The graphs were real.
When the CEO stood up at the all-hands and showed what had happened to engineering output over the previous two quarters, he wasn't spinning anything. Deployment frequency up. Feature throughput exploding. Two engineers shipping what had previously required ten. The numbers were clean. The success was genuine. Something extraordinary had happened, and he had seen it happen.
So when he said this changes everything—when he said AI wasn't a tool anymore, it was an operating system, when he said every team should be asking why a human needed to do this, when he announced that anyone requesting new headcount would first need to explain why AI couldn't handle the role—the room nodded. Of course it did. The evidence was on the screen. The logic felt not just reasonable but inevitable. Nobody wanted to be the person who looked at a 400% improvement in deployment frequency and said: yes, but.
The word he kept coming back to was velocity. The company had found velocity. The goal now was to spread it.
Velocity Is a Physics Word
In physics, velocity means something precise: speed in a given direction. A vector, not a scalar. It tells you how fast you're moving and where.
What it doesn't tell you is whether you should be moving at all.
We have built entire management philosophies around the idea that speed is health—that the organization running fastest is the organization running best, that friction is always the enemy, that smoothness is always the goal. This assumption is so embedded in how organizations think about performance that questioning it feels like defending inefficiency. Like being anti-progress. Like being the person in the room who doesn't get it.
AI didn't create this assumption. It gave the assumption a power it had never had before.
For most of organizational history, friction pushed back. Humans got tired, disagreed, asked questions, slowed things down in ways that were difficult to fully eliminate. The velocity ideology was always there, but reality imposed limits on how thoroughly it could be enacted. AI removed several of those limits at once—and in doing so, didn't just accelerate the organization. It made the ideology feel empirically validated. The graphs went up. The mandate felt earned.
There's a subtler consequence that follows from this. When a system can only evaluate certain kinds of work—work that is legible, measurable, fast-feedback—organizations begin reorganizing themselves around those kinds of work. Not through policy, but through accumulated preference. The things the system can assess get prioritized. The things it can't quietly lose ground. The metric system doesn't just measure the organization. Over time, it reshapes what the organization believes its job is.
What gets validated along with the graphs is a category error that was always latent in the velocity worldview—one that only becomes visible when the tool is powerful enough to act on it completely.
Two Kinds of Hard
Some problems have structure you can learn. Clear inputs, legible outputs, feedback that arrives quickly enough to correct your course. The feedback loop is tight enough that repetition produces competence. Code either compiles or it doesn't. A chess move is good or it isn't. The gap between action and consequence is short enough to teach you something real.
Other problems don't work this way. The goals shift. The feedback arrives late, or sideways, or not at all. You can do everything right and still fail, and often you won't know which it was. Managing a team through a restructure. Sensing whether a key customer is genuinely satisfied or performing satisfaction before they leave. Deciding which problems are worth solving versus which ones are just loud. These don't yield to repetition. What accumulates isn't pattern recognition—it's judgment. The capacity to act well when the pattern hasn't shown up yet.
Researchers who study expertise call these kind environments and wicked environments—kind being the ones where feedback is fast and rules are stable, wicked being the ones where neither is true. The terminology matters less than the distinction.
The AI engineering breakthrough happened entirely in kind territory—code generation, automated testing, UI scaffolding, debugging assistance. Tasks with tight feedback and legible outputs that AI could learn thoroughly and execute at speed. The gains were real because the domain was right for the tool.
The problem was the inference: if it works here, it works everywhere.
Give someone a hammer and every problem starts to look like a nail. Give someone a tool that just eliminated months of engineering backlog in a quarter and every organizational bottleneck starts to look like an engineering backlog. The category error was invisible because the success was so legible—and because wicked problems rarely announce their failures on a dashboard.
You can have perfect velocity toward the wrong thing.
Smoothness Is Not Health
Somewhere in the last decade, friction became a dirty word.
Lean methodology, agile development, growth hacking, process optimization—each wave brought its own vocabulary but the same assumption: friction is waste. The well-run organization is the smooth one. Resistance is dysfunction. Remove the blockers, accelerate the throughput, measure the velocity.
This is sometimes true. Real organizations accumulate real waste—redundant process, unnecessary approval chains, meetings that exist because they've always existed. The velocity advocates are not wrong that organizations carry drag.
But smoothness is not health.
Consider what friction actually does in systems where it's working. Brakes are friction. Pain receptors are friction—the body's way of detecting damage before it compounds. Peer review is friction. The awkward conversation a manager has with a team member heading off a cliff is friction. None of these are waste. All of them are how a complex system stays in contact with reality rather than with its own outputs.
Meetings, disagreements, escalations, conflicting metrics—these look like blockers to execution. Sometimes they are. But sometimes they're the organization surfacing something that doesn't yet have language: competing goals, hidden risks, a structural contradiction that hasn't found its crisis yet. Removing the friction doesn't resolve the contradiction. It removes the visibility of it. The problem persists underground, accumulating, until it's expensive enough to be undeniable.
Some friction is what reality feels like before an organization has admitted that something is difficult.
Speed removes the pause in which that admission might happen. In complex environments, the gap between action and consequence is where learning lives. The delay isn't the problem. The delay is the sensing.
A body with no pain receptors moves freely. It also destroys itself without noticing.
What the Metrics Cannot See
When AI gets deployed across an organization in the name of velocity, the first results are legible and good. Response times drop. Output volume climbs. Costs fall. The metrics move in the right direction and they keep moving.
What doesn't show up in the metrics: nobody is escalating strange customer stories anymore.
There had always been a person who noticed. Not the person who read the ticket accurately and resolved it correctly—AI does that faster and at scale. The person who noticed was the one who caught something underneath the ticket. The customer who was technically getting her questions answered and was clearly, subtly, heading for the exit. The account that had always been slightly difficult and was now slightly more difficult in a way that meant something different. The complaint that didn't fit any category and was therefore telling you something a category couldn't.
This noticing isn't a skill you drill into competence through repetition. It develops through sustained exposure to human complexity—through getting it wrong and feeling the weight of it, through years of operating in conditions where the right answer isn't retrievable, only constructable. It looks like intuition from the outside. From the inside, it's accumulated judgment about what things mean in context, applied to situations where the context is doing most of the work.
The ticket closed. The metric moved. The customer left three months later.
The same dynamic appears wherever wicked work gets reclassified as kind. The team lead who senses burnout before it surfaces—it doesn't announce itself in a performance dashboard. It shows up in the slightly different energy of a Tuesday standup, in the way someone answers a question they would previously have pushed back on. The automated synthesis of engagement data doesn't catch this. What catches it—when it gets caught—is a person in the room reading something other than the output.
Or consider what happens to disagreement. AI-generated meeting summaries are fluent and efficient. They are also, by design, synthetic—they resolve the noise of an actual conversation into clean action items and consensus language. A room where three people had fundamentally incompatible views of the roadmap becomes a document where everyone is aligned. The friction disappears. So does the information the friction was carrying.
The organization that eliminates this friction doesn't become more capable. It becomes informationally rich and interpretively poor. More data than ever. Less capacity to know what the data mean.
The Counterfeit
The current wave of AI deployment is not the first time technology has displaced workers. What is new is the nature of what's being simulated.
Earlier automation was legible in its limits. The machine did the mechanical thing and you could see exactly where it stopped. A factory robot cannot tell you whether the product strategy makes sense. The absence is obvious. Absence is legible.
What AI produces is fluent.
It synthesizes. It contextualizes. It adjusts its register, offers caveats where appropriate, confidence where expected. In kind environments—where right and wrong are quickly distinguishable—this fluency reflects genuine competence. The output is trustworthy because the domain allows trust to be verified.
But fluency doesn't stay in its lane.
The same system that correctly retrieves the answer to a product question will assess whether a long-term customer is at risk of churning—with identical confidence, identical tone, identical apparent authority. It cannot tell you which of these two tasks it actually knows how to do. It doesn't experience the difference between a problem with a retrievable answer and one that requires being in the room, reading the room, having been wrong about this before and carrying that knowledge forward.
This is what organizations are beginning to discover about AI-generated analysis and recommendation: the output arrives without the cognition that would have produced it. The summary looks thorough. The assessment looks considered. But something that looks like understanding and something that is understanding are not the same thing—and in wicked environments, the gap between them is exactly where the damage happens.
Counterfeit judgment is not the absence of judgment. It's something harder to reject than absence: a plausible, confident, well-formatted simulation of judgment, produced by a system with no access to the thing it's simulating. A system that cannot tell you when it's operating outside its competence because it has no experience of competence to compare against.
The counterfeit is dangerous in proportion to how convincing it looks. And the organization that has eliminated the friction—removed the slow humans, compressed the deliberation, optimized the throughput—has also removed most of the mechanisms by which the counterfeit might be caught.
The People Who Already Knew
Some people in that all-hands felt something they couldn't name. They said nothing. The evidence was on the screen, the logic felt airtight, and there was no language available for what they were sensing. Wicked problems resist articulation almost by definition. The thing that's wrong doesn't yet have enough consequence to be described precisely.
But the sensing was real. And it was exactly the kind of sensing the organization was in the process of treating as friction to be eliminated.
The support worker who flags a strange customer interaction. The team lead who senses a shift in the room. The PM who pushes back on a timeline because something about it feels wrong. None of this shows up in a velocity metric. From the dashboard, from the executive presentation, it looks like drag—the thing that speed is supposed to overcome.
Velocity has no way to distinguish between friction that is waste and friction that is sensing. Both look identical in the short term. Both slow things down. Both resist the dashboard. The difference only becomes visible when it's gone—and by then the organization has already reorganized itself around the assumption that there was no difference.
The graphs still look good.
That's what makes this hard to see. The dashboards still report success. Tickets still close. Response times still fall. Features still ship. The organization is still moving—faster than before, in fact—while beneath the metrics the weak signals accumulate: the customers growing quietly polite before they leave, the teams no longer surfacing the disagreements that used to slow everything down, the decisions that feel slightly off but that nobody slows down enough to question.
Velocity is real. So is drift. The two can coexist for a long time before the gap between them becomes impossible to ignore.
By then, the dashboards are describing a system that no longer exists.