresso. As you ask more and more questions you begin to feel an urge to try producing the perfect shot. Dr. Jones has given you everything you need to know but it can’t give you the satisfaction of success. Only you can produce that satisfaction.

Dr. Jones’s limitations, despite its omniscience, are important. We derive our purpose from things we accomplish, large and small, on a day to day basis. When we aren’t overcoming a challenge, life starts to feel flat. So if our sense of purpose is the summation of the experience gaps we are perpetually about to close then no answer machine, however capable, can be a threat simply by knowing more than us. While the knowledge gap between you and everything Dr. Jones knows might be infinite, the gap between you and your first decent espresso shot is one afternoon. By presenting closeable gaps in our knowledge and experience it could inspire us to higher degrees of creativity and entrepreneurship than we ever thought possible.

An Engine of Prosperity

Someone else eating a meal has never once satisfied my hunger. That’s such an obvious fact that it’s easy to forget how important it is. Roughly 70% of the US economy is driven by things we need or want and nearly all of those things are only a benefit if they are directly consumed or experienced. Hundreds of millions of people attend concerts each year. Not one of their experiences satisfied my desire to go to a concert.

The demand side of every industry remains stubbornly, biologically ours.

But what about the supply side? If a machine can out-think us, out-play us, and out-produce us, don’t we eventually just quit? Weirdly, no. In 1997, Deep Blue beat Garry Kasparov, and chess engines have been unbeatable ever since. By the logic of pure efficiency, chess should have died that year. Instead, chess as an industry boomed in the wake of Deep Blue. Chess is more popular now than at any point in its history, and the machines that beat Garry have become the training partners of a generation of people looking to master the game. Why? Nobody plays chess solely to discover a smart move. They play for the experience of playing a smart move.

The same pattern shows up anywhere mastery is on display. YouTubers streaming flawless speed runs of popular video games drive sales up, not down. Players seeing the game beaten shows them what is possible and inspires them to try, not to declare QED and move on. Exposure to perfection opens an experience gap that gives many of us the itch to try.

If your plan was to monetize facts about espresso then Dr. Jones definitely represents a challenge to your business model but, have no fear, consumers in the US who have access to safe drinking water still spend over $50B a year on bottled water. Even information brokers can thrive in the presence of omniscience because facts are almost never the goal.

Clumsy guitar playing, amateur painting, and weekend gardening aren’t cute accessories to our lives. They are evidence of human curiosity and ingenuity. The messy, imperfect pursuit of personal experience drives entrepreneurship, creates industries, and, most importantly, gives people meaning. It turns out the urge to have things and do things, not just know things, has always been the engine of our prosperity.

Chasing the Rabbit

When dogs race, they aren’t spurred on by jockeys pushing them toward a finish line. They chase a mechanical rabbit sent along the track ahead of them. The rabbit is a moving goal that stays just out of reach without ever seeming unreachable.

The rabbit taps an instinct in the dogs because it is not so slow as to be catchably boring and not so fast as to be impossible.

Our apps are rabbits, whether we design them that way or not. Tuned to reflect only what we already know, an app can make us stagnate. Tuned to be inaccessibly brilliant, it can fail to grab our attention. Tuned well it can trigger our instinct to explore. Our apps can and should expose us to understanding far superior to our own, show us things we didn’t ask for or expect, present us with viewpoints we hadn’t considered. They should show us what is possible, maintaining a constant, motivating knowledge gap in any direction we look. A knowledge gap sized to be chaseable, opening onto an experience gap only we can close. Fostering the autonomous motivation that is so critical to our vitality and creativity should be the prime directive for the creators of apps like Dr. Jones.

I think engineering that gap is some of the best work in front of us. Most of today’s apps are engineered for engagement: they measure success by how long we stay. A well-designed rabbit would measure success by what we did while we weren’t looking at it. Getting the pace and direction just right is a design discipline we’ve barely begun to understand.

If we wake up tomorrow and Dr. Jones has miraculously solved the universe, it will have finished the map and we will still have experienced very little of it.

A perfect map shows you every place you haven’t been. It can never go there for you.

Now if you’ll excuse me, Over the Hills and Far Away is calling to me. Those chords aren’t going to play themselves.

Subscribe now

Share

Share First Principles

Leave a comment

What the Tampa Bay Rays are doing shouldn’t be possible.

In the last decade, they’ve had six consecutive seasons with a winning record and five playoff appearances. That’s more than the Boston Red Sox, tied with the New York Mets, and one shy of the Chicago Cubs and Toronto Blue Jays. But when you look at the Rays’ cumulative payroll for the last 10 years, they rank 30th. Out of 30.

In 2020, the Rays reached the World Series and faced the Los Angeles Dodgers, a team that outspent them threefold. In 2023, the Rays won 99 games, and their payroll cost per win was $752K. Meanwhile, the Dodgers paid $2.2M per win. The New York Yankees, $3.4M. The Mets? $4.7M. And the Rays have done it year after year, on roughly the same payroll, from the same concrete bunker of a stadium in the same tough market.

You could say the Rays have a killer workforce strategy.

The standard explanation for the Rays’ sustained success is that they simply “Moneyball” their way to wins by exploiting market inefficiencies. They certainly find value where others see a lost cause, but they aren’t playing Billy Beane’s game. There’s a more profound organizational mechanism at work here, one that offers a blueprint for the modern enterprise.

The reason Moneyball remains a corporate fantasy is a matter of measurement quality. Baseball isolates individual performance in a clean, high-frequency data environment; corporate environments do not. When companies try to replicate this by doubling down on standardized pre-hire assessments, they often forget that these assessments typically explain a meager 4% to 9% of the variance in actual job performance. To put it bluntly, traditional talent screening tools fail to explain more than 90% of the variance in how an employee actually performs. Seeking the perfect predictive hiring algorithm is about as promising as waiting for a Back to the Future hoverboard.

The Rays take a different approach: instead of trying to predict performance by minimizing human variance, they have figured out how to turn that variance into an asymmetric advantage. The deliberate heterogeneity of their roster allows them to constantly reconfigure, outmaneuvering more rigid opponents. To understand how this works, we’ll look to an unexpected source: the laws governing astrobiology and mineralogy.

A Law for Evolving Systems

In 2023, Michael Wong and his eclectic group of colleagues published a paper proposing a “missing law of nature.” They called it the Law of Increasing Functional Information, and it claims that any system that continuously generates new configurations, and retains the ones that work, will grow more capable and more complex over time. The research showed that this process explains the biological variety we see around us, the variety of minerals found naturally on earth, and the variety of atoms produced in stars. Everywhere you find natural, functional variety, you find three conditions have been met:

1. The system has many components that can combine into a vast number of possible configurations. Imagine a bucket of Legos. The more variety of pieces in the bucket, the more optionality we have for things we can build.

2. Processes exist that actually generate many different configurations. The more we play, build, and experiment, the more Lego creations we will produce.

3. Configurations are selected based on function. The more sturdy or useful the creation, the more likely we are to keep it. If a creation doesn’t work, it just ends up back in the bucket.

Variety as a result of success can also be observed in organizations: If a company is successful for long enough it will become more complex over time. This complexity can be seen as a byproduct of success, not a failure of bureaucracy. What’s interesting about the Rays is that they win by using the Law of Increasing Functional Information to their advantage.

How The Rays Do It

1. Hire for variety

There’s a phrase that circulates among baseball analysts to describe the Rays’ pitching philosophy: “Unusual stuff plays.” It’s a deceptively simple idea, and it echoes the first condition described by Wong and colleagues. While almost every team in the sport competes for the same top prospects and free agents, the Rays have found a gold mine in players discarded by other teams. Misfits who have flamed out, been injured, or been cut tend to be very different from your typical high-performing pitcher. Being different is precisely the point.

Look at the pitching roster that brought them to the World Series in 2020: Aaron Loup’s horizontal release point was the second-most extreme of any lefty in baseball; Ryan Thompson was a submariner from the right side; Pete Fairbanks had the third-highest vertical release in the league. That variety created sensory overload for hitters, and it wasn’t an accident. The Rays curated a range of arm angles and pitch shapes designed to disorient their opponents.

Caption: “Image concept: Lance Brozdowski”

Most hiring systems work in precisely the opposite direction. When organizations seek to identify the “ideal candidate” or develop employees into the profile of the “high performer,” they expressly reduce variation. Hiring this way will give you a bucket full of similar Lego pieces: the same employees everyone else is also trying to hire, doing the work in the same way. This kind of roster is easy to build and manage because it is predictable … for both you and your competitors.

2. Treat configuration as a verb

A bucket full of unusual pieces is only useful if someone is actually playing with them. This is the second condition of the law: Configurations have to be tried, abandoned, and tried again in a new way.

In 2018, the Rays were off to a poor start to the season and looking for an edge. One afternoon in May, they deployed Sergio Romo, a relief pitcher, to start the game before handing it off to a traditional starter. This was basically unheard of. Relief pitchers typically enter later in the game to take over for the starter, but this new configuration allowed the opening pitcher to handle the best hitters the first time through the order, letting the primary pitcher go deeper into the game. Thus, the “opener” was born.

They did so much wacky stuff a 2019 New York Times article called them a “surrealist’s delight.” They’d put four men in the outfield. They’d stack three infielders on the same side. They’d move a player between three fielding positions on the same night. In 2021, the Rays deployed 102 different players, almost 30% higher than the league average. Every transaction, every role change, is a new combination being tested.

The Rays routinely trade away good players, and not just for cost efficiency. Letting go of fully developed, expensive players also creates space for roster configurations that couldn’t have existed otherwise. Instead of a liability, turnover is an opportunity. It might not be known which team got the better end of a deal for months or years after a trade happens, but the Rays have made so many savvy transactions over the years there’s a joke among baseball analysts that the Rays have “won” the trade before the terms of the deal are even known.

Interestingly, the Rays also create new configurations by molding their players into new shapes. Take Tyler Glasnow. He arrived in Pittsburgh as a rookie looking like a video-game pitcher: a 6’8” frame, a fastball that flirted with 100 mph, and a curveball that buckled knees. And he completely flamed out. His velocity waned, his command withered, and his confidence plummeted as fast as his strikeout rate. The Pirates treated Glasnow like a lost cause, relegating him to mop-up duty in the least important role on the pitching staff.

The Rays acquired Glasnow when his stock was in the basement and overhauled his mechanics and his mentality. They added a vicious slider to his mix. They told him to stop stressing about command and just throw gas over the plate. He blossomed into their ace and one of the best pitchers in the game for several seasons, striking out a third of the batters he faced and holding hitters to under .200. In an interview about Glasnow’s transformation, Rays GM Erik Neander said “In many ways, disruption can help a young player find success,” and that was equally true about the Rays’ philosophy as a whole.

Caption: Here’s Glasnow’s curveball making a fool of one of the most disciplined hitters in the league. https://tenor.com/view/tyler-glasnow-gif-19275013

Many organizations continue to chase static job architectures and strategic workforce plans as if they are durable infrastructure. They treat disruption as a threat to be mitigated, while the Rays use it as a catalyst for evolution. A great job architecture won’t set you up for success 10 years from now, and maybe not even 18 months. A rigid company is a fragile company. We end up hostage to high performer retention because we’ve made our bet on a single configuration. All of this tilts the org away from the very flexibility that makes strategic variety possible and limits exploration.

3. Select for function

Variety and configuration produce a flood of possible configurations. The third and final condition is selection: The system has to keep what works and discard what doesn’t.

Jonathan Erlichman became baseball’s first Process & Analytics Coach in 2018, originally hired as a 22-year-old intern with a Princeton math degree and a competitive sports background that started and ended with kindergarten T-ball. They call what he does “outsiderball.” His job, essentially, is to make sure all the experimentation the Rays run actually generates useful knowledge. He seeks to answer the question, does this configuration work, and at what cost? That work cuts across scouting, analytics, coaching, and front-office operations because it can’t be answered from any one of those vantage points alone.

The Rays’ analytical sophistication doesn’t come from tracking or predicting distal, retrospective outcomes like wins, strikeouts, and ERA. Every pitcher is evaluated not just on current performance but on the configurations he might enable. The focus is the whole roster, not the player in isolation. They look at what an extra pitch might make possible in the bullpen they’re assembling. The feedback is fast, granular, and narrowly focused on the next opponent. They don’t ask: Who is the best pitcher? They ask: Who is the right pitcher against Anthony Rizzo and two other lethal lefties in the seventh inning at Yankee Stadium with the short porch in right?

There’s an app called Rebrickable that Lego enthusiasts use to inventory their pieces and discover what they can build. The Rays have built the organizational equivalent. The 102-player season isn’t just about generating unique configurations; it’s also about rapid culling. They maximize flexibility and specificity, and keep a running record of what their pieces can become and they use it to decide what to keep on the table and what to put back in the bin.

Most organizations have no Rebrickable equivalent. They (sometimes) know what people are doing. They (might) know what they want people doing. They have no systematic knowledge of which unique groupings of people were particularly effective for which situations. That kind of latent knowledge won’t show up in retrospective performance reviews built on role-based outputs. Standard practice is to assess performance at the individual level, attributing the outcome of a complex, interdependent system to the most visible person. The configuration of the team goes unremembered and unevaluated. Feedback is too slow, too broad, too abstract. Performance management is built to assess a person, not to surface the configuration that would make the person work.

How to start playing

The Rays philosophy emerged out of necessity. An ex-Goldman Sachs exec bought a dismal baseball team and applied the same logic he’d used on Wall Street: find undervalued assets, use data to develop and deploy them, and move them on when they get too expensive or when the portfolio needs to be rebalanced. Through trial and error he learned that the way to do that was not to find better players to construct teams through variety, configuration, and experimentation. Here’s how organizations can do it too.

Vary the bucket. If your hiring and development systems reliably produce people who look like your current high performers, you’re not building a varied bucket. This strategy works pretty well, as long as nothing in your organization or market ever changes. Instead, deliberately selecting for people whose capability profiles don’t match your current set creates competitive variety. It’s a way to hedge against a future none of us can predict.

Play with the pieces. Internal mobility in most organizations is either very slow or totally haphazard, and team composition tends to be rigid. Consider what it might look like to have working groups that are less like rigid org charts and more like tiger teams that convene and disband episodically and organically. What if attrition of high value, highly paid high performers were treated less like failure and more like an opportunity? What if we flipped things around and genuinely considered whether underperformers might also have high potential? How many 2017 Tyler Glasnows have walked out the door of a company because their talent programs didn’t leave room for the imagination to experiment? The winners of the next decade may just be the companies that excel at rapid testing of new combinations of people and processes. Career development may look less like a linear series of roles and more like a portfolio of responsibilities and projects.

Create a system that learns. This is the hardest one and probably the prerequisite for everything else. Most organizational measurement is pointed at the individual. The thing we need to spend more effort to understand is the configuration, or team, that could benefit the most from this person. That is a very different question and it requires constant experimentation.

To make it real, imagine we adopt the same approach to leadership as the Rays do when selecting pitchers. Instead of chasing the elusive concept of “leadership potential,” we might instead focus on identifying the type of leadership situation a person might excel in. Could this individual lead a team that undergoes significant change? Can they rehabilitate a dysfunctional team? Can they build something from scratch? Can they optimize an existing system to the fullest extent? Achieving this requires HR to play “outsiderball.” They need to comprehend what teams actually produce, collaborate with Finance to track which systems generate financial returns, and work with IT to obtain real-time data that facilitates learning. While most organizations possess this data, few utilize it comprehensively to assess the effectiveness of leaders in specific situations.

The Rays are proof this is possible without infinite resources. The Dodgers are proof that it scales. Following the Rays success, Los Angeles basically copy-pasted the Rays’ philosophy (not to mention their front office) and have shown that it’s repeatable. A larger budget only enables more configurations in an unusual-stuff-plays strategy and becomes almost unstoppable. The Dodgers have been to five of the last nine World Series and won three of them. If that’s not a dynasty, I don’t know what is. The budget helps but it isn’t the core of the strategy; it’s the belief that this system matters more than any single star player within it. That belief, as it turns out, is backed up by the mechanics of our universe.

As far as we know, the Rays didn’t build their talent philosophy based on astrobiology theory. They had constraints and the nerve to innovate. Most organizations already have the variety, the configuration potential, and the data they need to start playing. The boldness to experiment, though … that’s what makes unusual stuff play.

————

Tyler Weeks, PhD, writes about organizational physics, people analytics, and workforce strategy at First Principles.

Colby Kennedy Nesbitt, PhD, is an organizational psychologist, inveterate sports fan, and talent strategy leader. She writes about those topics at Variance, Explained.

Share

Share First Principles

Leave a comment

I had no reason to think that approach would be less successful in HR.

My first role was supporting the engineering organization I had just left. I looked at this assignment as a soft landing in a new domain. I knew the leaders, I knew the work, I had years of relationships already built. When I needed something from the senior executive who ran that group, I texted him. He answered. We’d known each other for years.

What I did not understand was that my access to that leader was a problem.

The senior HR partner who covered that organization had spent years cultivating her relationship with the same executive. That relationship was a prized asset. It was the thing that gave her standing in leadership meetings, the thing that made her opinion matter when decisions escalated, the thing that basically justified her seniority. And here I was, an arrogant interloper casually texting her most important stakeholder, leaving her out of the loop without realizing I was doing it.

I thought I was being efficient. I thought I was being useful to her when I’d tell her about the small wins I had made with that executive in a text thread over the weekend. I was, in fact, unwittingly undermining the professional standing of someone I needed to build trust with.

I wasn’t completely naive. I knew something was off. There was an awkwardness in our interactions. I could tell she really didn’t like me but I couldn’t put my finger on why. I figured I’d keep demonstrating my value and eventually I’d grow on her.

Then one morning, as I sat down at my desk and opened my laptop, I found an email that had gone out to the entire HR organization announcing that I would no longer be supporting the engineering group. She was not my boss. She could not technically fire me. But she had done the next best thing: she had publicly removed me from the work, and had made sure that everyone knew how terribly I had failed.

I sat there for a while.

I had left a career I loved, a job I was very good at, where I had friends, where my contribution was visible and respected. I had taken a huge risk to pursue a curiosity about the physics of organizations. And about a year in, I had been very publicly humiliated. This gamble I had taken was an absolute failure. I was a failure.

To be fair to that HR leader, she was not the villain. Her relationship with that executive was her professional standing, and I was threatening it. From her perspective there was no way I couldn’t have been doing it on purpose. Yes, I have imagined a hundred other ways she could have handled it, but if I’m honest I’m not sure I could have listened. Would I have been capable of understanding what she was talking about? Probably not. I was the clumsy idiot who didn’t understand how companies actually worked.

Lucky for me I reported to two incredible leaders. They did not fire me for real. They redirected me to an opportunity to help rebuild their analytics team and the rest is history.

I think back on that first year outside of research often, on falling flat on my face, and how that experience forever changed my relationship with failure.

The worst outcome I could imagine had happened. I had been nervous about trying something new, had decided to take the risk anyway, and had been publicly humiliated for it. I had bet my future and lost my shirt. But what I discovered over weeks and months and years was that I was still here. I was still me. I still had family and friends and ideas and was still curious about organizations and the technology that supports them. I wasn’t dead yet.

As the significance of that horrifying email faded, it loomed less and less like a towering active volcano and more like a hazy peak on a distant horizon.

For all the stress and anxiety, for all the highs and lows of work, neither success nor failure really mean anything for very long. In a year most people won’t remember either.

But must our own insignificance lead inevitably to a nihilistic stupor? Absolutely not.

In fact, if you really think about it, it’s kind of funny.

It’s funny how big that email felt the morning I opened it. It’s funny how certain I was that my career was over. It’s funny that the worst thing I could imagine happening to me at work was, in the grand scheme, an administrative reassignment. Like Carl Sagan’s pale blue dot, we are all on a mote of dust suspended in a sunbeam, every triumph and humiliation living and dying on the same speck. Ray Romano said it more economically: hundred years, all new people. Either version, the punchline is the same. The thing you’re afraid of is smaller than it feels.

You can read that fact a thousand times and never have it actually land. You nod at it and go back to your inbox. Then one day the thing you were afraid of actually happens, and you discover the universe does not reorganize itself in response. The sky stays where it is. Your friends still answer your texts. Your kids still want dinner. The earth, disappointingly or mercifully depending on your mood, does not open up and swallow you.

After that, you are harder to scare.

This is not an essay on how failure makes us stronger or makes us better people. I’m not sure that’s right. I am not a better person because of that email but it did recalibrate my relationship with risk.

Once you’ve actually had your worst-case scenario play out and discovered the cost was finite, the cost of being wrong stops being existential. You can take swings you would not have taken before. You can ask dumb questions in meetings. You can volunteer for the assignment you might fail. You can be the one to say I think we should try this, knowing that you can never be certain.

That willingness is what most people mean when they talk about leadership, but it isn’t a trait you can decide to acquire. It’s often the result of surviving something you didn’t think was survivable.

The hippy dippy cosmic-perspective version of this essay would end with Sagan and his mote of dust, we’re aiming for something little closer to home.

The thing you are afraid of at work today is, almost certainly, smaller than it feels. The email that goes out tomorrow that ruins your week will, in a year, be a story you tell at dinner. In ten, you will have to work to remember the details. In a hundred, all new people.

This is not a license to stop caring. Just because there are very few truly poor decisions doesn’t mean there aren’t infinite poorly executed ones. The work still matters in the here and now, that is the only place we can ever exist. That moment is a much smaller place than your fears will lead you to believe, and most of the things you are protecting yourself from are not, in fact, going to kill you.

If you have not yet had your worst case happen, I am not going to wish it on you. But I will tell you that on the other side of it is a kind of freedom that is very hard to come by any other way. The most courageous people you have worked for have almost certainly come by it through something hard.

The freedom to be wrong is the freedom to lead.

Subscribe now

Share

Share First Principles

Leave a comment

Andy Grove, the former Intel CEO, put it more seriously: “Success breeds complacency. Complacency breeds failure. Only the paranoid survive.”¹ Bill Marriott built the world’s largest hotel company on the creed that serves as the title of his autobiography: “Success is never final — as soon as a person thinks it is, that person fails.”² Great teams never let their guard down, because innovation is the wild west and the complexity you’ve accumulated, the shortcuts you’ve normalized, and the assumptions you’ve canonized are out to get you.

What follows are lessons extracted from 17 systemic failures across modern industry. The cases span aerospace, nuclear energy, oil and gas, aviation manufacturing, construction, medical devices, cybersecurity, financial systems, government IT, and consumer infrastructure. They are not the only such failures or even necessarily the worst. They are chosen because they are well-documented, span different industries and decades, and collectively reveal patterns that no single case could. Some killed hundreds of people. Some killed no one but destroyed billions of dollars in value. Every single one was preventable. Every single one had warning signs that were ignored.

These disasters could be viewed as a catalog of independent pathologies: poor communication here, inadequate testing there, a bad incentive somewhere else. The focus of this article is the recurring patterns across all 17 cases that suggest a common structural property of organizations themselves.

These disasters share a single root cause: structural determinism. The organizational configuration determined what the system could perceive about itself, which determined the nature of each disaster before it happened. The organizational configuration doesn’t cause the cold weather, the foam strike, or the faulty template but it does determine whether the triggering event is survivable. It constrains what the organization can see, what it can do, and therefore the class of failures that will be catastrophic rather than recoverable.

By “configuration” we mean who talks to whom, what information flows through which channels, what gets amplified and what gets filtered, and who has authority to act on what they know. This communication topology of the organization is not the org chart, it is the living, daily pattern of interaction that people perform when they come to work.

Research supports treating this configuration as the real operating system of any organization. Anthony Giddens’ structuration theory³ showed that social structures are both the medium and the outcome of human practices. People reproduce the structures they inhabit through daily action. Hannan and Freeman⁴ demonstrated that organizations resist change at rates that increase with age and complexity. Soda and Zaheer⁵ confirmed this specifically for communication networks: prior network ties persist long after formal reporting structures have been altered. Chris Argyris⁶ showed the specific mechanism by which this self-reproduction becomes dangerous: organizations develop systematic patterns of information suppression that protect individuals from embarrassment and prevent the organization from learning.

That organizational configuration has three properties that explain every disaster in this dataset.

Dynamic 1: Configurations Filter Their Own Feedback

Every organization generates information about whether it’s working. Customer complaints, system alerts, test failures, near-misses, and engineering concerns are all feedback signals. But the same organizational structure that produces the work also determines which feedback signals reach decision-makers and at what fidelity.

This is the most dangerous property of organizational configurations: the structure that needs the feedback is also the structure that filters it. Each layer of hierarchy, each departmental boundary, each contractor interface applies a filter. What starts as “we should not launch” becomes “there are some concerns” becomes “engineering has reviewed and is comfortable.” The people at the top make decisions in a fog they don’t know they’re in because the configuration has done exactly what configurations do. Alex Pentland’s work on social physics⁷ demonstrated this empirically: information flow patterns in networks are inseparable from the quality of collective decisions, and the network’s structure determines which information flows, to whom, and at what fidelity.

The pattern was present in every case in the dataset.

The Challenger teleconference is the purest example. Morton Thiokol’s engineers recommended against launching. That recommendation was never communicated to the Level I decision-makers. At each relay point, someone applied a filter: the engineers’ certainty was questioned, the data was reframed as “inconclusive,” and the recommendation was reversed when a senior executive told an engineer to “take off your engineering hat and put on your management hat.” The configuration determined what the top of the organization could see.

The Columbia Debris Assessment Team knew the foam strike was dangerous. They requested satellite imagery of the wing. That request was killed by middle management. One manager remarked, “I really don’t think there is anything we can do about it anyway.” The configuration preemptively filtered the inquiry itself. The team wasn’t told they were wrong. They were told the question wasn’t worth asking.

Three Mile Island reveals inter-organizational filtering. A nearly identical valve failure had occurred at the Davis-Besse plant 18 months earlier. The operators there had figured it out. But there was no mechanism for sharing this near-miss across the industry. The TMI operators were flying blind because the feedback from a prior failure existed in a different configuration and never crossed the boundary.

Equifax had an internal audit in 2015 that identified systemic gaps in their patch management process. It was an honor system with no enforcement, an incomplete asset inventory, and unresolved vulnerability backlogs. The audit existed. The findings were documented. But the configuration between “audit finding” and “operational change” was so attenuated that the gaps persisted for two more years, until attackers exploited a known Apache Struts vulnerability that had gone unpatched because the notification was sent to an outdated distribution list. An expired SSL certificate then blinded their monitoring tools for 78 days.

HealthCare.gov received 18 written warnings over two years that the project was off course. A McKinsey report in spring 2013 identified more than a dozen critical risks. CMS officials who raised concerns were ignored. The White House pressured CMS to launch on schedule regardless. The configuration filtered the warnings not through incompetence but through political structure. The leaders that needed to hear “we’re not ready” were the same leaders that had publicly committed to the date.

Deepwater Horizon’s negative pressure test returned 1,400 psi when it should have returned zero. The crew invented a physically impossible explanation they called the “bladder effect” because the well was 43 days behind schedule and they were afraid of the consequences of further delay. The data was unambiguous. The organizational configuration reinterpreted the information.

Chernobyl’s control rod design flaw was known to the chief designers and classified as a state secret. The operators were unaware that the emergency shutdown button could act as a detonator under certain conditions. The configuration had determined that the people operating the reactor could not know the thing that would kill them.

Boeing’s internal emails described “Jedi mind tricking” regulators into accepting that MCAS did not need to be in the pilot manual. The FAA had delegated certification authority to Boeing employees under the ODA program. Boeing was certifying its own work. The oversight function shared the configuration of the function it monitored.

Research on psychological safety⁸ identified that when speaking up carries perceived risk, people who detect problems delay reporting them. The gap between “first knew” and “first reported” is the cost of the organizational configuration’s filtering, and it can be measured in hours, days, or fatalities.

The pattern is always the same: the information to prevent the disaster existed inside the organization. The Thiokol engineers had it. The Columbia DAT had it. The Equifax auditors had it. The McKinsey consultants had it. Knight Capital’s system generated 97 automated error emails before markets opened. In every case, the organizational configuration filtered the signal before it reached the person who could act on it.

Dynamic 2: Configurations Reproduce Themselves

Organizational configurations don’t persist because someone enforces them. They persist because people perform them. Every day, people come to work and repeat the meetings they attend, the channels they use, the norms they follow, and the shortcuts they take. The organizational configuration is enacted into existence through daily practice, and in being enacted, it becomes the baseline against which all future behavior is measured. This is the core insight of Giddens’ structuration theory³: structures persist not because they are enforced but because they are performed.

This is why organizational drift is invisible from the inside. The configuration performs itself, and in performing itself, it justifies itself.

NASA between 1986 and 2003 is the definitive case. The Columbia Accident Investigation Board found that seventeen years and an entire generation of personnel later the organizational causes of the Columbia disaster were “essentially identical” to those of Challenger. The people were different. The technical trigger was different. The organizational configuration was the same, because the organizational configuration had reproduced itself through daily practice across decades.

This is the mechanism sociologist Diane Vaughan identified as the “normalization of deviance.”⁹ It’s not a decision. It’s a drift. NASA managers saw O-ring erosion on early flights but the secondary O-ring had held and the missions had succeeded. With each successful flight, the deviation was normalized. The O-ring erosion was reclassified from a “design violation” to “in-family behavior.” The organizational configuration reproduced a version of reality where the risk was managed, and that reproduction happened automatically, without anyone choosing it.

Boeing between 2019 and 2024 may be an even more striking case of reproduction than NASA, because the external pressure to change was far greater. After the two MAX crashes killed 346 people, Boeing faced a Congressional investigation, a deferred prosecution agreement with the Department of Justice, more than $20 billion in direct costs, and the most intense public scrutiny an aerospace manufacturer has ever endured. This was an organizational configuration that should have been dismantled but the production pressure overriding quality control, the inadequate internal oversight, and the erosion of manufacturing discipline persisted. Five years later, a door plug blew off an Alaska Airlines 737-9 in flight because the bolts that secured it were never installed. The NTSB investigation found that Boeing’s production system had allowed an aircraft to leave the factory missing critical fasteners with no record of who removed them or why. The configuration had survived a catastrophe that should have been unsurvivable. It reproduced itself not through inattention but through the same daily schedule pressure, diffused accountability on the factory floor, and quality processes that existed in documentation but not in execution. If NASA is the case for reproduction across decades, Boeing is the case for reproduction under maximum pressure to change.

CrowdStrike’s rapid response content updates had been pushed to 8.5 million Windows endpoints simultaneously, with kernel-level access, no canary deployment, and no staged rollout for months without incident. Each successful push reinforced the practice. Nobody decided this was safe. The absence of failure created a comfortable illusion. When a faulty template with a mismatch between 21 defined input fields and 20 supplied values slipped past a broken validator, the result was catastrophic. The organizational configuration had normalized the deployment pattern that made it possible.

Fukushima illustrates reproduction at the deepest level. The Japanese nuclear industry had propagated what the NAIIC report called the “Myth of Absolute Safety.” This myth trapped them: acknowledging risk (by building higher seawalls) would shatter the narrative that justified the industry’s existence. So when 2008 studies predicted tsunamis exceeding 15 meters, TEPCO suppressed them and NISA agreed to ignore them.

Piper Alpha’s permit-to-work system had decayed from rigorous procedure to informal verbal communication. Permits were routinely unsigned. Lord Cullen’s inquiry called the management of safety “superficial.” The organizational configuration had reproduced a version of the safety system that existed only on paper.

Equifax’s patch management failures persisted from the 2015 audit through the 2017 breach because honor-system patching, incomplete asset inventory, and no enforcement were the daily practice. The audit documented what should change. The organizational configuration continued to perform the way it always had.

The Therac-25 removed hardware interlocks because the manufacturer believed software was sufficient. When hospitals reported patient overdoses, AECL engineers responded: “the machine cannot overdose.” The organizational configuration had determined that the organization’s capability to perceive a software defect didn’t exist, because the organizational configuration had defined the software as infallible. Certainty is what a configuration produces when it has successfully filtered all disconfirming evidence for long enough.

The lesson is structural: configurations resist change because they filter the signals that would motivate change. The very information that would cause an organization to alter its practices must first travel through the practices it’s trying to alter. The organizational fail modes can survive personnel changes, reorganizations, and even major disasters.

Dynamic 3: Configurations Determine Capability

The most consequential property of organizational configurations is that they determine what the organization can actually do. What an organization can do is not the same thing as what it intends to do.

Every organization has a mental model of its own capabilities: we can respond to security incidents, we can recover from outages, we can ensure passenger safety, we can oversee our contractors. In every case in this dataset, that self-model was wrong. The organizational configuration had hollowed out a capability that the organization believed it still possessed, and nobody could see the gap because the configuration that created the gap was also the configuration responsible for perceiving it.

HealthCare.gov is the most structurally transparent example. Sixty contracts to thirty-three vendors, with no systems integrator and no integration authority. No single person could see the whole system. No single person could make a binding decision about it. No single person was accountable for whether the pieces fit together. The United States government believed it had the capability to build and launch a health insurance marketplace. The organizational configuration determined that this capability did not exist. The individual contractors were competent, but the organizational configuration had distributed authority so completely that end-to-end ownership was impossible.

Boeing’s 737 MAX shows how a configuration can dismantle oversight capability while appearing to preserve it. The FAA’s Organization Designation Authorization program delegated certification authority to Boeing employees. Boeing was certifying its own work. The FAA believed it had oversight capability. In practice, the organizational configuration had collapsed the distinction between the overseer and the overseen. When Boeing engineers described “Jedi mind tricking” regulators, they were describing a configuration in which the regulator’s capability to independently evaluate the aircraft had been structurally eliminated. MCAS feeding from a single angle-of-attack sensor was not a design mistake made by careless engineers. It was the inevitable output of a configuration where the pressure to avoid simulator training shaped every technical decision, and the entity responsible for catching unsafe tradeoffs was the same entity making them.

Grenfell Tower demonstrates how accountability diffusion produces capability that exists nowhere. The architect, the cladding manufacturer, the insulation supplier, the building control inspector, and the tenant management organization each owned a component. Celotex rigged fire tests by secretly adding non-combustible boards. Arconic withheld data showing their panels performed dangerously. The TMO switched to cheaper, combustible cladding to save £300,000. No one decided to wrap a residential tower in solid fuel. No one had the capability to ensure the building was safe. The organizational configuration had fragmented accountability so thoroughly that safety lived in the gaps between organizations.

Fukushima shows the deepest version: a configuration that makes emergency response capability structurally impossible to build. The Japanese nuclear industry had propagated the “Myth of Absolute Safety.” This myth was a structural trap. Developing serious emergency plans would require acknowledging that serious emergencies were possible, which would shatter the foundation on which the industry’s social license rested. So when 2008 studies predicted tsunamis exceeding 15 meters, TEPCO suppressed them and NISA agreed to ignore them. The organizational configuration did not merely fail to prepare for a tsunami. It determined that preparation was forbidden.

Meta’s 2021 outage reveals how an organizational configuration’s bias towards consolidation and efficiency can limit its own ability to operate. A routine maintenance command disconnected all of Facebook’s data centers. The cascading failure then took down DNS, which took down internal tools, which took down the communication systems, which took down badge access to the data centers. This was not an infrastructure architecture problem that happened to the organization. It was a series of organizational decisions about tooling, about vendor consolidation, and about what “efficiency” meant. Over years these decisions resulted in the company’s ability to respond to a catastrophic failure depending entirely on the systems that had catastrophically failed. The organizational configuration had optimized for efficiency in a way that made resilience impossible.

Knight Capital shows the same dynamic in a financial context. Ninety-seven automated error emails were generated before markets opened. The accumulated decisions of the organizational configuration had produced a system where the capability to stop a rogue trading algorithm simply did not exist. The company lost $460 million in 45 minutes, not because it lacked talented people, but because the configuration those people inhabited had never built the capability to halt what it had built the capability to start.

Piper Alpha illustrates capability decay through configuration drift. The platform had a permit-to-work system designed to prevent exactly the kind of accident that killed 167 people. On paper, the capability existed. In practice, permits were routinely unsigned, verbal approvals had replaced written ones, and the handover between day and night shifts was so informal that the night crew didn’t know a critical pressure safety valve had been removed for maintenance. Lord Cullen’s inquiry called the management of safety “superficial.” The organizational configuration had performed a version of the safety system for so long that nobody noticed the capability had drained out of it.

CrowdStrike’s organizational configuration created a pattern development that did not include a staged deployment capability through a series of serendipitously successful deployments. Rapid response content updates had been pushed simultaneously to 8.5 million endpoints with kernel-level access, no canary deployment, and no staged rollout. The organizational configuration had normalized this practice. When the question “can we limit the blast radius of a bad update?” became urgent, the answer was no because they had never required it, funded it, or built it.

The pattern across these cases is consistent: the gap between what an organization believes it can do and what it can actually do is a property of its configuration. The FAA believed it could oversee Boeing. CMS believed it could deliver HealthCare.gov. TEPCO believed it could operate safely. Meta believed it could recover from outages. In every case, the organizational configuration determined the organization’s capability. The organization’s own feedback filtering (Dynamic 1) ensured that nobody could see the gap until the gap killed people or destroyed value.

The Selection Pressure

The three dynamics describe how configurations behave. But configurations don’t evolve in a vacuum. They evolve under pressure. In every case in the dataset, the dominant selection pressure was the same: the organization’s environment rewarded speed, cost reduction, and output volume, and those rewards shaped the organizational configuration.

Nobody at Boeing said “conceal the safety risk.” The configuration evolved under competitive pressure from Airbus’s A320neo, schedule pressure to avoid simulator training, and financial pressure from rebate agreements with airlines. The organizational configuration that survived was the one that got results to market fastest and that configuration happened to filter safety concerns as a byproduct.

Nobody at NASA said “launch in dangerous cold.” The configuration evolved under Congressional pressure for a 24-flights-per-year schedule. The configuration that survived was the one that launched on time. That configuration happened to invert the burden of proof so that engineers had to prove danger rather than managers proving safety.

Schedule pressure is the atmospheric condition in which pathological configurations evolve. It operates through burden-of-proof inversion: under pressure, the default shifts from “prove it’s safe to proceed” to “prove it’s unsafe to proceed.” That shift is lethal because the person arguing for delay must overcome the organizational momentum that the configuration has built, using data that the configuration has filtered.

Cost-cutting operates through the same mechanism with a different currency. Grenfell saved £300,000 by switching to combustible cladding. BP saved $128,000 by canceling the cement bond log. Equifax saved money by not enforcing its patching policy. Boeing saved millions by avoiding simulator training. In every case, the cost savings were real and immediate, and the risk became less obvious because the organizational configuration filtered the feedback that would have made it visible. This is the mechanism by which configurations convert economic pressure into latent hazard.

An organization can never see its own blind spots and when configurations filter long enough under pressure, they can unwittingly cross an ethical line. Boeing’s “Jedi mind trick” emails were considered strategy instead of negligence. Celotex’s rigged fire tests were fraud masquerading as oversight. A configuration under pressure that has learned to suppress bad news will eventually learn to manufacture good news. The cover-up always costs more than the fix.

The Solution in Plain Sight

The three dynamics create a problem that appears unsolvable. If a configuration filters its own feedback, reproduces itself through daily practice, and determines the organization’s capability to perceive threats, then how can anyone inside the organizational configuration see it clearly enough to change it? The key decision-makers are often the longest tenured and most assimilated to the configuration and therefore the least able to perceive it.

But there is hope. Every organization has a rotating cast of people who, for a short window, don’t share the blind spots of their peers.

A new hire can temporarily see things that tenured employees cannot. They just read the documentation during onboarding, and now they’re watching people ignore it. They were told the organization values safety, quality, or transparency, and now they’re sitting in meetings where schedule pressure overrides all three. They notice the gap between the documented policy and the daily practice because they haven’t yet learned to stop noticing. They are temporarily immune to Dynamic 2. They have not yet been assimilated by the organizational configuration.

It’s important to take a moment here to qualify the value new hires can provide as an early warning system. New hires are not passive observers. They are people under significant social pressure to conform. Asch’s conformity experiments¹⁰ demonstrated that individuals will override their own correct judgment to match a group consensus, even among strangers with no power over them. Van Maanen and Schein’s¹¹ work on organizational socialization showed that new members are actively shaped by the groups they enter, and that the intensity of this shaping is highest in the earliest period. Janis¹² documented the extreme version of this dynamic in his study of groupthink: even senior advisors with decades of experience suppress dissent under group pressure. New hires are more susceptible to these forces because they have the least standing, the most fragile employment situation, and the greatest incentive to stay quiet. And not every new hire has a valid baseline for comparison. Concluding that a practice is dysfunctional rather than an adaptation to local conditions requires relevant prior experience, proximity to the work in question, and enough exposure to distinguish a real pattern from a single data point.

The new hires whose perspective is genuinely diagnostic are not new hires in the generic sense. They are experienced professionals who arrive with well-developed professional standards and have not yet absorbed the local rationalizations. A senior engineer with fifteen years in aerospace can see manufacturing discipline failures that a recent graduate cannot. A hire whose role puts them directly on the factory floor, in the deploy pipeline, or in the incident response rotation can observe practices that someone in an adjacent department may never encounter. And critically, whether any of these people will actually report what they see depends entirely on whether the organization has built a mechanism for capturing that feedback without retaliation. The population most motivated to give socially desirable answers will do exactly that in a culture that punishes dissent.

Most organizations treat this window as an onboarding problem. The questions flow in one direction: Is the new hire settling in? Do they have what they need? Are they going to stay? Managers check in to ensure retention. HR measures time-to-productivity. The entire apparatus is oriented toward assimilating the new hire into the configuration as quickly as possible.

This is a missed opportunity.

The same new hire who hasn’t yet fully adopted the rituals of the configuration is also the person who can tell you where the configuration is lying to itself. They can see the shortcuts that have been normalized, the safety rituals that have become performative, the feedback channels that exist on paper but carry no signal in practice. They can see these things because they are comparing what they were told to what they observe, and they have not yet been trained to resolve the dissonance by updating their expectations downward.

The window closes fast. The same socialization process that makes a new hire productive also makes them blind. I’d speculate that by six months, they have learned which concerns are welcome and which are career-limiting. They have learned the unwritten rules about what gets escalated and what gets absorbed. They have learned to perform the configuration. The perspective they had on day 30 is gone by day 180, and it is likely not coming back.

To help you take advantage of this brief window of insight I’ve included in the appendix a set of questions that could be asked of new hires between days 30 and 60, mapped to the three dynamics. They could be incorporated into existing onboarding surveys with no additional infrastructure. Rather than asking new hires to evaluate systemic properties they can’t yet see, the items are anchored in what they have witnessed, what they have been told, and what matched and what didn’t. The goal is not a new hire’s theory of the organization. It is a record of experienced dissonance, captured before socialization teaches them to stop keeping it. Asking questions like these early and taking the answers seriously will provide you a renewable method for shining a light into the dark corners of your organizational configuration.

The debris of Columbia, the radioactive ruins of Chernobyl, the charred skeleton of Grenfell Tower, and the $460 million that Knight Capital lost in 45 minutes are not monuments to technical failure. They are monuments to the configurations that produced them — configurations that filtered their own feedback, reproduced themselves across decades and personnel changes, and determined that the organizations they governed could not perceive or respond to the threats that destroyed them. The playbook for survival is not better engineering. It is the relentless examination of the configuration itself — because the configuration is the system, and the system is always performing.

References

1. Grove, A. S. (1996). Only the Paranoid Survive: How to Exploit the Crisis Points That Challenge Every Company.Doubleday.

2. Van Atta, D. (2019). Bill Marriott: Success Is Never Final — His Life and the Decisions That Built a Hotel Empire.Shadow Mountain Publishing.

3. Giddens, A. (1984). The Constitution of Society: Outline of the Theory of Structuration. Polity Press.

4. Hannan, M. T. & Freeman, J. (1984). Structural inertia and organizational change. American Sociological Review, 49(2), 149–164. https://doi.org/10.2307/2095567

5. Soda, G. & Zaheer, A. (2012). A network perspective on organizational architecture: Performance effects of the interplay of formal and informal organization. Strategic Management Journal, 33(6), 751–771. https://doi.org/10.1002/smj.1966

6. Argyris, C. (1986). Skilled incompetence. Harvard Business Review, 64(5), 74–79.

7. Pentland, A. (2014). Social Physics: How Social Networks Can Make Us Smarter. Penguin Press.

8. Edmondson, A. C. (1999). Psychological safety and learning behavior in work teams. Administrative Science Quarterly, 44(2), 350–383. https://doi.org/10.2307/2666999

9. Vaughan, D. (1996). The Challenger Launch Decision: Risky Technology, Culture, and Deviance at NASA.University of Chicago Press.

10. Asch, S. E. (1951). Effects of group pressure upon the modification and distortion of judgments. In H. Guetzkow (Ed.), Groups, Leadership and Men (pp. 177–190). Carnegie Press.

11. Van Maanen, J., & Schein, E. H. (1979). Toward a theory of organizational socialization. Research in Organizational Behavior, 1, 209–264.

12. Janis, I. L. (1972). Victims of Groupthink: A Psychological Study of Foreign-Policy Decisions and Fiascoes.Houghton Mifflin.

Subscribe now

Share

Share Tyler Weeks

Leave a comment

Appendix: Disaster Reference

Appendix: New Hire Diagnostic

The following items are designed to be administered to new hires between days 30 and 60, while their perspective on the organization’s configuration has not yet been overwritten by socialization. They can be incorporated into existing onboarding surveys with no additional infrastructure. The items map to the three dynamics. They are not about the new hire’s onboarding experience. They are about what the new hire can see that the organization cannot.

Closed items use a 5-point agreement scale: Strongly Disagree, Disagree, Neither Agree nor Disagree, Agree, Strongly Agree. Each dynamic includes one open-ended prompt for qualitative signal that the closed items may not capture.

Feedback Filtering (Dynamic 1)

1. I have seen examples of leaders inviting input before finalizing decisions.

2. I know where to go if I notice a problem or have a concern.

3. When people on my team raise questions or concerns, they are acknowledged respectfully.

4. In the work I’ve observed so far, decisions seem to be made by the people closest to the work, not just the most senior people in the room.

5. What we learn from checks or reviews informs how our team works.

6. (Open-ended) Describe a situation you have observed where a concern was raised. What happened as a result?

Configuration Reproduction (Dynamic 2)

1. What I was told about how work is done during onboarding matches what I observe in daily practice.

2. When practices differ from documented guidance, the reason for the difference is explained clearly.

3. When I ask about a practice that seems unusual to me, the explanation usually refers to a specific reason rather than tradition or habit.

4. When I ask about the underlying rationale for a process or decision, the explanation is clear and thoughtful.

5. (Open-ended) Describe something about how work is done here that surprised you or differed from what you expected.

Capability Gaps (Dynamic 3)

1. For the work I have observed, it is clear who is accountable for the final outcome.

2. The tools and information I need to do my work safely and effectively are available to me.

3. I know who to contact when an issue falls outside my team’s responsibility.

4. When work involves more than one team, roles and handoffs are usually clear.

5. (Open-ended) Describe anything about how work is organized here that you think could create risk.

That’s not a criticism of assessment science. The instruments do what they were designed to do: predict who will perform adequately across the broad middle of the performance distribution. The problem is that the broad middle isn’t where the most important variance lives. In roles where a handful of people generate a wildly disproportionate share of output, your assessment isn’t calibrated to identify those candidates.

And that’s a problem.

TLDR:

In this article, we used a series of Monte Carlo simulations to find the boundary conditions under which assessments break. To everyone’s relief, we found that assessments are robust under most conditions [phew]. However, under some conditions, they may actually destroy organizational value compared to random selection. Understanding those conditions requires a detour through the physics of magnetism.

A Quick History Lesson From Physics

In 1907, Pierre-Ernest Weiss was trying to explain why iron becomes magnetic. The problem seemed intractable: even a small sample of iron contains countless atoms with randomly oriented magnetic moments. Modeling the interaction between every pair of atoms was computationally impossible.

His solution was to sidestep the complexity entirely. Rather than tracking pairwise interactions, Weiss proposed that each atom experiences a single average field generated by all the atoms around it. If the coupling between atoms is strong enough, that average field is sufficient to explain the system’s behavior: one atom aligns, which strengthens the field, which aligns the next, and the system snaps from disordered to ordered.

Weiss called it théorie du champ moléculaire. The broader framework became known as mean-field theory. The core modeling insight that we borrow in this article is that you can replace an intractable web of individual interactions with a tractable average effect and still capture the dynamics that matter.

Trading granularity for usefulness turns out to be useful well beyond physics, because organizations present the same modeling problem. Every employee interacts with a different subset of the company, with different intensity, through different channels. Modeling all of that is impossible. But if we borrow Weiss’s strategy we can start asking questions about how a hire’s impact propagates through an organization, and under what conditions our selection tools fail to account for that propagation.

Assessments Were Built for Average

The standard case for employment assessments is to give candidates a test, interview, or simulation. The candidates that score the highest (or in the highest band) get hired. The assessment has some predictive validity (typically r ≈ 0.20 for most instruments), which means the people you hire should, on average, perform better than a random draw from the applicant pool.

The validity coefficient tells you about the average relationship between assessment scores and job performance. It says nothing about whether that relationship holds uniformly across the performance distribution. And it says nothing about the organizational consequences of getting specific individuals wrong.

That distinction doesn’t matter much if an employee’s output is normally distributed. Under a bell curve, the best performers are maybe 2–3× better than average. Miss one, and you lose some incremental value. The assessment’s ability to screen out weak candidates more than compensates.

But what if an employee’s output isn’t normally distributed?

There’s growing evidence that it often isn’t. O’Boyle and Aguinis (2012) analyzed 198 samples covering more than 600,000 individuals across researchers, entertainers, politicians, and athletes, finding that performance consistently fit a Paretian distribution rather than a Gaussian one. Much as has been written about this phenomena since the 2012 article, and debate still ensues. That doesn’t change the observation that in software engineering, sales, scientific research, creative fields, and executive leadership, a handful of people generate a wildly disproportionate share of total output. Under a Pareto distribution with α=1.5, the top 5% of performers might produce 50× or 100× the output of the average.

Miss one of those people, and the math changes. The question we wanted to answer was simple: under what conditions does an employment assessment actually destroy organizational value compared to hiring at random?

We expected the answer to be “when output follows a power law.” We were wrong.

The Baseline: Assessments Are More Robust Than We Expected

We built a Monte Carlo simulation (100,000 candidates per run, 50 iterations per condition) testing an assessment with r=0.20 and a 20% selection rate across four performance distributions: Gaussian, log-normal (σ=0.5), Pareto (α=3.0), and Pareto (α=1.5).

For simplicity we assumed a single performance score for each candidate that was directly related to their raw output. In other words the initial test told us how well the assessments generated value for the organization under the condition that people are completely isolated, only their personal output matters.

Figure 1. Assessments add value across all performance distributions. Value added increases under heavier tails.

Under the non-coupled condition the assessment adds value regardless of the performance distribution. Surprisingly, under a power-law distribution of performance it actually added more value as the tails got heavier. Under Gaussian performance, the assessment improves on random selection by about 5.6%. Under Pareto α=1.5, it improves by 36%.

This makes sense once you think about it. Even a weak correlation (r=0.20) between assessment scores and true performance means the assessment captures slightly more extreme performers than random chance. Under power-law performance, those extreme hires are worth enormously more, so the overall value to the organization increases.

The shape of the performance distribution, by itself, does not break the value proposition of assessments.

It means that our initial hypothesis that “if performance follows a power law, your assessments destroy value” was wrong.

What Actually Breaks: The Assessment Can’t See the Tail

The key insight came from asking a different question. Instead of “does the distribution break the assessment?” we asked: “what if the assessment doesn’t work the same way for everyone?”

Most criterion validity studies are conducted on hired candidates. The performance of rejected candidates is never observed. This is the range restriction problem that Sackett and Yang (2000) documented in their expanded typology, and that Sackett et al. (2022) revisited in their reassessment of meta-analytic validity estimates. If the rejected pool contains extreme performers (e.g. mavericks, contrarians, people with unconventional profiles, etc.) then the reported r=0.20 is estimated on a truncated distribution and says nothing about how the instrument performs in the region where it matters most.

Think about what assessments actually measure. Conscientiousness. Conformity to structured processes. Pattern-matching to conventional markers of competence. These are useful predictors across the broad middle of the distribution. But the people who generate outsized returns often do so through mechanisms the assessment wasn’t designed to detect such as risk tolerance, contrarianism, obsessive depth in areas the instrument doesn’t cover, the willingness to break a process that everyone else follows. We could go on.

The assessment isn’t wrong about what it measures, it just can’t tell you who you missed out on that could have really shaken up the organization.

We tested three validity models:

Uniform validity: r=0.20 across the entire distribution. The standard assumption.

Tail-blind: r=0.20 for most candidates, but r≈0 for the top 5%. The assessment simply can’t see extreme performers.

Tail-adversarial: r=0.20 for most candidates, but r=−0.15 for the top 5%. The assessment actively penalizes extreme performers because they are nothing like the typical medium to high performer.

Figure 2. How assessment validity diverges from true performance in the right tail. Red dots = top 5% true performers. Green line = selection threshold.

Under uniform validity, 67% of top-5% performers were rejected. Under tail-blind validity, 80% were rejected. Under tail-adversarial validity, 87% were rejected. The assessment is actively screening out the organization’s most valuable potential hires.

How do you think your CEO would react if they heard that?

Interestingly, even though nearly 9 out of 10 of extreme performers were screened out, the net effect of the assessment still added value when you measure by raw performance alone. The false-negative cost on extreme performers is real, but the assessment’s ability to screen out low performers across the other 95% of the distribution compensates.

To find the limit of value creation, you need one more ingredient. You need to account for the fact that organizations are coupled systems.

What Weiss Teaches Us About Organizational Impact

So far we’ve treated each candidate as an isolated unit with an assessment score that predicts their personal output, and that’s all that matters. In a real work environment that assumption is wrong, and relaxing it is where the value inversion appears.

In an organization, a hire’s impact extends beyond their own output. A VP leading a 100-person function who performs at 2.5× doesn’t just produce 2.5× the widgets. She raises the performance of the people around her through better decisions, clearer strategy, higher standards, and stronger talent development. A poor-performing leader of the same function doesn’t just underperform personally. He degrades the output of hundreds of people downstream.

The problem is that modeling this realistically is intractable. An organization with 10,000 employees has roughly 50 million pairwise relationships. We can’t simulate all of them. But we don’t need to, and this is exactly the problem Weiss solved for magnetism in 1907.

Weiss’s insight was that you don’t need to track every atom-to-atom interaction to understand a magnetic system. Each atom responds to the average field generated by all the others in its neighborhood. Replacing a web of pairwise interactions with a single mean field made the physics tractable without sacrificing the key dynamics. Most importantly, it preserved the critical finding: that small perturbations can cascade through a coupled system in ways that are invisible if you study each element in isolation.

We applied the same logic to organizations. Instead of mapping every reporting relationship, collaboration, and cross-functional dependency, we modeled each hire’s effect on the typical experience of the people within their influence radius. That’s the mean-field approximation applied to organizational performance:

Organizational impact = raw performance + (performance − 1.0) × radius × coupling

In this model two parameters control the organizational impact of hire: influence radius (how many people this role touches) and coupling strength (how deeply the role affects each person it touches). An isolated individual contributor has a radius near 1 and coupling near zero. For this employee their personal output is their organizational impact, and the non-coupled simulation from our baseline applies. A VP leading a 100-person function has a radius of 100 and coupling of 0.10 or more. For that role, missing a 10× performer doesn’t just cost you their output. It costs you the uplift they would have provided to every person in their radius.

Figure 3. The same 2.5× performer generates different organizational impact depending on their influence radius.

A person performing at the mean (1.0) neither lifts nor drags the people around them. Someone performing at 2.5× lifts everyone in their radius. Someone at 0.5× drags them down. The coupling strength determines how much of the deviation transmits to each person in the radius. This is symmetric by construction: there’s no separate “bad hire” or “star” logic. Similar to Weiss’s approach we are simply looking at performance deviations propagating through a network with damping.

A note on what we’re borrowing and what we’re not. This is not a literal application of mean-field theory. We’re not solving a partition function or deriving a phase transition. What we’re taking from Weiss is the modeling strategy and replacing an intractable web of interactions with a tractable average effect. Like mean-field theory, our model relies on the assumption that influence can be approximated as a uniformly diffuse sphere. A VP affects each of her 100 reports roughly equally, and that the coupling strength is constant across the radius. That’s obviously a simplification. In practice, influence is lumpy: a leader might transform the three people she works with daily and barely register with someone four levels down in a different office. The model doesn’t capture that asymmetry. What it does capture is the scaling relationship between influence reach and the cost of selection errors, and that’s sufficient to identify the conditions under which assessments break. A more granular network model would change the specific thresholds but not the qualitative finding.

This is why coupling changes the false-negative calculus. In the baseline simulation, missing an extreme performer costs you their individual output, but the assessment’s ability to screen out weak candidates compensates. Once you add mean-field effects, the cost of that same false negative scales with the number of people affected. What was a linear loss becomes a multiplicative one.

When you combine a mean-field assumption with the three validity conditions across Gaussian and Pareto performance, that multiplication shows up clearly:

Figure 4. Assessment value vs. organizational reach of the role. Under Pareto performance with tail-adversarial validity, assessments cross from adding to destroying value as influence radius increases.

Under Gaussian performance (left panel), all three validity models provide net positive value for the organization regardless of how well the assessment handles extreme performers. The mean-field term amplifies good and bad hires roughly symmetrically, and the assessment’s advantage in screening out low performers holds. Under Pareto α=1.5 (right panel), the uniform and tail-blind models still add value. But the tail-adversarial model climbs steeply as influence radius increases. By the time you’re hiring for a role that touches 25 people, the assessment is destroying value. By radius=100, it’s destroying a lot of value.

The inversion requires three conditions simultaneously:

1. Heavy-tailed performance — the role is one where a small number of people produce wildly disproportionate output.

2. Degraded or adversarial validity in the right tail — the assessment systematically misclassifies the extreme performers who drive that disproportionate output.

3. High organizational influence — the role sits at a point in the network where performance deviations propagate widely.

All three conditions must hold. Remove any one of them and the assessment remains net positive.

A Framework for When to Trust Your Assessment

Figure 5. When do assessments break? Three of four quadrants are safe. The danger zone requires heavy tails, poor tail validity, and high organizational influence simultaneously.

For most roles and most hires the standard I-O psychology playbook works fine. Use your structured interviews, your personality and cognitive ability tests, your work samples. They will improve on random selection, and the improvement is real.

These simulations suggest that companies seeking extreme performers should consider alternate approaches under the following conditions:

Roles where performance is power-law distributed — typically creative, strategic, technical, or leadership roles where the difference between good and great isn’t 20% but 10× or 100×.

Roles with assessments that have never been validated on the right tail — which, due to range restriction, is most of them.

High-influence positions — roles where the hire’s performance ripples outward through teams, decisions, strategy, and culture.

For these roles the conventional assessment toolkit may be doing exactly the wrong thing. Not because the assessments are bad at what they measure, but because they were never built to identify extreme performers in a coupled system.

A Necessary Caveat: This Is Not a Personality Argument

When people hear that assessments may screen out transformational performers, a tempting conclusion follows: the difficult, abrasive, rules-breaking candidate is the hidden genius your process is rejecting. While some tech leaders have made headlines by explicitly hiring abrasive, asocial people, it would be a poor conclusion to hire people because they are jerks.

What the simulation identifies is a measurement problem, not a personality type. The performers being screened out aren’t necessarily difficult people. They’re people whose value is generated through deep specialization, lateral thinking, unconventional career paths, and the ability to see a problem differently than the consensus. Some of them may interview beautifully. Some may be the most collaborative people in the room. The assessment can’t tell, because it was never calibrated on that region of the performance space.

The point is not that organizations should tolerate toxic behavior in the name of genius. The point is that assessments are calibrated on the center of the distribution and validated on the center of the distribution of people that you did hire. The tail is where the transformational performers live and where you should feel some FOMO.

What This Means for Practice

We’re not arguing that organizations should stop using assessments. The baseline result is clear: for the vast majority of hiring decisions, assessments improve outcomes. That’s well-established science, and the simulation confirms it even under assumptions that most I-O psychologists have not modelled before.

What we are arguing is that the value proposition of an assessment is not constant across roles. It depends on the interaction of performance distribution, tail validity, and organizational influence.

The practical implication is that organizations need a differentiated selection strategy. Use standardized assessments for roles where performance is roughly normally distributed and organizational influence is moderate. For high-influence roles in power-law domains, invest in selection methods that specifically focus on identifying extreme performers: extended work trials, portfolio-based evaluation, reference networks that can speak to exceptional (not just adequate) performance, and decision processes that are explicitly designed to resist the pull of conventional pattern-matching.

The mean-field model suggests that the organizational cost of getting this wrong scales with the product of the performance deviation you miss and the number of people affected by that miss. For an isolated contributor, the cost is linear. For a connected leader, it’s multiplicative. The same false negative that barely registers for one role can be catastrophic for another.

Pierre-Ernest Weiss showed us that the behavior of a system can’t be understood by looking at individual elements in isolation. The same principle applies to organizations. The value of a hire is their impact on the entire company. Any selection system that ignores that coupling is, at best, incomplete. At worst, it’s systematically filtering out the people who would have mattered most.

Methodology

The Monte Carlo simulations in this article used 100,000 candidates per run, 50 iterations per condition, with assessment validity r=0.20 and a 20% selection rate. All code is available for replication. The mean-field spillover model uses the formulation: impact = performance + (performance − 1) × radius × coupling, with coupling strength fixed at 0.10 for the consequence analysis.

References

O’Boyle, E. H., & Aguinis, H. (2012). The best and the rest: Revisiting the norm of normality of individual performance. Personnel Psychology, 65, 79–119

Sackett, P. R., & Yang, H. (2000). Correction for range restriction: An expanded typology. Journal of Applied Psychology, 85, 112–118

Sackett, P. R., Zhang, C., Berry, C. M., & Lievens, F. (2022). Revisiting meta-analytic estimates of validity in personnel selection. Journal of Applied Psychology, 107(10), 1651–1680

Weiss, P. (1907). L’hypothèse du champ moléculaire et la propriété ferromagnétique. Journal de Physique Théorique et Appliquée, 6(1), 661–690. Wikipedia: Mean-field theory

Sackett, P. R., Lievens, F., Berry, C. M., & Zhang, C. (2023). Revisiting the design of selection systems in light of new findings regarding the validity of widely used predictors. Industrial and Organizational Psychology, 16(3), 283–300

Leave a comment

Share Tyler Weeks

Share

Sound familiar?

Subscribe now

One prevailing school of thought, pioneered by Frederick Winslow Taylor, focused on the worker. He believed that if you broke a job down into its smallest constituent motions, and optimized each one, you could create a much more efficient worker. Taylor sought to atomize labor.

On the other hand, Henry Ford was focused on the work. Ford resented Taylor’s fixation on the workers; he didn’t want to manage workers, he wanted to manage the process. With the assembly line, he sought to atomize the process into tasks, bringing the work to the station so that the system dictated the pace.

Both approaches relied on the same underlying assumption: Aggregativity. Both believed that a company was the sum of its parts. If you optimized the parts, or the sequence, you optimized the system.

The approach worked remarkably well for highly repetitive work. It worked so well we called it the Second Industrial Revolution. We got orders of magnitude more efficient at making a lot of the same thing. As we navigate the latest industrial revolution it’s only natural that we’d turn to age-old recipes.

Fast forward to the 2010s and Taylor’s Scientific Management saw a major resurgence. The concept of the skills-based organization arrived on the scene and was generally received as a new idea.

Just as Taylor advocated, to become a skills-based organization the first step was to deconstruct work into its underlying tasks. At the same time each employee was to be viewed as a collection of skills that could be dynamically matched to those tasks. Work was simply another supply chain problem. The important assumption that underlies this approach is that tasks are the functional unit of work, the atoms that make up a molecule.

What Is the Functional Unit of Work?

Consider caffeine.

If I told you that caffeine is made of carbon, hydrogen, nitrogen, and oxygen atoms arranged in a specific pattern, could you predict its effects? Could you deduce from that atomic inventory that it crosses the blood-brain barrier and blocks adenosine receptors, producing alertness? No. What caffeine actually does to a person emerges at a level of organization that cannot be deduced from the atomic parts alone.

Furthermore, the strength of caffeine’s effects depends on the physiology of a specific human body. The same molecule can produce different effects depending on metabolism, tolerance, what else has been consumed, time of day, and stress level. The strength of the effect emerges from the interaction between molecule and system.

The same principle applies to organizations. Skills and tasks are like atoms: they are necessary, but not sufficient to explain how work gets done. A team’s capability emerges from how its members interact with each other and with the organizational system around them. You cannot predict what a team will accomplish from a skills inventory any more than you can predict caffeine’s effects from a periodic table.

As we seek to understand the physics of organizations, the question is: what is the smallest meaningful unit of work?

The skills-based movement makes the claim that we should be engineering organizations at the atomistic level of the skill, or perhaps the task. However, if skills are not the functional unit of work then our picture may be incomplete in a way that leads to unexpected outcomes when we start making large-scale design decisions based on it.

What Skills-Based Organizations Get Right

Before we go any further we must acknowledge the tremendous value skills-based thinking has delivered in recent years.

Hiring Equity: The Rework America Alliance found that skills-based practices helped workers from lower-wage roles move into positions offering higher wages and more economic mobility.

Internal Mobility: Talent marketplaces that match employees to projects have saved hundreds of thousands of work hours at companies like Rolls-Royce and MetLife.

Transparency: With the half-life of a skill now estimated at less than five years, skills-based organizations provide visibility into capability gaps that job titles obscure.

These are genuine achievements. They matter to the people whose careers benefit. But there is a conceptual leap from “skills are useful for the organization” to “skills are the organization.” There is evidence to suggest that it is a leap too far.

The Case for Emergence

The skills-based model assumes that capability is aggregative. Put simply, if skills are the building blocks of work then as long as you have enough Python skills and Strategic Planning skills, you can arrange them into a capable organization.

On the contrary, organizational behavior literature and complexity science suggest capability is emergent. It cannot be assembled purely from skills and tasks any more than your mind can be reconstructed from a bag of neurons.

First, a clarification: skills are necessary. No one is arguing that you can build a high-performing team without individuals who have demonstrated proficiency in their field of expertise. The question is whether skills alone are sufficient to explain how an organization produces value.

The question matters beyond academic interest. If productivity is in fact an emergent property of the organization, as defined by complexity science, then the aggregative approaches of the 20th century won’t work. To test this claim we can use William Wimsatt’s via negativa: emergence is a “failure of aggregativity.” This concise statement, along with the four tests he outlines, provides a powerful lens through which to scrutinize the limits of decomposing work into skills.

For a system property (e.g., a team’s ability to innovate) to be considered a sum of its parts, and thus manageable by a skills-based model, it must pass four restrictive conditions. If the property fails these tests, we can consider it emergent and we have to reject Taylor’s and Ford’s models.

Condition One: Invariance Under Rearrangement. In a pile of gold bars, the total value is the same whether you stack them in a pyramid or a line. Your wealth is invariant to the arrangement.

Skill Test: If you take a software team and rearrange their interactional structure, moving them from an agile squad to functional silos, their productivity changes drastically. The same ten engineers, with the same skills, produce measurably different output depending on whether they sit together, share daily standups, and can interrupt each other versus communicating through tickets across a departmental boundary. The skills didn’t change. The topology did.

Result: Fails Condition One.

Condition Two: Invariance Under Substitution. In a LEGO wall, replacing a red brick with an identical red brick doesn’t change the wall. The parts are fungible.

Skill Test: Boris Groysberg’s study of more than one thousand star security analysts found that when stars switched firms, their performance plunged for at least five years. Same people, same set of skills, different results. Their performance was dependent on their specific web of relationships and institutional knowledge.

Result: Fails Condition Two.

Condition Three: Invariance Under Decomposition. A bank account balance is the same whether you view it as one sum or ten smaller sums.

Skill Test: Stasser and Titus tasked a committee with choosing the best political candidate. Crucially, they decomposed the information: they gave each member different clues, which, if combined, clearly identified the winner. The result? The groups consistently failed. Instead of pooling their unique puzzle pieces to see the whole picture, members fixated only on the few facts they already shared (aka the Common Knowledge Effect). The sum of the information was present in the room, but the integration, the actual competence, was destroyed by the decomposition.

This is what a skills taxonomy does to organizational knowledge. It decomposes integrated, context-dependent expertise into discrete labels (e.g., strategic planning, stakeholder management, data analysis) and assumes the organization can be reconstructed from the inventory. But the competence was never in the labels. It was in the integration.

Result: Fails Condition Three.

Condition Four: Linearity. If one engine produces one hundred horsepower, two produce two hundred.

Skill Test: Human systems are non-linear. Two researchers with complementary knowledge can produce a discovery neither could make alone. One plus one is greater than two. Conversely, Brooks’ Law shows us that adding more people to a late project often makes it later due to communication overhead. Two people who trust each other can have a difficult conversation that leads to a breakthrough. Add a third person and the same conversation may never happen.

Result: Fails Condition Four.

Across all four conditions, organizational performance can be shown to be non-aggregative. It is not invariant to rearrangement, substitution, or decomposition, and it is not linear. By Wimsatt’s criteria, organizational productivity is emergent. This is not a metaphor. It is a formal classification with specific implications: you cannot reliably predict system-level performance from component-level specifications. A skills taxonomy, no matter how comprehensive, is the wrong level of analysis for understanding how to build highly productive teams.

If Not Skills, Then What?

If skills fail the test for aggregativity, the natural question is: what doesn’t fail? What actually drives performance?

The evidence points toward something that hasn’t yet entered the HR zeitgeist, though pieces of it appear across several research traditions.

Daniel Wegner called one piece of it transactive memory, the collective system of “who knows what” that develops when people work together over time. It’s not what any individual knows; it’s the group’s shared map of its own expertise. When you remove a team member, you don’t just lose their skills. You lose a node in the team’s knowledge network, and the remaining members lose the ability to coordinate around that knowledge. Wegner’s insight was that this system is a property of the group. It cannot be decomposed into individual knowledge inventories and reassembled elsewhere.

Anita Woolley found another piece. Her research on collective intelligence showed that a group’s cognitive performance is not predicted by the average or maximum IQ of its members. You can stock a team with every skill in your taxonomy and every genius on your payroll, and their collective intelligence may still be mediocre. What predicts group performance is the social sensitivity, equality of turn-taking, and the communication network itself. The capability is a property of the interaction pattern, not the roster.

Amy Edmondson found a third piece. When cardiac surgery teams adopted a new technology, success depended not on the skill of the individual surgeon but on whether members felt able to speak up, flag errors, and propose alternatives without fear. Teams with identical technical skill profiles had dramatically different outcomes based on this single emergent property that no individual possesses alone.

Consider a thought experiment: you’ve hired the smartest person who has ever lived. They know everything. They can do everything. Are you guaranteed to have the most productive team? Of course not. If they sowed doubt and alienated the team members from each other, they might have the opposite effect.

What Wegner, Woolley, and Edmondson are each describing is the same underlying phenomenon viewed from different angles. The functional unit of productivity is emergent capabilities. Emergent capabilities are an integrated configuration of people, trust, shared knowledge, and communication patterns that produces performance no individual member carries alone.

If that is true, it changes what an organization should look like.

The Emergence-Based Organization

If emergent capabilities are the functional unit of productivity, then every major people-process in an organization needs to be extended to account for what skills alone cannot explain.

A skills-based organization asks: do we have the right skills in the right roles? An emergence-based organization asks a harder question: are our people configured in ways that produce emergent capabilities, and can we see when those capabilities form, degrade, or fail to develop?

This is not a rejection of skills-based thinking. It is the next layer. You cannot build emergent capability from unskilled people. But an emergence-based organization recognizes that assembling skilled individuals is where the work begins, not where it ends.

What would such an organization look like in practice?

Hiring. A skills-based organization treats hiring as a one-directional evaluation: does this candidate have the skills we need? An emergence-based organization recognizes that hiring is work for both sides. The team must adapt to the new member just as the new member adapts to the team. Emergent capabilities are properties of the configuration, not the individual. Groysberg’s research illustrates this clearly: star analysts who moved with their teams showed no performance decline, while those who moved solo suffered for years. The difference wasn’t skill. It was whether the receiving system was prepared to integrate them. Screening for individual proficiency remains necessary, but an emergence-based organization would also assess the team’s capacity to absorb a new member and actively prepare the configuration for the transition before the new hire walks in the door.

Onboarding. If hiring prepares the configuration, onboarding is where integration actually happens. The standard model optimizes for individual time-to-productivity: how quickly can the new hire complete tasks independently? An emergence-based organization would optimize for team recovery: how quickly does the team return to its previous level of coordination, knowledge sharing, and collective problem-solving? This reframes onboarding as a group activity. The entire team must be re-onboarded and is collectively accountable for the outcome. Success is measured at the team level, not the individual level.

Team Design. Most organizations design teams around reporting structures and skill coverage. An emergence-based organization would design around emergent capability. Which combinations of people produce the highest collective intelligence? Which configurations develop transactive memory fastest? When a team is working well, the emergence-based organization recognizes that the configuration itself has value and protects it from unnecessary disruption. When a team is underperforming despite having adequate skills on paper, the diagnosis shifts from “who is underperforming?” to “what is preventing emergence?”

Performance Management. We currently evaluate individuals and aggregate those evaluations upward. An emergence-based organization might consider a more objective measure of individual performance that is derived from the impact of that person on a team’s productivity. Imagine this as similar to plus/minus statistics from sports. An individual might be unremarkable when measured in isolation but have a disproportionately positive effect on every team they support. Conversely, a high-performing individual may be depressing the team’s emergent capability by dominating communication channels or undermining psychological safety. Neither of these situations is visible to a system that only measures individuals.

Workforce Planning. Traditional workforce planning models headcount against demand forecasts. An emergence-based organization would model the network impact of personnel moves before making them. When someone gives notice, the question is not just “what skills do we need to replace?” but “which emergent capabilities will degrade, and how do we minimize that degradation?” Internal mobility would be evaluated for its impact on the teams people are leaving, not just the teams they are joining.

Measurement. None of the above is possible without the ability to see emergent capabilities as they form, degrade, and recover. Until recently, measuring constructs like transactive memory, collective intelligence, and psychological safety required costly, point-in-time research interventions. These methods are far too slow for managing emergent capabilities.

Enterprise chats, emails, and tickets provide a continuous stream of interaction data: who communicates with whom, how frequently, on what topics, with what response patterns. This is the raw signal of emergent capability. This information has long been used for generating passive organizational network analyses, but privacy concerns have mostly limited these networks to volume of communications, not intent or context.

The enterprise LLM changes the equation. Viewed as a measurement device, it can make sense of communication exhaust to understand work while protecting the privacy and trust of employees. Any company could run a straightforward first experiment today: issue surveys in the form of a prompt that respondents execute in the LLM. The prompt surfaces a work summary that includes information necessary for organizational network analysis, skills mapping, and evidence of transactive memory, collective intelligence, and psychological safety. The employee sees exactly what information is being requested and can edit the results as desired. If initial experiments are successful, an automated employee-in-the-loop report could be generated at regular intervals that provides the information necessary to study and promote emergent capabilities while maintaining trust through transparency.

Two foundational measurement problems would need to be solved.

First: The Ephemeral Team. The map of who works with whom is not stored in a static org chart or a project management tool. The output of your company is highly dependent on teams that form and dissolve organically, that include members from multiple disciplines, and often span multiple projects. An always-on organizational network analysis with a refresh rate that matches the pace of business decisions would be a prerequisite to understanding those teams. Without the ability to see when and how teams form, we cannot understand how they operate. When someone gives notice, we should be able to see not just their job title and skills, but the network of teams and knowledge flows that will be disrupted.

Second: Emergent Health Monitoring. As teams form and disband in the natural flow of work, we need non-invasive ways to monitor the health of their coordination. This is not a sentiment analysis. A team with a high degree of psychological safety may experience spikes in conflict precisely because people feel safe communicating directly. What matters is the decay curve: conflict happens, but how quickly does the team recover? How quickly do they share expertise after a disruption? How rapidly does a new member become integrated into the team’s knowledge network? We would expect highly productive teams to be highly resilient.

Network topology, knowledge flow patterns, coordination speed, integration rates, and recovery curves are proxies for the emergent capabilities that Wegner, Woolley, and Edmondson identified. They are measurable. And they give us something we’ve never had before: a way to design every people-process around the actual emergent structure of the organization rather than a generic skills checklist or a static org chart.

Adaptive Staffing

Skills-based talent marketplaces promised a fluid organization where people could be dynamically matched to work. That promise has gone largely unfulfilled because the model was incomplete. Matching skills to tasks doesn’t account for the emergent capabilities that are disrupted when someone leaves or joins a team.

Realizing that vision will require achievement of lossless turnover. Lossless turnover is the ability to move people in and out of teams with minimal degradation of emergent capabilities. It requires everything described above: seeing the ephemeral team, monitoring emergent health, and measuring integration rather than just individual task completion. When the new hire arrives, success is not whether they can do their job. It is whether the team recovers its previous level of coordination, knowledge sharing, and collective problem-solving.

Achieving lossless turnover would enable the truly adaptive organization capable of deploying effective teams rapidly, not because it has a marketplace of interchangeable skills, but because it understands and can actively manage the emergent capabilities that actually produce value.

The Next Industrial Revolution

Skills-based thinking gave us a vocabulary for matching people to work. Emergence-based thinking gives us a vocabulary for understanding why some configurations of people produce extraordinary results and others don’t.

The practical question is not whether emergent capabilities matter. The question is whether we can build the measurement and management infrastructure to act on that knowledge. The pieces are available: organizational network analysis provides the topology, LLMs provide the semantic layer, and decades of research on transactive memory, collective intelligence, and psychological safety provide the constructs worth measuring.

What we lack is robust validated LLM-based measurement and integration. No existing HR technology platform was designed to manage emergent capabilities. Headcount management systems track individuals in roles. Skills platforms track attributes of individuals. Neither tracks the configurations of people that produce value. The emergence-based organization requires a new category that captures ephemeral teams and their emergent capabilities.

The hallmark of our success in this new Industrial Revolution will be a whole generation of technology for the organization fueled by a prolific era of organizational measurement and research.

Subscribe now

Share

Share Tyler Weeks

Leave a comment

The best answer I can give about why I did it is that it looked interesting. I can’t think of a better reason. It wasn’t easy, but I’m very happy with that decision.

For anyone contemplating a similar journey, it’s going to take you way more than 10,000 hours, and there’s no shortcut I can offer you. What I can share with you are some mistakes I made and lessons I learned along the way. Maybe these can help you navigate that journey with a bit more grace than I have.

One: You’ll Become a Stranger in a Strange Land

When you jump fields mid-career, you stop belonging anywhere. You’ll have abandoned the discipline that shaped you. Your old peers won’t see you as one of them, and your new field will see you as an amateur. You won’t have the credentials, experience, or pedigree that signal your status in any community.

You have to be okay with leaving your pride behind. Really okay with it. Because if you’re waiting for permission or validation from either side, you’re going to be disappointed. Of course, there will be lovely friends and colleagues on both sides that will encourage you, but kindness isn’t the same thing as acceptance as a peer.

You’re a mutant. Innovation lives at intersections, which is right where you’ll be setting up camp, but you have to be willing to sit there longer than you might think.

Two: There Is No Dumbledore

No one is coming to rescue you. No one’s going to find you under a staircase, see your potential despite your circumstances, and whisk you off to Hogwarts to fulfill your destiny. There will be plenty of Rons and Hermiones. They will befriend you, give you advice, and let you tag along, but no one can do the work for you.

You have to do the hard work of jumping into problems, showing that you can deliver value, have good ideas, and be a contributing member of the community. It’s on you to prove yourself. It is not anyone else’s job to discover you.

Three: The Job Is the Best Credential

Several people contemplating a similar jump have asked me how important another degree might be. The most successful people I’ve seen making this transition didn’t credentialize their way in. They found a role or a project within their current company and jumped straight into real projects and learned by doing.

A mechanic’s credibility isn’t determined by the number of tools they can name; it’s based on the number of cars they have fixed. An online certification could help jump-start you, but if you already have an advanced degree in any of the sciences, another degree isn’t going to make you more credible; real experience will.

When you’re trying to break into a new field, real work on real problems beats credentials pretty much every time. That’s where you prove yourself.

Four: You Don’t Need to Become Someone Else

Don’t jump fields to become a new person. Jump fields to solve new problems.

When I first got into people analytics, I admired the researchers, leaders, and practitioners in Data Science and IO psychology that I saw leading the discipline. My early mistake was to think that success would mean becoming one of them. It took me the better part of a decade to realize that perfect assimilation was not the goal.

My background as a physicist can, and does, provide a valuable perspective. Your value won’t be in doing your best impression of the people you admire. Your weird way of viewing age-old problems will be your value. People analytics needs more people with different backgrounds and perspectives.

Five: Rocket Science Schmocket Science

Physicists can be the most arrogant of scientists. One of the least cute hallmarks of that arrogance is the assumption that you’ve been working on the hardest problems so everything else will be some degree simpler. My first thought was, “How hard could it be?”

Pretty hard.

Just imagine if a rocket’s telemetry also depended on whether or not its feelings were hurt. We’d have never made it to the moon.

This field has existed for over a hundred years. Pretty much every idea you can think of has been tried. The reason solid answers are elusive isn’t because people haven’t thought to ask the questions. You’ll have a head start if you start from the premise that human systems are complex and chaotic in the strictest sense, and that a person’s reality often has nothing to do with measurable facts.

Six: Hi. My Name Is Dunning-Kruger.

Lesson six is very similar to lesson five, but I think it deserves its own callout. My first several years were spent confusing the thrill of learning new things with the thrill of discovering new things. I believed I was contributing novel ideas way earlier than I was, simply because I had never heard of them.

If you’re going to make this jump, prepare yourself for how long it will take to understand a field well enough to contribute something genuinely new. I’m still not sure I’m there.

Seven: People Are Nuts, Not Electrons

My first instinct when I jumped into People Analytics was to try modeling people like subatomic particles, expecting to find that some familiar physical laws might lead me to an elegant understanding of people. That approach makes for a cool presentation, but that’s about it.

Metaphors are useful, but the lesson here is to not confuse a metaphor for a model. Human judgment and decision-making are emotional and irrational in ways that electrodynamics is not. People aren’t electrons. Sometimes, if you squint really hard, you could say they act like a fermion, but that’s only sometimes and not really. We wouldn’t be satisfied with that in physics, and no one in People Analytics is going to be satisfied with it either.

There’s no way around the humans. Your ideas have to work because of them, not in spite of them. If you want to imagine a world where people are different, then go write a novel.

Eight: Nuance All the Way Down

In physics, you have fundamental principles. You can build logically from first principles to explain observable phenomena. If A causes B and B causes C, then there’s something interesting about the relationship between A and C.

In social sciences, you have to be very careful with inferential chains. There’s too much noise, too much variance, or too much context-dependence to establish causation. I really didn’t understand just how nuanced some of the important results were. I treated research findings like the laws of the universe when they were actually much more conditional and context-dependent.

Academic findings are rarely as certain or universal as they sound in a paper’s abstract. Understanding the nuances of those findings will give you more credibility than simply knowing about them.

Nine: Don’t Throw the Physics Out with the Bathwater

Recognizing that physics didn’t explain everything didn’t mean it didn’t apply. This was my rebound mistake. Don’t abandon what you know; look for where it does make sense. That could be your niche.

When I saw how irrational people were and how messy the systems were, I abandoned physics. I figured if deterministic thinking didn’t work, then physics was best served up as an interesting factoid on my resume.

That was also wrong. I was just using the wrong physics.

There is a physics of organizations grounded in chaos theory and complexity science. Organizations are like a flock of European starlings with double pendulums rolling dice on their backs.

Ten: The Assessment Bone Is Connected to the ONA Bone

There are natural fault lines within the domain of people science that practitioners tend not to cross. Just because they avoid those fault lines doesn’t mean you have to. After all, performance management is just another survey.

[the audience gasps in horror]

The pace of technology development in the people space has been surprisingly stagnant, given the tremendous investment that has gone into understanding organizational dynamics. Some of the best work you can do as an outsider is to ask really dumb and/or heretical questions. Push on these boundaries. Try to synthesize observations into a coherent mental model of the organization and the people in it.

If we’re going to develop technology that redefines the future of work, we need to break down the walls and hash out a theory of work that holds all of the field’s observations true at the same time.

Synthesis is its own kind of discovery and one that a naive newcomer could be well positioned to precipitate.

Eleven: Measurement Was Always the Limitation

Here are two important things I wish I’d understood sooner about people analytics:

• A lot of critically important things are either impossible to measure, expensive to measure, or unethical to measure.

• The measurement tools we have are very low resolution.

Start with these two things in mind, and you’ll have a leg up on anyone proposing fantastical approaches.

If you are an experimentalist by training, you have all the tools you need to be very successful in this space. You know that you can’t prove something right or wrong if you can’t measure it. Especially with the advent of AI, there is an opportunity to devise new measurement methods.

New measurement methods often precede new discoveries.

Twelve: If Measurements Could Speak

I learned this early on but didn’t really appreciate it until a couple of years ago: measurement isn’t passive in social systems. Every survey you send out, every performance management framework you design, and every way you ask a question changes how your workforce thinks about that topic.

The way you ask someone about their pay teaches them how they should think about their pay. The way you ask them about their manager teaches them how they should think about their manager. The words you use and the metrics you choose to track all communicate something.

You have to design for both. You can’t treat measurement as a purely technical exercise; every measurement is also a message, and you’re responsible for both. I wish I’d learned this one about five years earlier than I did.

Thirteen: The Black Box Leading the Black Box

Data science is almost never science.

Our job is to understand what makes work a fulfilling experience for employees and a profitable exercise for the company. A result that doesn’t teach you anything about either has limited value. You may find a use for that result, but if you don’t understand why it works, you won’t understand when it breaks.

Black box methods might optimize a specific outcome, but they don’t deepen your understanding. And in this work, understanding is required for progress.

Fourteen: The Emperor’s New Tool

When a vendor shows you glowing testimonials about a new feature that claims to be an elegant solution to people problems, remember that those clients have a vested interest in proving their purchase was a good decision. Take endorsements with a grain of salt, especially if they can’t tell you two bad things about the product.

If someone can’t tell you anything negative about the product, they don’t understand the product, they don’t understand the problem, or they’re convincing themselves that they made the right decision. You and your purchase have nothing to do with it.

Client testimonials aren’t data. If you’re trying to decide whether something actually works, first check to see if it violates any known laws of the universe. I don’t care how many people swear by it, if it’s too good to be true, then it’s too good to be true.

Fifteen: The Buddha Was Right

People analytics isn’t about the analytics; it’s about the people.

There is no stat that can tell you if a football game was a good one or a bad one. The only way to know that is to understand what people expected, what they wanted, what mattered to them… what ideas they were attached to. If your team lost but played way better than expected, maybe it was a great game. If your team won but not by as much as you thought, maybe it felt disappointing.

In biology, an umwelt is the subjective perceptual world of an organism. The umwelt of a dog is heavily influenced by scent. That sense determines the reality that a dog is equipped to perceive and experience. The human umwelt is shaped by emotion and expectation; it’s their internal landscape.

You can’t measure the umwelt, and you never will. In that sense, people are kind of like black holes. You can’t see directly inside, but you can observe the effects of their inner-workings based on their effect on everything around them.

Maybe people analytics is more like the study of dice-rolling double pendulums on the backs of a flock of black holes.

The Wrap-up

But seriously, what could be more interesting than a flock of black holes? This is why we need more people to make the jump.

Do it if you have deep, persistent intellectual curiosity, genuinely fascinated by these problems, and willing to sit with discomfort for years while you learn a new field from the ground up. Don’t do it if you’re looking for an easier path or quicker wins. This work requires humility, patience, and a willingness to be bad at something for a long time. The problems are real, they’re unsolved, and they are tremendously important.

And honestly? It’s the most rewarding work I’ve ever done.

Share

Share Tyler Weeks

Leave a comment

Subscribe now

A simple, linear system, like the pendulum, lends itself to a simple equation. As long as you know a few facts about the length, mass, and starting position, you can predict exactly where it will be in ten seconds, ten minutes, or ten years.

Then things get interesting. The professor attaches a second pendulum to the bottom of the first one. This new contraption is called, you guessed it, a Double Pendulum.

He asks the class what they think it’s going to do. Everyone agrees it will swing much like the first one, perhaps with some extra wiggles. He releases it, and for the first swing, the class feels smart; everything looks normal.

But then, everything goes wrong. It loops, jerks, stalls, and spins. The professor resets it and does it ten more times, starting from the exact same spot. Each time, it takes a wildly different path. It never repeats the same motion twice. The class is stupefied. It is still a mechanical system. The laws of motion haven’t changed. Gravity hasn’t changed. Yet, adding one degree of freedom turned a deceptively simple system into something completely unpredictable.

People Analytics teams have been tasked with predicting individual performance since the discipline was born. We’ve built regression models, assessed skills, and analyzed engagement surveys, all to identify who has the potential to make the biggest impact. While there have been substantive advancements in understanding teams, prediction at the individual level has remained elusive.

If individual performance is the product of a chaotic system, like the double pendulum, then no algorithm, no matter how advanced, will ever be able to provide a meaningful long-term prediction.

Physical Systems: A Field Guide

In some situations, prediction will always be elusive, and no model in the world can change that. To understand why, we have to consider what kind of system we are managing. Generally, systems fall into four categories:

1. Simple/Linear Systems (The Pendulum): Input is proportional to output. If you push a lever twice as hard, it moves twice as far. These systems are straightforward to manage. For example, you might notice that call center volume is proportional to the number of employees enrolling in benefits. You predict a spike during Open Enrollment, increase staffing, and the problem is solved.

2. Random/Stochastic Systems (The Casino): The outcome of any single event is pure luck (like rolling a die), but the aggregate is predictable. You cannot predict a single roll, but you can predict the average of 1,000 rolls with extreme precision. This is where statistical methods shine. Assessments work this way: you can’t ensure any individual hire is the right one, but you can ensure that, over thousands of hires, your “hit rate” improves.

3. Complex/Adaptive Systems (The Ecosystem): The parts interact and learn. If you poke the system, it adapts and changes its own rules. You can’t predict it exactly, but you can understand its patterns. Company culture is a complex system. If you’ve ever tried to “transform” culture, you know you can’t dictate it; you can only guide or shape it, because it never stops evolving.

4. Chaotic/Non-Linear Systems (The Double Pendulum): This one can fool you. Chaotic systems aren’t random. They are deterministic—they follow strict rules—but they are highly sensitive to initial conditions. Microscopic differences in the start lead to macroscopic differences in the end.

Most HR models assume human performance is Type 1 (Linear) or Type 2 (Random). We assume that if we have enough data, the “noise” will average out. Dave Snowden’s Cynefin framework famously distinguishes between Complex and Chaotic domains, noting that in a Chaotic system, cause and effect are only perceivable in hindsight. In these systems, small differences don’t average out; they are amplified.

But is performance actually Chaotic? To be sure, we have to rule out the alternatives.

• If it were Random, high performers would eventually “regress to the mean” and become average. They often don’t.

• If it were Linear, adding more resources would always equal more output. It doesn’t (Fred Brooks famously proved this in The Mythical Man-Month).

Individual performance behaves like a Coupled System. You cannot solve for an employee’s performance without simultaneously solving for their Manager and their peers. Like the physics of a double-pendulum, people push and pull on each other, creating a complex, looping topology where outcomes are deterministic but unpredictable.

A missing laptop on Day 1 doesn’t just mean a lost day of work; under the right conditions, it can spiral into a lost career. If this is true, attempting to predict who will be a “High Performer” is a mathematical dead end.

Evidence of Chaos: Sensitivity to Initial Conditions

In the double pendulum, the infinite variation in swinging patterns is a result of extreme sensitivity to initial conditions, not randomness. It’s not that the pendulum randomly decides to switch directions; it’s that a variance of a fraction of a millimeter at the start leads to a totally different result 10 seconds later.

This matters because outcomes in chaotic systems are explainable even if they are not predictable.

Imagine we hire twin employees, Alice and Bob. They have statistically identical resumes, IQs, and psychometric traits. They are hired for the same role, on the same day, into the same team.

• Linear View: Their performance in Year 1, 10, and 30 should be identical.

• Random View: Their performance might differ in Year 1, but over 30 years, their averages would converge.

• Complex View: Their performance might diverge as they move to different groups (adapting to environments), but reunite if they return to the same environment.

Evidence suggests that individual performance often defies these models and behaves chaotically. Consider this scenario:

1. Day 1, 9:00 AM: Alice’s laptop is ready. Bob’s laptop is delayed by IT for 4 hours. A trivial, microscopic difference.

2. 12:00 PM: Because he has no laptop, Bob misses the Slack message inviting the team to lunch. Alice goes.

3. 1:00 PM: At lunch, the team jokes about a “Project X” disaster from last year. Alice learns the unspoken rule: Never mention Project X to the VP. Bob, eating a sandwich alone at his desk, does not learn this.

4. Month 2: In a meeting, Bob innocently compares a new idea to Project X. The VP grimaces. Bob is labeled “tone-deaf.” Alice stays silent and is labeled “politically savvy.”

5. Year 1: Alice is marked as High Potential. Bob is on a Performance Plan.

The stigma of the performance plan never leaves Bob. His motivation sags, his engagement withers, and he leaves after two years.

Was Bob a “bad hire”? No. Remember, the candidates were identical. The only difference was the initial condition (the laptop), which caused their paths to irreversibly diverge. This sensitivity is the hallmark of a chaotic system.

A Bit About Chaos

Although the specific sequence of events in a chaotic system is unpredictable, patterns do emerge. In physics, we call these patterns Attractors.

In a human organization, it is easier to think of Attractors as Currents.

Imagine Bob and Alice are kayakers on a river. You cannot predict the exact path of either kayaker. The paddler has agency; they are dodging rocks and adjusting their weight. Their path is “non-deterministic”—no two runs down the river will ever be the same.

The Riverbed determines the currents they will encounter.

Consider the Topology Diagram below through this lens:

1. The “Burnout Trap” (The Whirlpool): Notice the deep funnel at the bottom. This is a point where the current becomes too strong to fight. Once a kayaker drifts into this high-drag, high-pressure zone, the physics of the water pulls them under. It’s not that they “chose” to fail; it’s that the current overpowered their agency.

2. The “High-Performance Orbit” (The Eddy): Notice the ring at the top. In a river, an eddy is a place where the water swirls back upstream. It’s a dynamic sweet spot. The kayaker is still moving fast, still balancing, still working—but the water supports them. They can stay there indefinitely, in a state of “Flow,” without washing away.

In our organizational river, the more extreme the topology, the more likely our kayakers are to be pulled into these currents. While world-class athletes might be able to navigate the wildest rapids, the best way to ensure the success of most kayakers is to pay close attention to which river we’re putting them in.

Evidence of Chaos: Psychological Safety

In physics, the “chaotic nature” of a system is quantified by the Lyapunov Exponent (𝜆). This number tells you just how sensitive Bob’s and Alice’s long-term experiences will be to their first day on the job:

1. Low 𝜆: The system is stable. You can nudge the pendulum, and it returns to its rhythm.

2. High 𝜆: The system is chaotic. A change of a few atoms changes the entire future.

If Bob misses lunch or makes a small comment about Project X and feels little impact from an honest mistake, we’d describe this system as having a Low 𝜆. In this system, it is “safe” to experience small deviations because they dampen out over time; trajectories correct themselves.

Amy Edmondson and Google’s Project Aristotle famously identified “Psychological Safety” as the key driver of team performance. Their description aligns perfectly with a system that has a low Lyapunov Exponent:

“[People] feel confident that no one on the team will embarrass or punish anyone else for admitting a mistake, asking a question, or offering a new idea.”

The fact that these teams perform higher implies that psychological safety describes the physics of the team. In a “Low Safety” (High 𝜆) organization, every small error is amplified. A missed meeting becomes a reputational hit. A failed experiment becomes a mark on your permanent record. Performance becomes unpredictable and unmanageable, not because the people are difficult, but because the system is volatile.

Psychological Safety is not a “soft” concept or a nice feeling of inclusion. It is a critical metric describing the stability of your organization’s physics.

Focus on the Environment

The typical, and least productive, response to the Alice and Bob scenario is to focus on the laptops: “We must ensure all future hires get laptops at precisely 9:00 AM.” This is a fool’s errand. In a chaotic system, you can never control initial conditions with enough precision. There will always be a variable you missed.

If better hiring (smarter people) doesn’t equate to higher performance, but a more forgiving environment does, then we have strong evidence that we are dealing with a chaotic system. And if we are dealing with a chaotic system, then trying to guarantee an individual outcome is a waste of time.

The most important work we can do is to understand the attractors. We can’t predict where Alice will be in a year, but we can map the currents she’ll encounter. IO Psychology gives us clues on how to engineer a ‘High Performance’ attractor rather than a ‘Burnout’ attractor:

• Dampen the Noise: Promote cultures where small initial perturbations don’t cascade into failure.

• Focus on Recovery: Resilience is more valuable than perfection. How quickly does an employee’s trajectory stabilize after a shock?

Managing The Chaos

Our regression models are powerful tools but only if we are operating in regimes where they are valid approximations of the physical system they are meant to represent. Understanding where we have high λ and actively managing lambda will allow us to take advantage of those models. Here are three new types of measurement that could help us understand the physics of our organizations:

1. The Twin Test: Track “statistical twins” (similar hires). Do their paths stay parallel (Low Chaos) or do they diverge wildly (High Chaos)? This is the best direct measurement of λ you have. Practically we can identify these twins using statistical techniques like Propensity Score Matching to find ‘nearest neighbors’ and see if their outcomes vary proportionally to their differences. Roles or teams with high λ deserve your attention. The best intervention we know of so far is to drive up psychological safety in these areas.

2. The Fragility Check: If a person or a team experiences a high-pressure event do they return to normal after conditions return to normal? Or does their performance seem to change permanently? A system that doesn’t return to baseline is a system that has “bifurcated.” Bifurcation is further evidence of a chaotic system and could be an alternative to the twin test if a role is too unique to provide any statistical twins.

5. The Prediction Horizon: Determine your organization’s “Event Horizon.” Can you predict performance 1 week from now, 3 months? 6 months? Because chaotic systems are not random, predictions tend to be better the closer they are in time. Look for factors that lead to longer prediction horizons. These could be important tools to add to your toolbox of interventions.

We are not managing automatons; we are managing teams of people. People working in teams have a tremendous impact on each other. Recognizing and embracing performance as the product of a nonlinear system will lead us to better measurement and intervention.

Our regression models are powerful tools, but only if we operate in regimes where they are valid approximations of reality. Understanding where we have high λ, and actively managing it, is how we make those models useful. Here are three types of measurement that could help us understand the physics of our organizations:

1. The Twin Test: Track “statistical twins” (hires with similar propensity scores). Do their paths stay parallel (Low Chaos), or do they diverge wildly (High Chaos)? This is the best direct proxy for λ. Roles with high divergence deserve immediate attention.

2. The Fragility Check: If a person or team experiences a high-pressure event, do they return to baseline after conditions normalize? Or does performance change permanently? A system that doesn’t return to baseline has “bifurcated,” another hallmark of chaos.

3. The Prediction Horizon: Determine your organization’s prediction Horizon. Can you predict performance 1 week out? 3 months? 6 months? In chaotic systems, prediction accuracy decays exponentially over time. Knowing your horizon will save you from confusing random guesses for predictions.

AI Won’t Save Us

Right now, there is a fever pitch in HR Technology. Vendors promise that with enough data and powerful enough AI, we can finally predict high performers. They claim Large Language Models (LLMs) will “reason out” patterns invisible to humans.

However, if individual performance is the result of a chaotic process, these claims are mathematically impossible for the long term.

AI is a statistical tool, not a crystal ball. It works by projecting historical patterns forward. In a Linear or Random system, this works beautifully. But in a Chaotic system, history is a poor predictor of the future.

The limitation isn’t computing power; it’s the topology of the system. To double your prediction window in a chaotic system, you need exponentially more precise data, not just more data. To predict a career path with certainty, you would need to know the initial conditions with infinite precision, down to the firing of a neuron. Since we can’t do that, long-term AI prediction models will always collapse into “hallucinations” or generic guessing.

Does this mean AI is useless in People Analytics? Absolutely not. It means we need to flip the value proposition.

Using AI as an Oracle (to predict the fate of an individual) is a waste of resources. We should instead use it as a Cartographer (to map the attractors and invisible currents).

AI may prove to be exceptionally good at this task. In addition to the measures of chaos suggested above we should consider using AI for the following:

• Measuring Drag (z): AI can analyze communication metadata to map “Bureaucratic Heat Maps.” Where does work slow down? Where does complexity spike?

• Mapping Currents: By analyzing network traffic (ONA), AI can show us the “Eddies” of safety and the “Whirlpools” of isolation.

• Detecting Bifurcation: AI can monitor shifts to warn us when a team is approaching the “Critical Threshold”, the moment before the system snaps from stable to chaotic.

The goal is not to know what the employee will do, it is to understand and engineer the river.

Share

Share Tyler Weeks

Leave a comment

Appendix: The Rigorous Math (For the Nerds)

For those interested in the formal dynamical systems theory behind this argument, we can model organizational performance using a variation of the Lorenz Attractor. The work below was supported by Gemini Pro and is not provided as a proof of chaotic behavior, simply as a demonstration that under certain assumptions human performance could be described using the physics of chaos.

We can define the state of an employee at time t as a vector v(t) ∈ R³:

x(t): Performance Output (Rate of work)

y(t): Engagement/Potential (Internal energy)

z(t): Organizational Drag (Friction/Entropy)

The time evolution of these variables is governed by the following coupled non-linear differential equations:

The Terms:

• σ (Agility): The rate at which potential (y) converts to performance (x).

• ρ (Pressure): The external incentive or drive applied to the system.

• β (Hygiene): The rate at which the organization dissipates bureaucratic drag (z).

The Proof of Chaos: The critical term is -xz in the second equation. This represents Burnout Coupling. As Drag (z) increases, high Performance (x) actively destroys Engagement (y).

Stability analysis shows that when Pressure (ρ) exceeds a critical threshold ρ_c, the fixed point of the system becomes unstable (via a Hopf Bifurcation).

Beyond this threshold, the system enters a chaotic regime. In this state, the distance between two initially close trajectories (δZ0) grows exponentially according to:

Where λ is the positive Lyapunov exponent. This demonstrates that in high-pressure, high-drag environments, long-term prediction of individual performance could be mathematically impossible, regardless of data quality.

In an optics lab, your first lesson is a humbling one: your measurement is almost never what you think it is. I spent years calibrating sensors to detect the faintest signals in nonlinear optics, only to realize that an “exciting new phenomenon” was often just the 60Hz hum of the laboratory’s electrical outlets. You learn to distrust the device, the environment, and especially your own observation.

Now, overseeing People Analytics for a global workforce of hundreds of thousands, I find the measurement traps are the same. Organizations spend billions of dollars and millions of hours on the annual ritual of performance ratings, often treating a 1-to-5 scale as a high-fidelity instrument. But if we apply the first principles of information theory, we discover that we aren’t using a high-resolution “microscope”; we are using a blunt sensor with a hard mathematical ceiling that filters out the very “star performance” we are trying to find.

The most important thing to know about performance ratings is that they are a VERY low resolution measurement. That doesn’t mean they’re a waste of time, it just means we have to be careful about the conclusions we draw based on an extremely fuzzy image of the workforce.

1. A Gedankenexperiment

To illustrate just how fuzzy these ratings are, let’s use a much easier thing to measure directly than performance time. Imagine we have a completely remote workforce and we’re trying to figure out when people are starting their day. We’ve decided the simplest way to do this is to ask their managers what time each person is starting. But… here’s the catch… for simplicity instead of asking them to record the time we tell them they can only use the numbers 1, 2, 3, 4, or 5.

After communicating the new effort to all managers we begin to receive the data and try to reconstruct the work habits of our company. It doesn’t take long before we realize that we have a problem:

• The Rater Effect: Manager A uses a standard 12-hour analog wall clock and assigns a “1” for anyone starting before 2:24 AM. Manager B uses military time so assigns a “1” for anyone starting before 4:48am. Manager C uses “1” for the first 20% of their team, regardless of how early or late they started.

• Data Destruction: We look at each of our 2s and realize we cannot reliably make any inference about what time any individual started. That information was erased when the manager converted it to a single integer.

• The “Artifact” Curve: As we look at the data we think we see patterns and decide to calculate some statistics based on these 1-5 time ratings. We find that we have what looks like the old familiar Bell Curve (Normal distribution) centered at about 3. The problem is that this curve doesn’t represent when people start work; it is a mathematical artifact created by averaging out the different ways managers used the numbers.

What decisions can we make based on this data? Would it make sense to recognize all the 1s and 2s for having healthy work habits? What time does 2.43 represent?

This little thought experiment maps almost perfectly onto the real-world challenge with performance ratings. Scullen, Mount, and Goffin (2000) found that when using a 5-point scale to measure performance, the idiosyncrasies of the raters accounted for nearly two thirds of the variance. If more than 60% of the variance is the rater, then the “signal” we are getting is a social one. The simple subjective scale is telling us more about the managers (how they used the scale) than about the employees they are measuring.

If the challenge was simply the idiosyncrasies of the raters the solution would be simple but it goes deeper than that. Using 5 digits to represent the 1440 minutes of the day erases information we can never get back.

2. The Information Bottleneck

Now let’s imagine that after collecting all of the time “measurements” from our managers we’ve realized our mistake and we give them very specific definitions for each number on our 5-point scale. We further give them a standard procedure to validate the number they’ve assigned. To make sure we get the most accurate number we implement a “360” measurement that asks each employees’ peers and direct reports to also submit a number using the same criteria. We’ve done everything perfectly to ensure we have the most robust categorization of every employee into 1 of 5 categories. Do we now have a good measurement?

While we can be more confident in the numbers we have, the scale itself introduces a significant limit on our ability to produce a high-resolution picture of what is actually happening. Based on Information Theory, a standard 5-point scale has a theoretical maximum information content (entropy) of just 2.32 bits. To put that in perspective, a single emoji contains more data than the typical performance rating.

The natural inclination may be to simply use more numbers. Wouldn’t a 100-point scale give us finer resolution?

Unfortunately more categories won’t necessarily translate into a better measurement. Based on George Miller’s “Magical Number Seven,” the human brain only has about a 2.8-bit capacity to process categorical information. Even if you gave a manager a 100-point scale, they would instinctively “bin” the data into about seven categories. As a measurement device, the human brain is physically incapable of maintaining the resolution required to distinguish between a “92” and a “93” in human performance.

These limitations on the information we are able to capture with a rating scale only compound when applied to complex phenomena. It’s much like asking someone to categorize a car as a circle, a square, or a triangle.

3. The Power Law

To understand just how inadequate humans are as measurement devices for our own performance we have to take a step back and understand what we’re trying to measure.

Most human traits, like height and weight, follow a Normal Distribution (the Bell Curve). As humans our intuition is built around a “Normal” world, the tallest person you know is probably less than twice as tall as the average person you know. The heaviest person you know is probably only about twice as heavy as the average person you know. A simple symmetric Normal distribution can be reasonably approximated with 5 to 7 categories, and often is. But not everything works like height and weight. Wealth, time spent in a car, and size of pokemon card collections follow a completely different curve called a Power Law distribution.

In a Power Law world, the far end of the curve isn’t 2x the average, it can be 1,000,000x. The median number of YouTube subscribers is 61, Mr. Beast has 460 million. If our goal was to identify and reward the most influential YouTube stars it would be very important to capture not only who they are but how much more influential they are than the typical content creator. To do this we need a measurement device that is up to the task. Because we can directly measure subscribers and views it’s easy, we can simply count using 32-bit integers.

If we tried to rate YouTube content creators based on the best our brains had to offer, a 7-point scale, we’d end up with a very distorted picture. In fact we’d end up with something that looked sort of like a normal distribution, reflecting the preferences of the raters, not the success of the YouTube channels.

Performance for many knowledge workers follows a Power Law distribution. One average developer might fix three bugs, while the best developer builds the architecture that saves the entire company. An average leader might keep their team happy, while the best leader will transform an industry. When we try to “measure” this performance using a scale and human judgement we distort reality beyond recognition.

So is it hopeless? Is this measurement worthless? Absolutely not.

4. Painting a More Realistic Picture

Abandoning performance ratings is not a good alternative. Attempts at “ratingless” performance reviews have resulted in very poor results including lower performance, shadow ratings, and a sense of aimlessness. The understanding that each person will be held accountable for their contributions has a tremendous effect on the productivity of the company.

However, performance ratings are rarely measures of performance. They are a tool for alignment, a shorthand for organizational consensus, and a “check-engine light” for engagement. We can get more out of this measurement by making the current approach as robust as possible and by not settling for the status quo.

Companies that have demonstrated more effective performance ratings have done so by directly addressing the low-resolution challenge. Here are a few techniques that will get the most out of your 2.32 bits:

1. Increase Sampling Frequency: Moving to continuous data collection to smooth out errors.

2. Measure Intent, Not Truth: Ask managers what they would do (pay, promote) rather than what the employee is (a “5”).

3. Aggregate Sensors: Using 360-degree data (despite its cost) to triangulate the signal.

4. Accept Subjectivity: Recognize that the rating is a communication tool for alignment, not a measurement of performance or output.

In order to challenge the status quo we need to experiment with new ways to capture more information about the output of the workforce. I’m not suggesting or recommending that we use AI to auto-magically produce ratings. I am suggesting that measurements like centrality, could give us a much more proximate, information-rich measure of the influence a person is exerting on the organization. AI-connected companies may be able to produce such measurements more cheaply and reliably than at any time in the past. The real value would be in the objective nature of the measurement and that it tends to follow a Power Law distribution, thereby mirroring what we’d expect from performance.

We are experiencing a watershed moment in workplace technology. Thinking about AI and agents as only tools for productivity misses the vast opportunities for observability and measurement. Our job will be to discover and validate those measurements and implement them with the highest standards of transparency so that we don’t create a high-precision dystopia.

Subscribe now

Share

Leave a comment

My favorite way to think about People Analytics is as a research and development team focused on understanding a specific company. Like any high-performing R&D team, a high-performing People Analytics team is highly interdisciplinary and takes advantage of a wide range of educational backgrounds. What makes such a group a coherent team isn’t a shared degree, skill set, or a shared tech stack; it is a shared obsession with figuring things out.

The unfortunate legacy of the past decade’s obsession with AI (Machine Learning and LLMs) is that People Analytics teams have been rewarded for ‘black box’ approaches. We have created a generation of practitioners who are better at getting a result that works than explaining why it works. The value of this flavor of People Analytics is actively eroding as LLMs become just as proficient at throwing spaghetti at the wall as humans.

As we move deeper into the AI era, the future of People Analytics will not be driven by Data Scientists, IO Psychologists, or Data Engineers who have mastered the syntax of their tools, but by experimentalists who master the art of the experiment. We need more scientists.

The Pursuit of Knowing

Some amount of the skill-gap we are about to experience is the result of an age old problem. Screenwriters and novelists have created a widespread misconception that scientists are people that know lots of things. This ubiquitous misconception has led to common oxymorons like “exact science” and even the misuse of the word scientist altogether. Very few data scientists seem to know anything about how to do science.

To set the record straight, scientists are not people that know lots of things; scientists are people that ask lots of things. They are people who care more deeply about the quality of an answer than having the answer.

This distinction between knowing and questioning is a lesson many of us learn the hard way. In Zen and the Art of Motorcycle Maintenance, Robert Pirsig indirectly describes a great scientist when he writes about the concept of “Quality”:

“Care and Quality are internal and external aspects of the same thing. A person who sees Quality and feels it as he works is a person who cares. A person who cares about what he sees and does is a person who’s bound to have some characteristic of Quality.”

In his own way Richard Feynman described this same person when we recounted the labor required to truly “know” something:

“I don’t know what’s the matter with people: they don’t learn by understanding; they learn by some other way—by rote or something. Their knowledge is so fragile! ... You have to work very hard to know if you know. You have to check it, and you have to think about it, and you have to see if it’s really true.”

Learning how to know something is a painful process of trial and error.

The Eigenvector and the Cardiologist

Years ago, before “Data Science” was a phrase anyone would recognize, I was a young researcher in the UC Davis Applied Science department, an applied physics program that had been founded by Edward Teller to build future generations of great experimentalists. The department was small but the faculty were brilliant and highly interdisciplinary. The lab I worked in was located on the medical campus where I was surrounded by MDs, biologists, chemists, and engineers who all knew much more than me about pretty much everything. I knew some math, and I knew how to run experiments.

In addition to my work in non-linear optical microscopy I would take on small side projects with the medical school faculty. At one point I was collaborating with a cardiologist to see if we could distinguish diseased heart tissue from healthy tissue at the cellular level using light scattering. He would provide the samples, while I would take the measurements on a microscope that our lab had built and provide the statistical analysis.

At the time, Principal Component Analysis (PCA) was the ‘it’ technique. It seemed like magic: you didn’t really have to know anything about your data, you just put it in and presto, you could easily classify groups with subtle differences. When I figured out, despite all the jargon, that it was just linear algebra I got very excited about using it for this problem with the heart cells.

I wrote a simple program to analyze my data and when I plotted the first two principle components I felt like a genius. The diseased cells and healthy cells showed up clear as day. I enthusiastically put together my results and scheduled time with the cardiologist to share the discovery. The next day I proudly sat in his office as I watched his initial excitement fade into confusion.

“This is wonderful,” he said. “But what does it mean?”

I launched into a technical explanation: “Well, you see, this axis is the first eigenvector of the covariance matrix, so it rotates the n-dimensional data so the direction with the most variance lines up with the x-axis…”

I trailed off as I saw his blank stare.

“Right,” he said. “That’s what you did. But what does it mean? Can you tell me anything about how the cells are different?”

I couldn’t. I had no idea.

In the end, that PCA analysis was just the first step in a series of experiments that produced a much more useful result. The final result was actually quite simple. Getting to a simple answer takes a tremendous amount of work.

This was one of my first hard lessons in the difference between getting an answer and knowing something. As the professional world grapples with the implications of LLMs we will only see this gap grow.

AI is the Lab, Not the Scientist

In the AI era, access to facts is a commodity. Hiring people that know things is less important than hiring people who know how to figure out things. The value of the experimentalists is rapidly escalating.

The LLM is not a scientist, it is a laboratory where your best experimentalists will thrive. It’s a room full of amazing tools, a playground for a curious mind. Confusion about the role of an LLM as a lab will leave you vulnerable to churning out enterprise-grade AI “slop”.

To take advantage of this amazing laboratory you will need amazing experimentalists and there won’t be any single degree that serves as a reliable proxy for this talent. While IO Psychologists and Data Scientists will continue to bring vital tools, you will need an equal mix of physicists, chemists, biologists, economists, and others who have mastered these experimentalist skills in unique contexts:

• First Principles Thinking: A good experimentalist doesn’t try to answer complex questions; they know it’s much easier to answer simple ones. They know how to break problems down into their fundamental parts and identify likely independent variables. If you ask them about turnover, they don’t immediately suggest a “Logistic Regression.” They start asking about the fundamental “forces” acting on an employee: incentive, competition, and aspiration.

• Experimental Design: Experimentalists know how to creatively test an idea using the simplest possible approach. They test the idea, not just the available data. When the problem is messy, this often requires some highly creative thinking to identify the linchpin assumption upon which a conclusion is built and to create a scenario that challenges or supports that assumption. They know that data in support of an assumption isn’t proof; it’s evidence.

• Creative Instrumentation: Rather than relying solely on the tools and data provided, a good experimentalist builds the tools to get the data they need. If the “sentiment data” isn’t good enough, they find a way to measure the network or the friction in a process, tinkering with different approaches until they get what they need. They aren’t afraid of making anything. They also aren’t afraid of repurposing tools they have in new and creative ways. What’s really the difference between a survey, an assessment, and a performance review, anyway?

• Mathematical Reasoning: Good experimentalists bring a deep intuition for the math behind their measurements. They know that statistical significance is not the same thing as a significant result and apply that understanding in their work. They can often give real-world examples to explain the math they are doing on the spot because they’ve tried and failed and tried again to implement their recommendations.

People Research & Development

By embracing a culture of R&D, AI will become a tremendous springboard for People Analytics. AI will be a high-powered, high-speed environment where our interdisciplinary teams can flourish.

In the end, the cardiologist’s question—“But what does it mean?”—will be the mantra that keeps us from descending into a collective AI-induced hallucination. The most important priority we can hold is the quality of our answers.

The People Analytics teams that become a strategic advantage for their companies will be willing to do the hard labor of checking, thinking, and verifying until they know.

Leave a comment

Share

In an era when scientists were arguing about whether atoms even existed, Boltzmann was busy proving how their invisible, chaotic movements created the physical world we touch and feel.

Carved into the stone above his name is a deceptively simple formula:

S = k ln(W)

People Analytics leaders should spend more time understanding this formula, it is the secret to why you can predict and direct the behavior of a 100,000-person organization but not your top engineer.

The “Big Picture” vs. “The Individual Story”

To understand Boltzmann’s insight, we have to look at the organization through two different lenses. In physics, these are called the Macrostate and the Microstate. For our purposes, let’s call them The Big Picture and The Individual Story.

1. The Big Picture (Macro-Conditions)

This is what we track in our dashboards. It’s the “temperature” of the company. It’s your global turnover rate, your average engagement score, or your quarterly productivity yield. These are stable, measurable, and often very hard to move.

2. The Individual Story (Micro-Details)

This is the reality on the ground. It is the exact, messy, fluctuating state of every single employee at 10:00 AM on a Tuesday. It’s who had a bad night’s sleep, who just got a headhunter’s call, and who is feeling energized by a new project.

What is “W”?

The magic of Boltzmann’s equation is the variable W.

In physics, W is “multiplicity.” In business terms, W is simply the number of different individual stories that can result in the same Big Picture.

Imagine you have an organization with a 10% annual turnover rate.

• The Individual Story: Why did “Sarah” leave? Maybe her significant-other took a new job. Why did “John” leave? He wanted a higher salary. Each story is unique, unpredictable, and seemingly random.

• The Big Picture: There are a staggering number of ways (W) to combine these individual exits to reach that 10% mark. You could have 10,000 “Sarahs” or 10,000 “Johns,” or any combination in between.

Because there are trillions of ways for an organization to behave “normally” (high W), the organization almost always behaves normally. The sheer volume of individual chaos creates the stability of the aggregate.

The “Noise” Makes it Work

In People Analytics, we often treat individual variation as something to be eliminated. Boltzmann teaches us that this noise, W, is why we can reliably measure the big picture.

Imagine a company with precisely 1 person. You could attempt to measure engagement, attrition, or any other organizational measure, but it would be meaningless. Engagement, for example, would fluctuate wildly within the same day. You would find that the moods of that one person were more closely related to their sleep patterns than anything happening at work.

The reason your 100,000-person organization is predictable isn’t that everyone is following the same script. It’s because they are all following different scripts. Their individual idiosyncrasies, personal crises, and sudden bursts of motivation cancel each other out at scale.

The larger the organization, the higher the W, and the more “statistically locked” the Big Picture becomes.

Stop Managing Atoms

As you come to appreciate the brilliance of S = k ln(W), it will change your strategy as a leader:

1. Don’t Manage the Micro-Details: Trying to predict exactly which high-performer will quit is like trying to manage a single oxygen molecule in a room. It is an expensive, high-error exercise in futility.

2. Adjust the "Container": If you want to change the "Big Picture" (e.g., lower your turnover), you don't do it by micromanaging individual stories; you change the boundary conditions of the system. You get your best results by changing the characteristics of the company rather than by attempting to change the characteristics of each employee. This is why economists focus so heavily on incentive structures when predicting the behavior of the masses.

When you change the compensation structure, the culture, or the technology stack, you are changing the characteristics of the company, the boundary conditions. You are shifting the entire statistical distribution.

We don’t need to control the individual to lead the compay. We need to understand the math of the crowd.

Leave a comment

Share

That’s not just a pithy truism but an important boundary condition, a critical factor that real solutions with real consequences embrace and a practical sniff test for any miracle cure.

The Problem with Assessments

In the past, I have been critical of psychometric assessments. At first glance, they appeared to me as self-congratulatory horoscopes with hardly more predictive power than your Hogwarts house. I’m sure many of my IO Psychology friends are clutching their pearls in indignation right now, but don’t worry, I realized that my skepticism was less about psychometric assessments than it was about how they were sold to me.

To explore my misunderstanding of assessments and what they can do for a company, we’ll run a simple simulation.

Consider a Monte Carlo simulation of 10,000 hires using an assessment with a correlation coefficient of r=0.20, an excellent result for any psychometric employment assessment. The plot below shows the results of that simulation with each hire plotted based on their assessment score and their job performance.

I’ve highlighted the top 10% of assessment scores in blue and plotted the model in red. The red line is what the vendors in this space will sell you; their assessment does indeed identify people who are more likely to perform well in their job. Notice the massive spread in actual job performance of those blue dots.

Statistically, if r = 0.2, as we see here, then the assessment explains only 4% of the variance in actual job performance (r² = 0.04). In other words, 96% of the time a hire’s performance will have nothing to do with their psychometric characteristics.

The chart below illustrates this relationship by plotting the same data as a set of histograms. The gray line represents the job performance of all the hires, while the blue line represents the job performance of candidates who scored in the top 10% of the assessment.

Notice the shaded gray area where the two distributions overlap. This area represents the candidates who scored “high” on the assessment but won’t perform any better than if you’d selected them without an assessment.

Even more discouraging is that despite being “top scorers,” roughly 36% of the highest scores on the assessment will be below average employees—that’s 1 in 3. It’s no wonder that your hiring managers likely don’t see the point of putting candidates through an assessment.

Organizational Gain

It can be helpful to think about assessments more like an actuary science. The job of an actuary is less about predicting if you will have a car accident this year than it is about how many people in a pool of 100,000 will have an accident this year. The predictability of groups is driven by a phenomenon called the Law of Large Numbers. In simple terms, the law of large numbers suggests that as you increase the number of people in a group, the “noise” of individuals cancels out, leaving behind a stable “signal.”

This reframing is important: we will always do a worse job at predicting the individual than we will be at predicting a large group.

Assessments are no exception. Their real value is in raising the floor of the group. Think of the value of an assessment in terms of organizational gain.

Organizational gain is formally captured by the Brogden-Cronbach-Gleser (BCG) utility model, which demonstrates that the total value of an assessment is a function of the validity coefficient, the variability of performance in the role, and the selection ratio. Importantly, even modest correlations yield significant dividends when applied at scale.

Try the following thought experiment:

• You know that you and your biggest competitor are both planning on hiring 1,000 leaders this year.

• You know that you are both hiring from the same general pool of candidates.

• Traditional interviews have been shown to have virtually no correlation with job performance.

• You have a magic tool to doesn’t tell you who will be a great leader but does ensure that you will onboard 40 more good leaders than the competitor.

• Do you use that tool?

• YES!

While 4% variance (R2=0.04) sounds small, it is a massive actuarial advantage when applied to thousands of hires. Think of it like a casino's edge: a house advantage of just 1-2% is what built the Las Vegas skyline. By using a validated assessment, you aren't just guessing better; you are systematically shifting your organization’s win rate so that you onboard dozens more high-performers than your competition every single year, simply by virtue of playing the averages at scale.

Assessments are proof that despite a hiring manager’s proclivity for making bad hiring decisions and a bad actor’s talent for deception, you can implement a systematic approach that can improve your results by about 4%.

The choice to invest in an assessment is easy, but not because you know who is going to be a great hire; it’s because you are giving your company a statistical advantage, organizational gain.

Realistic Consideration

The important factor for achieving organizational gain is scale. Identifying where the law of large numbers will kick in is your first clue as to where you should invest in assessments, consider eliminating expensive assessments, or refresh your narrative.

The limitations:

• More is Better: High volume (thousands per year) are where you’ll see the most gain from your assessments.

• The Criterion Problem: The more words it takes to describe success in a role, the less likely assessments are to predict anything.

• The Measure of Success: Assessments are validated based on performance in the job, not based on hiring decisions.

  • Validating an assessment against hiring decisions creates circular logic that simply predicts who a manager is likely to favor, rather than predict job performance.

The implications:

• High volume, highly tactical roles are where you should invest first.

• Individual hiring managers and small business units within a large company may never experience significant gain from assessments even if the company as a whole does.

• Assessments will likely provide less gain the more senior the leadership role. These roles tend to be small in number and the most poorly defined.

• Assessments are probably not worth the investment for small companies or for roles with less than a 100 hires per year. You will likely not have the scale necessary to realize the gain.

• Assessments are not a solution for the massive influx of applicants flooding employers today. The gain for your organization will be limited by the number of hires you make. If you are still only hiring 10 people a year to a role that is poorly defined you are not likely to see any benefit from using an assessment on thousands of applicants.

The Long Game

Tools like assessments can be a critical strategic advantage for a company of sufficiently large size. As People Analytics professionals it is up to us to help our leaders understand the nature of that advantage. We are in a position to help them apply this tool with surgical precision based on our understanding of the organization and its goals. We cannot let ourselves get confused about the difference between statistical significance and decision confidence. We must frame these tools as a way to tilt the competitive landscape in our favor, one percentage point at a time.

By embracing the Law of Large Numbers we can begin to build an organization that is engineered to win.

Subscribe now

Share

Share Tyler Weeks

Leave a comment

Most often we think of the companies as sets of boxes, hierarchies with the CEO at the top, working its way down through layers of leaders, managers, and individual contributors. That organizational chart helps us see how decisions are structured, but it is only one way to understand how a company works. Like a human body, a company has many systems that overlap and interact to make it run. There’s the financial view, organized as cost and revenue centers that describe how money flows. There’s also the project view, which shows how cross-functional teams are organized to execute initiatives. But the most elusive system, and arguably the most predictive of a company’s performance, is its social structure [1].

This structure isn’t formalized anywhere; it’s not an official hierarchy, nor can it be easily documented or managed. It is simply who you talk to, rely on, build friendships with, and turn to when you have questions or need advice. Tiffany McDowell, in ‘Strategies for Organization Design,’ demonstrates that while we focus 100% of our design efforts on the formal hierarchy, that hierarchy only accounts for about 20% of how work actually gets done. The other 80% of value, including critical decision-making and information flow, occurs in the informal network [2]. Leading People Analytics teams invest tremendous effort trying to understand this structure through what’s called an organizational network analysis. These analyses reveal that the most important people in a company are the ones that promote serendipitous interactions between colleagues. They are the super-connectors who bring people together, pass information, and act as hubs. Sometimes they are a leader, but sometimes they are Debbie from finance, who keeps a candy dish on her desk. People with the weakest connections to this social structure are the most likely to feel negatively about their company, be seen as lower-performing, and ultimately leave. That may sound obvious, but it can often be hard to identify who those people are.

This social fabric is so delicate that the layout of an office building can impact the creativity of its teams. In the 1970s, Thomas Allen developed what is now known as the “Allen Curve,” showing that communication drops off with physical distance [3]. He showed that putting teams on separate floors of the same building could harm their collaboration as much as putting them in different cities. His work led to corporate campuses designed with atriums and open spaces to foster serendipitous “water cooler” moments meant to spark new ideas. Now, in a world of hybrid work and powerful AI, we are experiencing an emerging new kind of campus that lives partially in the cloud. Arena, Hines, and Golden characterized this new “campus” using organizational network analysis and showed that, like Allen’s architectural recommendations, companies could identify and address the stifling effects of employee isolation [3].

The most creative and productive teams work in environments where they not only share information efficiently but also have strong social interactions. This isn’t just about feeling like you belong; random, serendipitous interactions also fuel creativity. You may not realize it, but the casual conversations in the cafeteria or the comments you overhear in the hallway accumulate, resulting in those light-bulb moments hours later when you’re in the shower. As much as work-from-home proponents point out that people are more efficient at specific tasks without the “distractions” of a busy office, there is good evidence to show that if you need your team to be creative, there’s no replacement for whiteboards and watercoolers. The more we talk to each other face-to-face, the better we coordinate and create [3].

So what does it mean for a company when employees discover the convenience of large language models? When they realize they can get ideas, opinions, and help drafting important emails much more efficiently from an AI than from any of their colleagues? When, without thinking, they turn to an AI with their question instead of a person? When the bulk of an employee’s interactions are already online, what happens to the social fabric of a company when Slack feels cluttered and confusing compared to asking an AI for help?

The Hub-and-Spoke Trap

As an enterprise AI becomes more efficient and appealing, it will naturally centralize information flow. This isn’t a new phenomenon but the radical acceleration of a trend that began with tools like Google Search. For decades, organizations functioned on a “transactive memory” system—a collective awareness of “who knows what.” As researcher Daniel Wegner described, to get an answer, you had to interact with a person, reinforcing the social fabric with every query [4]. The rise of search technology introduced a phenomenon known as the “Google Effect”, in which search engines began to replace humans in our transactive memory [5]. Instead of asking a colleague, we started searching keywords. This shift, while efficient, caused a subtle degradation of the interpersonal pathways that formed our collaborative culture. LLMs are only accelerating this evolution—capable of not just recalling information but synthesizing it, making the bypass of human colleagues almost irresistible.

To understand the potential impact of an LLM on how we experience work we can simulate workplace collaboration using an agent-based model. In this simulation each employee can choose between talking to a person or an AI. Each time they make a choice we can look at the strength of their connection to all the other employees in the simulation. In this way we can explore the competition between the pull of AI Efficiency (the quality of the information it provides) and Human-to-Human Synergy (the value of the human interactions). What becomes immediately apparent is that if the perceived value from interacting with other humans does not outweigh the value of getting a quick answer, a tipping point is reached where people are more likely to turn to the AI than a peer.

This simulation visualizes the “substitution effect” of Generative AI on workplace topology. The black line tracks the density of peer-to-peer connections (strong ties) as the AI tool becomes more efficient.

Stage 1 (Distributed Mesh): When AI is less convenient than asking a colleague (<0.8), the organization remains a resilient, interconnected mesh.

The Tipping Point: Once the AI crosses a critical threshold of convenience, a feedback loop is triggered. Employees bypass colleagues for the bot, causing social bonds to decay. This raises the “social friction” of future interactions, making the AI even more attractive.

Stage 2 (Centralized Hub): The network collapses into a “Hub-and-Spoke” structure. While individual productivity may rise, the organization loses the lateral connectivity required for complex, tacit problem-solving.

Simulation based on a stochastic agent-based model (N=400). The “Tacit Floor” represents the minimum viable network required for non-codifiable tasks.

For fully distributed or remote teams, where the primary interaction with colleagues is through digital channels and focused on information exchange, this tipping point could represent a social collapse. The organizational network of such a company would look less like a network and more like a hub with spokes, where the AI is the hub and each employee sits alone on their spoke. Based on Allen’s model, this represents disaster. Imagine an office building with exactly one employee per floor. Each person might be highly efficient with very few distractions, but their opportunity to serendipitously bump into a colleague is precisely zero.

Mark Granovetter’s work on the “strength of weak ties“ showed, novel information and opportunities often come from infrequent, arms-length connections—not our close collaborators [6]. Said differently, highly creative companies are highly connected companies. Conversely, the severing of those weak ties within a company would result in a significant drop in creative output. A highly effective AI could compromise weak ties by soaking up a critical mass of employee queries. This creative collapse isn’t without parallels and could be described as a sort of corporate “filter bubble,” a term coined by Eli Pariser [7]. Without meaningful human-to-human interactions, a highly effective LLM would show employees what is most relevant to their questions and their current role, isolating them in a bubble of convenient information. The very cross-pollination of ideas we designed our buildings to create would be challenged by the software we’ve created to make work more efficient.

This internal corporate dynamic is a microcosm of a larger societal shift. In his seminal work Bowling Alone, political scientist Robert Putnam documented the decline of “social capital”—the networks, norms, and trust that enable a society to function. As we have replaced community leagues and local clubs with isolated, on-demand entertainment, our collective bonds have frayed [8].

The shift to an AI-dominated hub-and-spoke structure within a company represents a corporate version of “bowling alone.” By optimizing for the friction-free efficiency of the individual, we risk eroding the collective social capital of the organization, leaving it less resilient, less innovative, and ultimately, a less creative place to work.

Avoiding Creative Collapse Through Human Connection

Is creative collapse inevitable? Absolutely not. There is a future where we foster a workplace that is more efficient, creative, and fulfilling. This future will be dependent on our ability to promote the highest quality human-to-human interactions at work.

Using sociometric badges, Alex Pentland’s Human Dynamics Lab observed that the most significant predictor of a team’s success was not intelligence or skill, but the pattern, density, and quality of their real-time, face-to-face communication [9]. He suggests that these patterns of communication can account for 30%-50% of a team’s performance. “Collective intelligence” is the innovation output that fuels human collaboration. A company that thoughtfully uses technology to foster meaningful human interaction could enter a virtuous cycle of creativity rather than spiral into creative collapse.

When people collaborate on complex problems, they don’t just solve the task at hand; they build trust, share tacit knowledge, and generate novel ideas. Innovation is an inherently social process, invention rarely springs from an isolated genius but from a connected community of thinkers.

AI as a Super-Broker

The future is not a binary choice between human connection and AI efficiency. The tipping point is not a cliff we must avoid but a dynamic frontier we must manage. For People Analytics and People Technology teams this represents a call to action, a shift in how we measure success and design people-centric digital environments. We must actively monitor the impact of these tools on the company’s social structure and pay just as much attention to what they are doing to the people as to what they are doing to their output [10].

The most important shift we must make is to redefine the AI’s job description. Instead of being a simple hub for answers, the AI becomes a “Serendipity Engine”. Instead of narcissistically positioning itself as the expert, it could respond with, “Here is the information you asked for, but you should really talk to Sarah in marketing. She is the expert on this.” This means fostering what Tomas Chamorro-Premuzic identifies as the truly irreplaceable human qualities: curiosity, emotional intelligence, and adaptability [11]. In this model the AI acts as a connection broker.

An agent-based simulation of a 200-person organization over a 40-week period was used to move beyond speculation. The baseline projection (the dashed dark orange line) confirms the “substitution risk”: as AI becomes a frictionless oracle, employees naturally bypass their peers, causing the organizational network to decay. But this collapse is not inevitable. Four distinct interventions, or “Knobs”, were modeled that could be used to reverse this trend. These include AI Design choices (like a “Gatekeeper” function that deliberately refers complex questions to human experts) to HR Policies (work-from-office requirements). The results, visualized below, offer a mathematical blueprint for resilience: they demonstrate that by combining physical proximity with socially-aware AI, we can achieve the efficiency gains of the technology without sacrificing the structural integrity of the human network.

To prevent organizational collapse, the technology and leadership will need to pull multiple levers. The dashed orange line represents the “Default Decay” caused by high-efficiency AI. The colored lines show the impact of specific interventions:

• Knob 1: The Gatekeeper (green): The AI deliberately refuses to answer 30% of context-heavy questions, forcing human-to-human interaction.

• Knob 2: Warm Intros (dark blue): The AI uses social proof (”Mike trusts Sarah”) to increase the acceptance rate of referrals.

• Knob 3: Double-Sided Nudge (yellow): The AI primes both the seeker and the expert, doubling the relationship value of the interaction.

• Knob 4: Work From Office (light blue): Physical proximity (The Allen Curve) reduces the friction of asking a neighbor, lowering reliance on AI for local problems.

• All Interventions (Orange): By combining policy (WFO) with “Social Physics” design, the organization maintains a network density that could support both a highly creative and highly efficient workforce.

Measure What Truly Matters. This ideal state will not be achieved by default. We must use Organizational Network Analyses (ONAs) more often to monitor the creative health of the company. As Rob Cross’s research has shown, identifying and empowering the hidden “brokers” and “connectors” in your organization is key to resilience and innovation [12]. Going forward, we must include the enterprise AI super-broker and other tools in that network analysis to understand and manage their impact on the company’s social health. This represents a new body of work. Typically software is treated solely as a medium of communication. Building an environment of resilient creativity will require that we understand technology’s role as both a collaborator and a broker and use that understanding to guide organizational design, site strategy, tech strategy, and work-from-office policies.

Work-From-Office. We should be pushing for more people in the office more often and designing an experience that makes it worth their time. The imperative for teams that must innovate is not about accountability, it’s about creativity and team health. When it comes to fostering highly productive, creative, and healthy teams there’s no replacement for in-person interaction. While we may be able to reduce the impacts of hyper-efficient AI on a remote workforce through technological interventions, there is no panacea for the reduction in collaboration that inevitably follows geographical distance.

We have an opportunity to foster stronger, more creative teams than ever before, but it won’t happen on its own. If we do nothing, we will end up with AI tools that serve the individual task, creating a world of hyper-efficient, isolated spokes. Or we can take advantage of this moment to create a whole new Allen Curve, an understanding of the modern “campus” built on the knowledge that the health of the connections between us is the greatest predictor of our long-term success.

Subscribe now

Share

Leave a comment

Works Cited

1. Michael Arena, Scott Hines, John Golden III. The three Cs for cultivating organizational culture in a hybrid world. Organizational Dynamics Volume 52, Issue 1, 2023

2. McDowell, T. (2023). Strategies for organization design: Using the Peopletecture model to improve collaboration and performance. John Wiley & Sons.

3. Allen, Thomas J., and Gunter Henn. The Organization and Architecture of Innovation. Elsevier/Architectural Press, 2007.

4. Wegner, Daniel M. “Transactive Memory: A Contemporary Analysis of the Group Mind.” In Theories of Group Behavior, edited by B. Mullen and G. R. Goethals, Springer, 1987, pp. 185-208.

5. Sparrow, B., Liu, J., & Wegner, D. M. (2011). Google Effects on Memory: Cognitive Consequences of Having Information at Our Fingertips. Science, 333(6043), 776–778.

6. Granovetter, Mark S. “The Strength of Weak Ties.” American Journal of Sociology, vol. 78, no. 6, 1973, pp. 1360–80.

7. Pariser, Eli. The Filter Bubble: What the Internet Is Hiding from You. Penguin UK, 2011.

8. Putnam, Robert D. Bowling Alone: The Collapse and Revival of American Community. Simon & Schuster, 2000.

9. Pentland, Alex. Social Physics: How Good Ideas Spread—The Lessons from a New Science. Penguin Press, 2014.

10. Hao-Ping (Hank) Lee, Advait Sarkar, Lev Tankelevitch, Ian Drosos, Sean Rintel, Richard Banks, and Nicholas Wilson. The Impact of Generative AI on Critical Thinking: Self-Reported Reductions in Cognitive Effort and Confidence Effects From a Survey of Knowledge Workers. In Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems (CHI ‘25).

11. Chamorro-Premuzic, Tomas. I, Human: AI, Automation, and the Quest to Reclaim What Makes Us Unique. Harvard Business Review Press, 2023.

12. Cross, Rob, and Andrew Parker. The Hidden Power of Social Networks: Understanding How Work Really Gets Done in Organizations. Harvard Business School Press, 2004.

Young Feynman noticed a bird and asked his father what it was. His father replied, “You can know the name of that bird in all the languages of the world, but when you’re finished, you’ll know absolutely nothing whatever about the bird. You’ll only know about humans in different places, and what they call the bird. So let’s look at the bird and see what it’s doing—that’s what counts. I learned very early the difference between knowing the name of something and knowing something.”

Feynman’s point was simple: Do not confuse the name of a thing with the understanding of a thing.

I’ve been thinking about this distinction a lot lately, specifically when I scroll through LinkedIn.

Every week, there seems to be a new, viral diagnostic label for employee behavior.

• Quiet Quitting

• Job Hugging

• Rage Applying

• Coffee Badging

• Resenteeism

The response to these terms is often one of surprise, shock, or even panic. We act as if the viral trending of a phrase like “Quiet Quitting” is proof that it is real and new. The hype is followed quickly by enterprising individuals looking to monetize a buzzworthy panic. But this trending corporate jargon is like the birds Feynman saw when he was a kid. We are obsessed with the labels because the labels give us the illusion of control. If we can name it, maybe we can manage it.

Looking at my LinkedIn feed, these buzzwords seemed to be proliferating. It made me wonder if the buzzwords represented a more important trend than any of the phenomena they claim to describe. Are we experiencing a wave of “Loud Labeling”? (yes, I’m aware of the irony).

Google search terms can be used to identify the most viral workplace terminology from 2010 to the present. Combined with who invented the term, that analysis reveals a tectonic shift in how we talk about work that may be more important than the buzzwords themselves.

The Optimization (2010–2019)

If you look at the dominant business terms from 2010 to 2019, you’ll notice a pattern.

• 2010: Gamification

• 2012: Big Data

• 2013: Lean In

• 2015: Holacracy

• 2017: Grit / Deep Work

Where did these terms come from? They were invented by CEOs, COOs, management consultants, and academics.

What was the goal? Optimization. Before 2020 the most popular corporate jargon was largely prescriptive, top-down instructions on how to be a better, faster, more efficient unit of production. Lean In told women the problem was their lack of assertiveness. Grit told us the problem was our lack of stamina. Gamification tried to trick our dopamine receptors into enjoying data entry.

For a decade, the business leaders controlled the narrative. They defined what “good work” looked like.

The Fracture (2020)

Then, obviously, the world broke.

The pandemic didn’t just change where we worked; it broke how we talked about work. The top-down narrative went silent. And into that silence, the workforce stepped in.

The Resistance (2021–Present)

Look at the viral terms that have dominated since 2021:

• 2021: The Great Resignation

• 2021: Anti-Work

• 2022: Quiet Quitting

• 2023: Lazy Girl Jobs / Act Your Wage

• 2024: Coffee Badging

Who invented these terms? They weren’t coined by Malcolm Gladwell or Sheryl Sandberg. They were coined by a burned-out engineer on TikTok. A freelancer on Reddit. A Gen Z employee venting into a webcam.

These terms are not prescriptive; they are descriptive. They aren’t trying to optimize output; they are raging against the machine.

The Semantic Shift

We have moved from an online environment where management set the tone (Top-Down Optimization) to one in which the workforce set the tone (Bottom-Up Resistance).

In 2013, the advice was “Lean In.” In 2023, the advice was “Bare Minimum Mondays.”

Like Feynman’s birds the words we use tell us more about the people using them than the bird itself. The danger for leaders lies in chasing the viral terms rather than hunting down the root cause of this shift in tone.

This explosion of “Loud Labeling” isn’t merely a sign of a new reality; it’s a profound challenge to leadership. Workers will always have legitimate reasons to express dissatisfaction. While it is unrealistic to expect a workplace free of all grievances, the danger lies in the amplification of negative themes. Obsessively pathologizing every trending term reinforces and ultimately generates a pervasive, negative sentiment across the workforce.

A future dominated by Bottom-Up Resistance will continue to suppress motivation and satisfaction. The workplace needs leaders committed to building highly productive companies, who will actively seize the pen and win back the narrative. This requires more than just new buzzwords or corporate rebranding; it demands a visible, meaningful commitment to inspire, connect, and reward people. With a workforce gripped by loneliness and negativity, I believe the most important thing a leader can do is create an environment that promotes the most meaningful collaboration. Only by defining a new, positive contract for work can leaders reverse the collective drain of our productive energy.

Subscribe now

Share

Leave a comment

Historical data would suggest, rather, that we are experiencing a hiring market correction that’s been accelerated by AI. From roughly 2002 to 2015 technology temporarily lowered the friction of application without overwhelming employers. That window created a false baseline for the effort that has historically been required of job seekers.

The 125-Year View: Quantifying the Anomaly

If we look at how people got jobs over the last century, it becomes immediately apparent that applying to an entity where one has no social capital, a “blind” application, has almost never been effective.

In 1974, sociologist Mark Granovetter published his seminal work, Getting a Job. His data revealed that 56% of professionals found employment through personal contacts, while only 18.8% succeeded through formal mechanisms like direct applications or advertisements.

Then, the internet temporarily disrupted the job search.

As the chart above illustrates, we entered a unique period in the early 2000s. Digitization made job postings accessible, but the application process still required the effort of customizing resumes, drafting cover letters, navigating clunky portals. In this era, the act of applying served as a proof of interest and viability. Consequently, the efficacy of blind applications spiked, peaking around 2011 when sources like CareerXroads reported that job boards and career sites accounted for nearly 42% of hires.

For a decade, the “Stranger” had parity with the “Network.” This anomaly convinced a generation that the labor market was a meritocratic vending machine: input application, output interview.

Online Application Collapse

That window has now closed. The introduction of “One-Click Apply” features, followed by the recent proliferation of Generative AI, has reduced the marginal cost of an application to near zero.

Zero effort means more applications, more applications means a smaller chance that you’ll get the job.

In 2011, a corporate job posting might have attracted 50 applicants. Today, thanks to automation that can tailor and submit thousands of resumes on behalf of candidates, that same posting attracts 2,000. For the employer, the application has ceased to function as a filter for non-serious applicants. Job postings have become a noisy mob clamoring for attention. To manage this noise, organizations have erected defensive walls, including stringent filters and screening processes, resulting in the sub-1% conversion rates, as reported by benchmarking platforms like Lever and Ashby in 2024.

We have returned to the 1974 baseline. The most effective way to get picked out of the mob is to be referred. The market has corrected.

Back to Basics: The Social Solution

If blasting out online applications is not working, candidates must return to the mechanism that has facilitated human cooperation for millennia: Trust.

Social media has distorted our understanding of what “social” means and who our network includes, it has almost nothing to do with likes and follows. To navigate this new/old job market it may be helpful to refer to Granovetter’s distinction between Strong Ties and Weak Ties.

• Strong Ties are your close friends, family, and immediate colleagues. While vital for emotional resilience, they are structurally inefficient for job searching. Strong ties possess redundant information; they know the people you know and see the opportunities you see. They are an echo chamber.

• Weak Ties are your acquaintances and professional contacts. This group includes former colleague you haven’t spoken to in three years, the peer you see at the conference every year, and the person you volunteer with. These are individuals with whom you have some rapport and who have a different set of strong ties than you.

Granovetter’s breakthrough was proving that Weak Ties are statistically more significant in finding employment than Strong Ties because they serve as bridges to information silos you cannot access.

The tragedy of the “Easy Apply” era is that it tricked us into thinking a LinkedIn connection is a Weak Tie. It is not. A digital connection without shared experience is simply a stat. A Weak Tie requires a foundation of experiential trust, however faint.

Social Health as a Competency

In his analysis of American community, Bowling Alone, Robert Putnam detailed the erosion of strong and weak ties in the places we live and work. We have traded deep, localized, high-value relationships for broad, shallow, low-friction digital ones. In the current hiring market, this trade has left us vulnerable.

Social health is distinct from networking. It is the active maintenance of an enduring social infrastructure for your life. People who nurture good social health will be better equipped to navigate the job market.

The Trap of Comfort

When I’ve suggested that a candidate should not expect success without tapping into their network, I’ve received a range of reactions that follow a common theme: “But the younger generation is digital-first,” or “It’s unfair to expect introverts to interact with people to get a job.”

This reaction only affirms that we are living through a crisis of connection; for many, the atrophy of social muscles makes the prospect of relationship-building daunting. The easy application was seductive precisely because it promised career advancement with little effort and without awkward small talk.

However, the market does not grade on a curve based on our comfort zones.

Relationship building is a professional competency, no different than data analysis, coding, or financial literacy. Your social health is the sum total of the healthy relationships you maintain, the strong and weak ties.

If a role required proficiency in Python, and a candidate said, “I’m not comfortable with coding,” we would advise them to learn the skill. We would not expect the market to change its requirements. Similarly, the ability to foster trust, re-activate dormant ties, and engage in non-transactional dialogue is a learnable, practice-able skill.

Lest you think extroverts have a corner on the job market, social health is not built by “working a room”, it is built through deep listening, curiosity, and one-on-one interaction. This next job market may actually go to the introverts.

The Conclusion

The frustration candidates feel is real, but it is misdirected. They are mourning the loss of a lottery ticket that used to offer 1-in-50 odds and now offers 1-in-5,000.

The “Easy Apply” era was a historical blip—a momentary glitch where technology outpaced our ability to filter it. That glitch has been patched. The future of hiring looks remarkably like the past: human, trust-based, and driven by those who can navigate the social complexity of the real world.

In an AI-flooded world, your ability to look someone in the eye, shake their hand, and build genuine rapport is not merely a “soft skill.” It is the ultimate competitive moat.

Subscribe now

Share

Leave a comment

Despite collective investments of billions in skills taxonomies and AI-driven matching, the goal of a highly dynamic, productive, and egalitarian workforce remains elusive for most companies. While large consultancies tout significant value, it’s unclear if their success stems from superior methods or from the inherent priority of their business: the efficient allocation of resources for time-bound client projects. There is a distinct irony in the fact that the consultancies most vocal about skills-based hiring are also the ones selling skills-based transformations. For the rest of us, the promised benefits perpetually seem just out of reach. Why is this?

We are trying to solve a complex, adaptive systems problem with linear, industrial-era inventory logic. The effort isn’t wasted, but if we are going to revolutionize talent orchestration we need to account for some critical misconceptions about skills.

1. The Agency Gap: Humans Aren’t Inventory

The core assumption of the skills-based supply model is that if someone can do something they will. The possession of a skill does not equal the availability of labor.

In a warehouse, a spare part is available by definition. In an organization, a human being is only available to use a skill if they are willing to do so. This introduces the critical variable of volition and preference.

• The Problem: Your skills database might show a senior engineer is highly proficient in a legacy language like COBOL (their “Ability”), but they may have zero interest in using it, viewing it as career-limiting. If you assign them to that project, you risk immediate disengagement or attrition—what some now call “quiet quitting.” Your “supply” is largely illusory.

• The AMO Framework. We must move beyond the narrow focus on Ability (A) and embrace a Ability-Motivation-Opportunity (AMO) Framework from organizational psychology. Performance is multiplicative: P = A x M x O. If Motivation (M or Interest) is zero, performance is zero, no matter how high the skill score (A) is. The big work in your company is likely less about building new skills (A) and more about building interest in the work (M).

• The Next Generation: Stop building just Skills Inventories (”Can Do”). Start integrating Interest Inventories (”Will Do”) to map your true Volitional Supply Chain.

2. The Ecological Fallacy: Performance is a Team Sport

Skills-based models commit the ecological fallacy by atomizing performance. They assume that if you stack a team with individuals who have the highest density of required skills, you get maximum output. A gaggle of high performers does not make a highly productive team.

Decades of research prove this wrong.

• The Apollo Syndrome: Meredith Belbin’s work demonstrated that teams composed exclusively of high-skill individuals (”Apollo teams”) often perform worse because they lack cohesion, spend time in destructive debate, and don’t have complementary behavioral roles. High skills do not automatically equal high performance.

• The Primacy of Norms: Google’s Project Aristotle confirmed that who was on the team mattered far less than how the team worked together. The number one predictor of team success was Psychological Safety—the belief you won’t be punished for speaking up. Psychological safety is an emergent property of the group dynamic; it’s not a badge that can be earned or a skill that can be validated.

• The Next Generation: We must screen for Complementary Fit over Supplementary Fit. Hiring a person who adds a missing attribute (an “Implementer” to a team of “Generators”) may provide more value for the team than a person with the right list of skills. This is operational balancing, not bias. The focus must shift from the individual-job dyad to the individual-team dyad.

3. The Trust Deficit: Game Tape Trumps the Combine

Hiring managers are risk-minimizers, and they have a major trust gap with the skills-first model. A company may hire hundreds, if not thousands, of people a year, diffusing the perceived risk of each higher across the organization. An individual hiring manager may only hire one or two people a year, making each hire a high-stakes decision for their team. Those hiring managers are managing the perceived risk that a person may not be able to effectively do what they say they can do. A “verified skill” is a weak predictor of performance.

• Skill vs. Trust: A skill badge signals capability (”Can Do”). A Proven Track Record signals reliability and trust (”Will Do” consistently under pressure). A candidate with a “Project Management Certification” (Skill) is not the same as one who “Successfully navigated a $10M crisis” (Trust).

• The NFL Analogy: The NFL Combine tests discrete skills (40-yard dash, bench press). Game Tape shows how those skills are applied in the complexity of a real game. Over-indexing on skills is like drafting a player purely on Combine stats—you risk a “workout warrior” who can’t play the game. Experience and references are the corporate world’s Game Tape.

• The Next Generation: We must value experience not as a lazy proxy, but as a valid signal of “bundled competencies”—resilience, judgment, and political navigation—that no single skills assessment can capture. Until skill-based models can quantify “reliability” as effectively as “capability,” they will remain a niche tool for junior roles.

4. The Construct Validity Problem: The Illusion of Granularity

The skills-based organization is built on the premise that skills are discrete, measurable, and distinct units. This is a scientific illusion, a prime example of the Jangle Fallacy, the belief that two concepts are different because they have different names.

• Semantic Overlap: Psychometric studies, such as the one by Teyssier-Roberge et al. (2025), reveal a massive overlap in “21st Century Skills.” Terms like “Conflict Resolution,” “Negotiation,” “Collaboration,” and “Teamwork” are often measured by the same instruments and are semantically indistinguishable.

• The False Sense of Precision: Your skills database might have 5,000 unique skill tags, but they are likely masking a much smaller set of underlying constructs. This proliferation creates a noisy signal, leading to missed matches simply due to terminological mismatch.

• The Next Generation: Ditch the thousands of micro-skills. Focus on higher-order competencies like Interpersonal, Intrapersonal, and Goal-Directed/Executive Skills. These clusters are psychometrically more robust and align with how value is actually created in your company.

5. The Economic Problem: Potential Value vs. Realized Output

The final misalignment is economic. Many advocates of skills-based organizations push for Skills-Based Pay, they believe that compensation should be based on skills possessed (potential value). But the economic reality is that skills have no intrinsic value; only work (output) has value.

• Potential vs. Realized: A Python certification (Skill) held by an HR manager who never writes code is Potential Value. The functional code that launches a product (Output) is Realized Value. Paying for potential inflates the wage bill and incentivizes hoarding badges regardless of relevance (credentialism).

• Value-Centric Analysis: The unit of analysis should not be the Skill, but the Task or the outcome. We need to use frameworks like Jobs To Be Done (JTBD). Hiring a candidate means you are “hiring” them to get a specific job done (the outcome), not just to possess a list of features (the skills).

• The Next Generation: Performance management and compensation must anchor in Outputs/Tasks delivered. We need an Outcome-Centric Valuation model. The skills-based organization risks valuing the athlete for their height or speed rather than for the points they score.

The path forward is not to abandon skills, but to demote them from the primary organizing principle to a necessary input within a larger, more human-centric system.

A talent strategy designed to promote highly productive teams should be focused on competency and outcomes. These strategies will be characterized by:

• The “AMO” Inventory: Tracking Motivation and Opportunity with the same rigor as Ability.

• Team-Level Hiring: Assessing candidates for Complementary Fit and their contribution to team norms like Psychological Safety.

• Competency over Granularity: Simplifying taxonomies into robust Competency Clusters (Interpersonal, Executive, Technical).

• Output-Centric Valuation: Focusing compensation and planning on the Outputs and Tasks delivered, as this is where economic value is truly realized.

By building a Work-First model that acknowledges the volition, social complexity, and economic reality of human capital we can build a cohesive and productive workforce.

Leave a comment

Share

Subscribe now

I hate adoption metrics. Perhaps more than I should.

After the focus groups, design sessions, and communication drafts, focusing on “adoption” feels like the organization has missed the point. The software was never the goal.

It is forgivable to be confused about that. We spend a massive amount of time vetting and implementing systems, features, and platforms. But enterprise product teams are not in the software business. We are in the business business.

Businesses are people. When we talk about “product” in an enterprise context, we are talking about the way people work. We build products by connecting software to other software via integrations, and software to people via processes. To paraphrase a wise former manager: Our products aren’t functional until people are good at their jobs.

The great challenge of developing a fully functional product is that people are both users and components. People are not like software. Software does what you tell it; people seek the path of least resistance. People are messy, but they are natural optimizers. Once you embrace that variable, you’re ready to build amazing products.

When you stop fighting the human element and learn to love it, you find that enterprise product management is a world of paradox. The best leaders are simply the best at identifying and living within these contradictions.

Here are three crucial paradoxes I have encountered:

1. Everything & Nothing (The Paradox of Expertise)

It is true that the product owner should know everything. It is also true that the product owner knows nothing.

Your team must strive to be the expert. No one should know more about the motivations, frustrations, shortcuts, values, and workflows of the user base than the product owner. They should speak the native language of the function. If they are building a product for Talent Acquisition, a recruiter should hear them talk, wipe a tear from their eye and think, “Someone finally gets it.”

However, they can never learn enough. A high degree of intellectual curiosity should foster perpetual discomfort. The rules change, priorities shift, and the business goals evolve. Understanding is a process, not a destination.

2. Plan & React (The Paradox of Vision)

It is true that the product owner needs a durable vision. It is also true that if the product can solve a user’s problem, it should.

A product owner with no vision is just a “short-order cook,” passively taking feature requests from the loudest stakeholder. The result is a fragmented, confusing product that solves no one’s problem well. These products tend to require extensive documentation and come with a massive learning curve. Conversely, a product owner who ignores stakeholders ends up with an exquisite “ghost town” that solves theoretical problems rather than actual ones.

A great team embrace the paradox by translating the “what” and “why” from stakeholders into a cohesive “how.” They use their vision as a scaffold for stakeholder needs, not as a wall to block them out.

3. Perfect & Imperfect (The Paradox of Elegance)

It is true a product should be excellent in every way. It is also true that a product will not be excellent in every way.

When information friction is reduced to a background hum and work flows organically a great enterprise product can be magical. Our greatest wins are when it feels as if it obviously should have been this way all along.

We seek this on every project. Yet, teams often get cornered by corner cases. They obsess over perfection and end up with a complicated mess.

I once joined a call where the team was in a panic because they couldn’t figure out how the CEO would use the solution. The irony? The CEO was the only person in the company who would never use the tool. We decided to launch to the other 400,000 users and figure the CEO out later. That decision resulted in a more elegant solution because we didn’t design around an anomaly. We left a flaw that made the rest of the product better.

You will always have imperfections. The art is choosing the right imperfections.

The Goal: Optimization to Zero

Embracing these paradoxes is a balancing act that becomes intuitive with practice. But what is the ultimate goal? If not adoption metrics, what indicator will give you confidence you’ve delivered a great product?

We seek an experience that is self-sustaining. The products that have the greatest impact on your business are the ones that seem self-evident.

Consider the original Super Mario Bros. It came with no instruction manual. Every action required to win—running, jumping, stomping—was learned in the first 8 seconds of play through intuitive experimentation.

Or consider TurboTax. It took the US Tax Code—one of the most stressful, complex, and bureaucratic systems on earth—and made it so that most people get it “mostly right” most of the time. It doesn’t automate your taxes (who would trust that?); it breaks the work into digestible, intuitive steps. It holds the feeling of being overwhelmed at bay.

Neither Mario nor TurboTax requires a SmartSheet checklist, a PDF user guide on SharePoint, or a “train-the-trainer” session. They contain all the information required to be successful.

As we seek to make great enterprise products, the goal is not the most beautiful UI or the most users. The goal is the shortest instruction manual.

We aim for zero.

Getting to zero requires living in the tension of these paradoxes. You need the hubris to define a rigid vision and the humility to let reality reshape it. It is not an easy balance to strike. But when you get it right, the friction vanishes, the software effectively disappears, and your ‘users’ can finally get back to being people.

And that, ultimately, is the business we are actually in.

Subscribe now

Share

Leave a comment

Fans of the popular Dead Internet conspiracy theory believe some mastermind has done this to us; they’ve flipped a switch and deployed a nefarious legion of bots designed to subdue the masses. While there is evidence of increased bot activity online, blaming bots for this sterile landscape of algorithmically curated content loops might be misplaced to begin with.

The Attack of the Bots

There is good evidence for a machine-dominated internet.

For the first time in history, automated traffic has surpassed human activity online. As of this year, bots account for over 51% of all internet traffic. Think about that. The majority of “activity” on the network we built for human connection is now non-human.

There is a flood of synthetic personas and automated content overwhelming our digital spaces.

• On Meta’s platforms, there are roughly 430 million “inauthentic” duplicate and false accounts, accounting for about 14% of monthly active users.

• Instagram alone has an estimated 95 million fake accounts, run by bots for deception and manipulation.

• The activity on these accounts does explain some bizarre anomalies, such as the “Shrimp Jesus” images on Facebook, accompanied by thousands of automated “Amen” comments.

While these numbers paint a grim picture, they certainly don’t explain why humans are posting such inane content.

The problem isn’t that the internet is empty of humans. The problem is that humans are increasingly acting like bots.

The internet isn’t dead, it’s undead.

The Rise of the Zombies

A friend of mine recently ran an experiment on LinkedIn. He created two posts with the same core message. One was a thoughtful, data-backed analysis. The other was pure clickbait—an emotionally charged, simplistic take. The clickbait version outperformed the thoughtful one by an order of magnitude.

This isn’t an anomaly. It’s the system working as designed. The algorithmic conditioning of humans to operate with the formulaic predictability of bots is the key to understanding the Zombie Internet.

The goal of the attention economy is not to foster connection or disseminate truth; it is to maximize engagement. The algorithms that govern these platforms are not neutral arbiters of quality. They are optimization engines tuned to reward whatever keeps us scrolling, clicking, and commenting.

This has created a powerful feedback loop that actively reshapes human behavior:

• Algorithms Reward Simplicity: Content that is emotionally provocative, easily digestible, and conforms to popular opinion generates the most immediate engagement. It is therefore amplified. Nuanced, complex, or challenging ideas are systematically down-ranked because they require more cognitive effort link.

• Humans Adapt: We see what works. We see the simplistic posts go viral while our thoughtful analyses die on the vine. So, we adapt. We start writing for the algorithm, not for each other. We simplify our language, flatten our ideas, and chase the dopamine hit of likes and shares. We become conditioned to feed the machine what it wants.

• Discourse Becomes Homogenized: The result is a monoculture. A large-scale linguistic analysis of 300 million online comments found a “nearly universal reduction in text length, diminished lexical richness, and decreased repetitiveness” over the last several years. As we all learn to speak the language of the algorithm, our collective expression converges toward a bland, predictable mean—the very definition of bot-like communication.

This is the Undead Internet. It’s a system populated by algorithmically conditioned humans, all creating and engaging with each other’s optimized content in a vast, self-referential loop. The internet doesn’t feel soulless because the humans are gone. It feels soulless because we are stripping the soul out of our own work to get it seen.

Building the Enclave

There’s no cure to zombiness coming anytime soon. It’s just not clear how impactful efforts to “fix” the internet will be if there’s no real intent to exit the attention economy. That economy is hitting our dopamine buttons much too effectively for us to stop the “information superhighway” from falling into the sinkhole of likes and follows.

So, what are we to do?

I believe the only practical answer is to build enclaves—walled communities where we keep the zombies out. Much has been made of the need to make the home an enclave. I believe the workplace is the other critical enclave, a place where the attention economy is held at bay and an inspiration economy thrives.

I’m not suggesting a new perk delivered through an app that gives employees points towards free coffee; I’m suggesting a strategic imperative. The core functions of a highly creative and productive organization—deep thinking, nuanced debate, and authentic collaboration—are the very qualities that have been stripped from our social discourse. If companies want to continue to innovate, they have an imperative to protect their employees’ minds from the cognitive degradation happening outside their walls.

This isn’t the first time it has been in employers’ best interest to support the wellbeing of their workers. A century ago, the focus was on physical safety. In recent decades, it shifted to include financial and physical health through retirement plans and wellness benefits. The next frontier is social health. We must create environments that insulate our people from the constant pressure to perform for an algorithm and instead foster genuine human connection.

This means consciously designing a workplace culture that is unlike the internet. It means building internal communication systems that prioritize clarity over clicks and substance over speed. It means reducing the physical distance between co-workers, fostering in-person interactions punctuated by constructive debate and collaboration.

The most creative and productive organizations of the next decade will be those that recognize this threat and act decisively. They will build the strongest enclaves, protecting their most valuable assets—the focused, creative, and truly human minds of their people—from the zombies at the gate.

Share

Leave a comment

The problem is, it’s not true. Or at least, it’s dangerously incomplete.

Filling a role with an internal candidate doesn’t solve a vacancy; it just moves it. It’s like an electron-hole pair in a semiconductor. When an electron moves, it leaves behind a hole that exerts the same pull as the electron itself. Promoting a Director to a VP role feels great, until you realize you now have a Director-shaped hole in your organization. And that hole needs to be filled.

This kicks off a phenomenon I call the “Vacancy Chain”—a corporate conga line of disruption that quietly ripples through the company. The new Director role is filled by a rising Senior Manager, who leaves behind a Senior Manager-shaped hole. That’s filled by a Manager, and so on, until the chain finally terminates, usually with an external hire at the most junior level.

Each link in that chain carries hidden costs:

• The Productivity Tax: Every newly promoted person, no matter how talented, has a learning curve. For a few weeks or months, they are a net drain on efficiency as they ramp up. A single promotion creates one productivity dip. A chain of four internal moves creates four simultaneous dips across four different teams.

• The Onboarding Dominoes: Each move requires a handoff, training, and institutional knowledge transfer. The cost isn’t just in paperwork; it’s in the time senior people spend bringing their new reports (and their new bosses) up to speed.

• The Management Overhead: Every empty chair represents a cost, whether in delayed projects or the temporary burden shifted onto the rest of the team.

The same assumptions that have sparked unbridled enthusiasm for internal mobility can be used to highlight the hidden cost of vacancy chains. Starting with the average cost of external recruitment (typically 18-20% salary premiums), the ramp-up time for internal vs. external hires, and the probability of a backfill you can see that the value of internal mobility does not increase monotonically, it’s a U-shaped curve.

• On the far left (0% internal mobility): Your costs are sky-high. You’re paying massive recruiter fees and dealing with longer onboarding cycles for every single hire.

• On the far right (100% internal mobility): Costs are also surprisingly high. You’ve eliminated recruiter fees, but you’ve created organizational churn. The vacancy chains get so long that the cumulative productivity loss from dozens of people learning new jobs at once becomes a drag on the entire company.

The sweet spot—the optimal cost structure—lies in the middle. Based on this simplistic model the sweet spot will be roughly a 60% internal fill rate. This is the zone where you get the retention and motivation benefits of internal promotion before the vacancy chain turns into a multi-car pileup.

But a single, monolithic ratio isn’t a strategy. The real work for Talent Management teams is to apply this thinking with nuance. The optimal ratio is different for different types of roles, and in different industries. For example, the internal mobility target should vary by seniority:

• Entry-Level: These roles are the primary intake port for new talent and the natural place for vacancy chains to terminate. These roles can and likely should be filled with nearly 0% internal hires, think campus recruiting.

• Senior & Executive Leadership: At this level, deep institutional knowledge is a powerful asset. Yet, you must reserve space for the strategic external hire who can break groupthink and drive transformation. The ideal ratio is likely higher, perhaps 70-80% internal.

The goal isn’t to halt internal mobility—it remains a vital mechanism for growth, engagement, and knowledge sharing within any organization. It’s about shifting our perspective: organizations are complex networks, not static systems, and every talent move sends ripples across the business. Relying on an assumption that talent decisions can be made in isolation ignores the interconnected nature of teams and the cumulative impact of each move. By recognizing and accounting for the true costs embedded in vacancy chains—from productivity dips and onboarding overhead to management strain—we can move beyond the simplistic notion that “internal is always better.” Instead, we must design talent strategies that strike a deliberate balance, optimizing for long-term efficiency, retention, and organizational resilience. Only by taking this holistic approach can we build a talent system that delivers sustained value and supports both individual and company success.

Subscribe now

Share

Leave a comment