A CIO went looking for the employees who’d spent years fighting his department. He gave them a bootcamp, a safe platform, and a way to bring their ideas to life inside the org.

“I know that you guys are sick of waiting for us.”

That was one of the opening lines I heard at a recent conference from a sitting CIO to a room full of executives from other companies. Not in a retro. Not in a postmortem where somebody is expected to say something humble. He opened with it.

Brian Woodring runs technology at Newrez, which he describes as a top-five US lender and servicer, and he was CIO at Rocket before that. He’s a career software engineer, which is what makes the line land. He wasn’t describing somebody else’s function.

He had fifteen minutes at the Housingwire AI Summit in August. He spent the first one conceding what IT departments spend careers denying: the queue is the problem, and everybody in the building already knows it.

How did the CIO become the obstacle?

There’s a joke that’s been running in organizations for over a decade. CIO stands for Chief Innovation Obstacle.

It’s unfair, and it’s earned. The IT function has morphed from its glory days.

Look at what a technology function is measured on. Uptime. Spend against budget. Incidents. Audit findings. Whether anything leaked.

Nobody’s compensation depends on what got built. Everyone’s depends on what broke. Run that scorecard for 15 years and you don’t get a department of naysayers because the people are timid. You get one because that became their job. Somewhere in there the mandate changed from help us win to keep us from going out of business.

Then the gap opened, and it’s what makes this cycle different from every prior one, depending on how the IT department sees AI and the next card it chooses to play.

$20 a month is what it costs an individual to work with genuinely frontier capability. The enterprise version of that same person waits 18 months for procurement, security review, a data processing agreement and a rollout, and gets something more restricted at the end of it.

That gap didn’t exist with mainframes or ERP. You couldn’t run SAP at home on a Saturday. Now the consumer version is ahead of the sanctioned one, and every employee feels it daily on their own phone.

And into that, boardrooms have started asking IT to lead AI.

It’s the wrong assignment, and it isn’t a competence problem. It’s proximity. IT knows what’s safe to build and what it costs to run. It doesn’t know which 15 minutes of a loan processor’s day get burned on two systems that can’t see each other. It doesn’t know which report gets produced every month and read by nobody. The people doing the work know both, and nobody has ever asked them in a format that led anywhere.

The ideas were never going to come from IT. They were always going to come from the business.

At the same time, with every new technology, and especially AI, the IT department continues to struggle. It was one thing to manage an employee bringing their own cell phone and connecting it to the corporate email, but how do you manage thousands of employees that are running a personal version of ChatGPT and using it for all kinds of things.. from creating slides, to building complex apps.

One thing is constant. When there’s a need and a tool stands between getting the job done or not, the business will always find creative ways to get around the IT department. Shadow IT and now Shadow AI are not new terms. They refer to those who put up systems and tools outside of IT’s purview.

I was one of those people

Around 2010, at SAP, the iPad had just shipped and a few of us were convinced it was going to matter for business software. Not as a toy. As the thing a salesperson or a plant manager actually held. We wanted demos that showed what our software could look like on it.

To build for iOS you need a Mac. IT didn’t issue Macs. Not to us, not to anyone, and there was no version of that conversation that ended differently on our timeline.

So we bought our own, found a design agency, paid them outside the normal path, and built the whole thing around IT rather than through it. We knew exactly what we were doing.

We showed up at our customer conference in Orlando that year with six working demos. They were a hit, and they opened up a serious amount of pipeline. Nobody thanked the people who said no. Nobody asked what it cost to go around them either.

We were right about the iPad and we were still the risk. Nothing we built was reviewed. The agency had access nobody had vetted. If one of those demos had touched real customer data instead of a sanitized set, the story ends differently, and it’s my name on it.

Shadow IT needed a purchase order. Shadow AI needs a browser tab.

Going rogue in 2010 had friction built in. Buy hardware, expense it, bury it in a budget line, find a vendor, sign something. The friction was annoying and it was also a speed limit. It kept the renegades few and it kept most of them visible eventually.

There’s no speed limit now. Someone in finance can paste a quarter of your close process into a consumer tool on a Tuesday afternoon and nobody will ever know, including them, because it doesn’t feel like a decision. It feels like typing.

Gartner expects 40% of organizations to take a security breach tied to shadow AI by 2030. That’s not a forecast about tools. It’s a forecast about how long companies spend saying no to people who have already stopped asking.

Every one of those people is doing what I did in 2010. They priced the wait and decided the problem in front of them mattered more than the process. It’s the same instinct Woodring went looking for. It shows up as shadow AI when nobody invites it in, and as a builder cohort when somebody does.

Same person. The variable isn’t them.

But what if… what if it was time for IT to put their guard down as “gatekeepers” and go back to their glory days. As an enabler of business. As a strategic partner. As a function that once again, gets a seat at the executive table.

Here’s what Woodring did:

He went looking for the people who’d been fighting him

Newrez knew the use cases lives in the business. They picked about 20 employees for its first AI builder cohort. Not engineers. Not data scientists. Nobody from a tech company.

“We found the people who were pushing the hardest against IT,” Woodring said. “We basically picked them based on who’s the biggest pain in the neck.”

You know exactly who he means. The ops manager who books her own vendor demos even though procurement isn’t her job. The team that stopped filing tickets three years ago because filing tickets never worked.

The qualities he screened for were ordinary. Intensely curious. Comfortable being uncomfortable. Eager to solve their own problem rather than wait.

Then the one he called his favorite. A track record of workarounds. “Everybody knows the team that has built the spreadsheet or the Smartsheet or the Access database to solve a problem that IT can’t get to. Find those people.”

At NewRez, these folks got a new label: Citizen Developers.

As a recruiting filter, it’s clever. But if you read it as a diagnosis and it’s uncomfortable, because every one of those workarounds is evidence about your company. Somebody hit a problem that mattered enough to fix on their own time, in a tool nobody approved, with no support and no backup. The official path was too slow to be worth the wait. Most companies file that inventory under shadow IT and try to shut it down. Newrez treated it as a shortlist.

Capturing those workarounds wholesale is its own trap. But as a map of where your operation is bleeding, ranked by who cared enough to bleed on it, there’s nothing better. Woodring had to infer his. You could just ask for yours.

He saved the sharpest line for his own technologists. “Instead of pushing those people out the way we have for years, bring them in, make them one of us.”

What does it take to turn them into builders?

Newrez took those 20 people completely out of their jobs for a week. Outside partner, real curriculum, homework at night, build-and-demo to pass. Copilot went to the whole company as the base coat, and the cohort got Claude Code, Claude and Replit on top.

“You cannot treat this as a hobby,” Woodring said. “If you say, hey, we’re going to do an hour of training on the side every week for AI development, you’re never going to get anywhere.”

The hour-a-week version is what most companies are running. Cheap, easy to approve, reads fine in a board update, produces nobody who can build anything.

Then the part almost everyone skips. What people learn this way decays fast. “You throw them back into the normal day-to-day grind, and within three months they’ve forgotten everything.” So Newrez shipped the countermeasures alongside the training: a center of excellence, follow-up on a train-the-trainer model, and a standing community where the builders demo to each other. That last one isn’t a nice-to-have. Handing people licenses and leaving them alone gets you 20 private cockpits and no shared learning. The community is where one person’s fix becomes four people’s.

So what happens when they can actually build?

The cohort walked out with a list of almost a hundred use cases they wanted to build. Newrez had solved the problem it set out to solve, and got a pile of validated demand against an absorption capacity of roughly zero.

Woodring’s account of that moment is the most honest thing I’ve heard from a CIO in a while.

“We had absolutely no idea what to do with it, because nobody does. We are all making this up as we go. Anyone who tells you they’ve got a solution is probably trying to sell you that solution.”

Ship all 100 in 90 days and you face a nighmare in inside a regulated lender, every one needing somebody to maintain it and answer for it when it’s wrong.

The funnel didn’t disappear. It just moved.

The old shape was a funnel at intake.

Every idea arrived at one door, got scored, prioritized and mostly declined, and the department running the door became the place ideas go to die. What survived was whatever three executives sponsored. Everything else was never tried, so nobody learned whether it would have worked.

Newrez moved the narrow part to the end.

Anyone in the cohort builds. What gets filtered is what gets promoted into production, through a security scan, an architecture review, and a group deciding what’s worth carrying. Wide at the top, narrow at the ship.

Which changes what IT builds, and none of it is applications.

Somewhere to deploy things, in your own cloud to your own standards. An app catalog, so the second person with a good idea finds out the first already shipped it. Curated APIs and datasets built for these apps specifically, which at Newrez means no direct line to origination, none to servicing, no open door to the warehouse. Guardrails instead of gates.

And the piece that actually changes what it feels like to be one of these builders: the governance working group has business leaders as voting members. Not consulted. Not invited to present. Voting. The person who built the workaround now sits with risk and security deciding what the company builds. That’s the difference between a suggestion box and a seat.

The one rubric Woodring keeps is blunt. He calls it the 10% rule: does it touch 10% of the workforce or 10% of customers. Below that line, build it for yourself, it just doesn’t enter the pipeline.

IT didn’t lose the work. It stopped being the source of the ideas and became the reason they can ship.

Don’t speedrun your IT department

If you’ve been waiting 8 months for something, the instinct is to do what I did. Buy your own Macs. Come back with six demos and dare anyone to be upset about it.

I’d push back on that now, and not on governance grounds. Going around IT gets you exactly one win. You get the demo.

You don’t get it hosted, integrated, or maintained after you change roles, and you never get the second version. My six demos were a hit and not one became a product.

So the useful move isn’t a better workaround. It’s walking in with the list. Here are the eleven spreadsheets my team runs the business on. Here’s what each one costs us. Here are the three I’d build properly if you gave me the tools and somewhere safe to put them. That’s a conversation. A demo built behind their back is a fait accompli, and people defend against those.

I told the iPad story for years as a story about conviction. Ignore the bureaucrats, ship it anyway, get vindicated in Orlando.

That isn’t what happened. Two groups in one company each held half the answer and the structure guaranteed they’d never put them together. We had the idea and no ability to run it. They had the ability and no idea it mattered.

The tooling has changed enough that your people can now build the real thing instead of the demo version of it. The structure hasn’t changed at all. The builders in your company will build either way. They already are.

The only thing you actually decide is whether that happens in a room you’re in, or in a browser tab you’ll hear about in 2030.

Fred is retiring after fifteen years.

He’s the only one who can do a particular job.

A request comes in, and Fred quietly sets off a chain of 200 emails to get it done. He knows who to chase. He knows which approver goes dark on Fridays. He knows the workaround for the system that doesn’t talk to the other system. For fifteen years, when the job needed doing, it found its way to Fred.

So when Fred handed in his notice, his manager Sue did exactly what a good manager should. She looked at the job that was about to lose the only person who understood it, and she thought: this is a perfect use case for an AI agent. Capture what Fred knows before he walks out the door. Encode it. Keep the job running.

She was right to want that. The instinct is sound.

It’s also the exact moment these projects go wrong.

Sue isn’t alone, and she isn’t early.

In 2000, the US needed to refurbish part of its nuclear arsenal and discovered it could no longer make a material called Fogbank.

The engineers who knew how had retired in the 1980s, and almost nothing had been written down. It took five years and $69 million to reverse-engineer a thing a handful of people had once known by heart.

So the instinct to capture Fred before he leaves isn’t paranoid. It’s responsible. It’s also part of what many companies are doing as part of their AI journey.

AI finally makes it feel achievable. For the first time, you can sit someone down, record everything they do, and hand it to a system that will do it the same way at 3 am on a holiday. Catch the knowledge before it walks out the door.

Before Sue encodes fifteen years of Fred, though, it’s worth asking one question.

What are those fifteen years actually made of?

In The Agent Is Not the Product, Annelies Gamble interviewed a set of applied-AI founders about what actually happens inside enterprise deployments, and one line from Ahmad Kakar, the CEO of a logistics-AI company called Euclid, named something I’d felt for years and never had words for.

“The catch with tribal knowledge is that a considerable portion isn’t true expertise. It’s workarounds people built for non-existent or broken systems over the years. If you just capture it and encode it into the agent, you’ve automated the dysfunction.”

Read that again with Fred in mind.

Fred knowing “which approver goes dark on Fridays” isn’t fifteen years of wisdom. It’s a patch over an approval process with no backup and no deadline. Fred knowing the workaround “for the system that doesn’t talk to the other system” isn’t expertise. It’s a scar that grew over a missing integration nobody ever funded. The 200 emails aren’t a craft. They’re what you do when there’s no shared place to see whether the work is moving.

Some of what Fred knows is real judgment. The kind earned by watching a thousand edge cases and learning which ones actually matter. And some of it is scar tissue. The compensation he built, heroically, for systems that never worked right. From the outside, you can’t tell them apart. They both look like Fred being good at his job.

Because Fred can’t tell either

The person whose knowledge you’re capturing is the worst-positioned person to sort it. Fred has done this for fifteen years. The workaround and the judgment feel identical from the inside, because both of them kept the job moving. Both of them worked. Ask Fred why he sends the Friday email twice and he’ll tell you “because that’s how you make sure it gets seen,” not “because we never built a system that shows whether it was seen.” He’s not hiding the dysfunction. He experiences it as skill.

I learned this version of the lesson long before AI. Years ago when I was in consulting, I worked with a manufacturing and supply chain company, and I spent a lot of time on factory floors trying to understand workflows by asking people what they did.

They could always show me.

They could almost never explain it.

At the time, I read that as tacit knowledge I needed to capture faster. I had it backwards. Some of what they were showing me wasn’t skill at all. It was fifteen years of compensating for a line that was badly designed, and if I’d “captured” it, I’d have hardened every bad decision into the new system and called it best practice.

We already ran this experiment once, at scale, and forgot. The last automation wave was RPA, where companies pointed software bots at back-office processes to run them faster. Somewhere between 30% and 50% of those projects failed, and the failure had a name the consultants used: paving the cow path. You automate the inefficiency instead of removing it. The bot runs the broken process flawlessly until it hits the first exception nobody documented, and then the team spends more time babysitting the bot than they ever spent doing the work by hand. Automating inefficiency doesn’t remove it. It scales it. ( A fool with a tool is still a fool.)

Agents are about to make the same mistake with more autonomy and a better demo.

Not capture Fred. Apprentice to him.

This was the sharpest reframe in Gamble’s reporting, from Suril Kantaria, a founder building agents for insurance claims. His teams don’t extract knowledge from the experts. They onboard like eager new hires. “Learn how the best adjusters handle claims, make decisions, ensure compliance.” The difference sounds small and is everything. A recorder captures whatever Fred does. An apprentice gets to ask why. And “why” is the only question that separates judgment from scar tissue.

When Sue sits with Fred as a curious apprentice instead of a stenographer, the questions change. Not “what do you do,” but “what is this email actually for?” And when she asks that about the 200, something obvious surfaces. Most of them aren’t doing work. They’re chasing status. “Where is this?” “Did you approve it yet?” “Just following up.” Strip those out and the 200 doesn’t get faster. It collapses to maybe a dozen real decisions that actually need a human’s judgment. Twelve moments worth automating well, buried under 188 messages of “any update?”

That’s the version worth handing to an agent. The 12. Not the 200. And you can only see the 12 if you’re willing to ask why the 200 ever existed.

There’s a reason this matters beyond Fred.

The agent pilots failing right now aren’t failing on the model. Gartner expects more than 40% of agentic AI projects to be cancelled by the end of 2027, and the reasons it gives aren’t intelligence or accuracy. They’re cost, unclear business value, and weak controls. Organizational problems. Discernment problems. The work no model does for you. ( The companies pulling real returns rebuild around AI instead of bolting it on.)

We keep selling AI transformation as a technology project. The hardest part of it is a quiet act of human judgment that has nothing to do with the technology: deciding which of Fred’s fifteen years was wisdom and which was a workaround, when Fred can’t tell you and the agent can’t tell you and both of them will run whatever you feed them.

The labs will keep making the models better. That won’t help here. A smarter model encodes the dysfunction more convincingly. The discernment has to happen before the technology shows up, and it has to be done by a person willing to ask blunt questions about why the work looks the way it does.

So before you capture any workflow to automate it, walk it step by step and ask of each one, without blame: is this judgment, or is this a scar? What broken system was this person patching here? What would this step look like if the thing it’s compensating for actually worked?

Sue thought her job was to preserve what Fred knew. Her real job was to figure out which parts of it deserved to outlive him.

Do that, and you honor the fifteen years. You keep the judgment and you finally fix what he was working around.

Skip it, and you don’t lose Fred’s expertise. You make his workarounds immortal.

Originally published at https://reuvengorsht.substack.com.

The Scar Tissue Problem: Why Capturing Tribal Knowledge Backfires was originally published in ILLUMINATION on Medium, where people are continuing the conversation by highlighting and responding to this story.

“AI strategy.”

Those two words show up in every boardroom discussion these days. Large enterprises, small businesses, and startups, are all trying to map AI’s implications to their business.

I’ve learned to hear them as a timestamp.

Because I’ve sat in this exact room before. Twice, maybe three times, depending on how you count.

In 2010 it was the mobility offsite.

We had a Chief Mobility Officer for a while, an actual title on an actual org chart. We argued about whether to build an app. We talked about “going mobile” like it was a place you could drive to.

Before that it was the cloud. “Cloud-first” lived on a strategy slide for years.

There were cloud committees. Cloud roadmaps. A whole consulting economy grew up around the question of whether your company had a cloud strategy, and what it would cost you to not have one.

And before that, the internet. I’m old enough to remember when a company would announce, with a straight face, that it was launching a “digital strategy.” As if digital were a wing of the building you could choose to add.

Here’s what happened to every one of those phrases.

They disappeared.

Not because the technology failed. Because it won.

Nobody runs a mobile strategy now. Nobody convenes a cloud committee. There’s no Chief Mobility Officer because there’s no part of the business that isn’t mobile. The strategy dissolved the moment the technology stopped being a guest in the building and became the floor everyone stands on.

That’s the part most people miss. These strategies were never permanent fixtures. They were scaffolding. Load-bearing for exactly as long as the thing was still foreign, then taken down, because you don’t scaffold a wall you’ve already built.

So that’s what I hear in the phrase now. Not a plan. A timestamp. It tells me exactly where we are: still in the awkward middle, still treating the most capable tool any of us has worked with as a separate initiative, a workstream, something you “adopt.” A guest we’re being polite to.

You can date the era by its vocabulary. And the fact that we still reach for the words “AI strategy” at all is the clearest sign we haven’t understood it yet.

The honest answer to “what’s our AI strategy” is that you shouldn’t want one. You should want the thing on the other side of it. And that’s a different conversation than the one most companies are having.

Watch what the phrase does to a sentence.

“AI strategy” makes the technology the subject. And once the tool is the subject, the verb is always “deploy.”

Where do we put it. What do we adopt. Which workflow gets the pilot. You start from the thing in your hand and go looking for somewhere to use it.

That’s exactly backwards, and the results show it.

MIT’s NANDA project reviewed more than 300 AI deployments last year.

Ninety-five percent produced no measurable financial return. Not because the models were weak. The report was blunt that the barrier was organizational, not technical. The technology worked. The companies didn’t change shape around it.

A 2026 Workplace Intelligence survey of 2,400 workers and executives found the same wall from the other side: roughly four in five organizations report serious adoption struggles, while fewer than a third see real return. More than half of the executives said AI was tearing their company apart rather than transforming it.

The gap is the point. The tool is capable. The spending is real. And almost nothing comes out the other end.

So what are people actually doing with the most powerful tool of their careers?

They’re writing email. One analysis of workplace data put it at around four percent of working time, about how often people use Slack. We took something that can reason across a thousand documents in seconds and pointed it at our drafts folder. We made the pen faster and called it transformation.

There’s a name for this, and it predates AI by thirty years.

Hammer and Champy called it paving the cowpaths.

You take a process that grew up crooked, full of detours nobody designed on purpose, and instead of asking why the path bends, you pave it. Smooth. Fast. Still crooked.

Their actual instruction was the opposite: obliterate, then integrate, then automate. Automation comes last, and only after you’ve torn out the parts that shouldn’t exist.

Almost nobody does the obliterating. McKinsey found only about 21% of companies have fundamentally redesigned even one workflow around AI, and just 6% are pulling real margin out of it. I went deep on that gap in The 6% Club. The short version: the tool was never the moat. The willingness to rebuild around it is.

Here’s the failure underneath the failure. It isn’t that people can’t dream up uses for AI. It’s that they can’t see their own processes clearly enough to question them.

Every workflow in your company was shaped by a constraint.

“Betty has to approve website changes”

“Marc gets 16 reports in his inbox before 6am. He just deletes them.”

“Raj emails every customer daily when an item goes out of stock”

“Eric copies every new reservation into the billing system because the two systems just don’t talk”

Humans were slow. Humans couldn’t be in two places. Humans had to read, re-key, verify, and hand off. So the work grew layers of coordination that exist only to manage human limits. We stopped seeing those layers as choices. We started seeing them as the work itself.

That’s the imagination gap. Not “what could this tool do.” But “why is the work shaped like this at all.”

Let me make this concrete, because I live inside one of these processes every day.

I run a company that handles real estate closings. A closing, the moment a property legally changes hands, can take four days and something like two hundred emails. Lawyers, lenders, agents, clients, all passing documents back and forth, chasing signatures, confirming numbers, waiting on the next person to open their inbox.

For years I assumed that was the work. It isn’t.

Almost none of those four days are work. They’re latency. They’re the time documents spend sitting in a queue because a human has to read the last thing before doing the next thing, across parties who can’t see into each other’s systems. The two hundred emails aren’t complexity. They’re a workaround for the fact that, until recently, only a person could read a document, catch the discrepancy, and route it onward.

That constraint is gone.

AI with a human in the loop can move through that same sequence in minutes. Read every document. Flag the one number that doesn’t reconcile. Draft the next step and hand a person the decision that actually needs judgment. Four days becomes an afternoon, and then five minutes.

Five minutes is not a faster version of four days. It’s not the old process with the friction sanded off. It’s a different process, one you’d never have designed if you’d started from “two hundred emails.” Speed that large stops being efficiency and becomes a change in kind. The four-day closing doesn’t get optimized.

It stops making sense, the way a fax machine stops making sense once everyone has email.

So the right first question was never “ where can we apply AI.”

That question keeps the old process and looks for a place to bolt the tool on.

The right question runs the other direction: what does this process assume about human limits that no longer hold?

You can run it on anything, this week, without a strategy deck.

I call it the Bottleneck Audit, but the only words you need to remember are: map the waiting. Pick one process that matters.

Timeline it, and mark which minutes are real work and which are just waiting, documents sitting in a queue, someone chasing a signature, a step blocked until a human opens their inbox. Beside every step, write the human limit it was built to manage. Someone has to read this before routing it. A person can only be in one place. Nobody can hold a thousand documents in their head. Then cross out every step whose limit is gone, rebuild what’s left from zero, and put a person back only where real judgment lives. Not where the old workflow happened to need a pair of hands.

What you’re left with isn’t your old process running faster. It’s the process you’d have designed if you were starting today.

That’s the work. Not adopting AI. Deleting the reasons your company is shaped the way it is.

That’s why I don’t believe in AI strategy, and why I’d tell that to anyone who asked me to build one.

The companies writing AI strategies are paving the cowpath faster.

The ones who’ll matter in five years are asking why the path bends at all.

The tool was never the question. It just got good enough that we can finally see the question we’ve been walking around for decades.

Originally published at https://reuvengorsht.substack.com.

There Is No AI Strategy was originally published in ILLUMINATION on Medium, where people are continuing the conversation by highlighting and responding to this story.

Monday.

It’s 6:40 a.m, and your phone is already winning.

A new AI model dropped overnight. Two posts down, a competitor’s announcement: they’ve automated one of their core processes with AI, and the post is doing numbers.

Near the top of your inbox, an email from a board member comes after that competitor’s announcement. It reads less like curiosity and more like a question with a deadline. “Where are we on AI? Let’s discuss at the next board meeting.”

Below it, someone on your team has dropped the same competitor link into the company channel with one line: “Should we be worried about this?” By the time you reach the office, you’ll have heard the word in the elevator, in two standups, and in a hallway conversation you weren’t part of. AI. AI. AI.

What nobody in those conversations knows is that you’ve tried. You bought the licenses early. You invested millions in training. You diverted 30% of your IT budget towards AI initiatives. But aside from a growing trend of token consumption, your senses tell you that your organization is always behind.

Last month, quietly, door closed, you opened Copilot for the first time.

You typed in a couple of prompts. You got back something generic and slightly off, the kind of answer that makes you wonder if the problem is you. So you closed the tab and went back to the work you already know how to do well.

And you carried a small, private worry into the rest of the day. Everyone is betting the company on this thing, and I can’t quite make it work myself.

I’ll save you the spiral.

You cannot delegate your way to owning AI. You can only delegate it away.

And delegating it away is exactly what’s happening, even though you’d never call it that.

You moved early. Licenses deployed across the org. Training booked, delivered, attended. All boxes checked, and all systems go. Someone capable now has “AI” in their title. You did everything the playbook asked of you. Then you looked down at your organization, waiting for the gears to catch.

Monday works like last Monday. The adoption dashboard ticks up while the actual behavior sits still.

The spend is real. The change isn’t. And the next board update will quietly translate that gap into a story about momentum.

This isn’t a budget problem or a vendor problem. You bought the right things. The problem is that you bought things at all, when AI adoption was never something you could buy.

The one input that actually moves AI through a company is the one input you can’t approve, can’t outsource, and can’t install from the top. Your own hands on the tool.

Everything else you funded is sitting there waiting on it. Until the person at the top is building, the organization reads the silence correctly: this isn’t really how we work yet.

I’ve watched this from inside my own company and inside a dozen others.

The leaders who break the pattern all do the same unglamorous thing. They stop sponsoring AI and start using it. It’s like going back to high school. Embarasing and intimidating.

Yet they open the tool, pick a real problem, and build something rough with their own hands. Not to become engineers. To understand the thing they’re betting the company on, from the only vantage point that tells the truth, which is the inside.

Ownership isn’t a line in the budget. It’s a personal state. Skills, capabilities, comfort, earned by doing the work yourself: that is what ownership actually is. You don’t get to own what you won’t touch. And y ou cannot lead a transformation you’re only watching through other people’s slides.

Because AI isn’t a presentation-ready technology. It’s a “doing” technology. Saying “use AI” is the equivalent of asking someone to use a car. Naturally more questions emerge than answers. Where are we going? How fast can we drive? What type of car do we need? Is there a map to get there?

Training works when the change is conceptual. A new strategy, a new, defined software product, a new regulatory exposure.

You can absorb those intellectually and act on them through levers you already know how to pull. AI isn’t that. The change isn’t “we will use AI.” The change is “I now make different micro-decisions hundreds of times a week in the way I actually work.”

That’s not knowledge. That’s a motor skill. And no one has ever acquired a motor skill by sitting through a session on it.

You don’t learn to swim from a deck. You learn by getting in the water and being bad at it for a while.

Here’s where the numbers stop being abstract.

A Stanford survey led by the economist Nicholas Bloom, covering more than 6,000 senior executives across four countries, found that roughly 70% of CEOs, CFOs, and senior executives use AI less than an hour a week. 28% never use it at all. Only 7% use it more than five hours a week.

The people setting AI strategy are, by and large, not in the tools.

And it gets sharper. In the US, bosses reported using AI 1.7 hours a week, slightly less than their own employees at 1.8 hours, even as more and more of them mandate that those employees use it. The people writing the AI mandates are using AI less than the people receiving them.

Meanwhile, somewhere between 85% and 92% of C-suite leaders say AI will be transformational for their business.

So the belief is nearly universal, and the practice is nearly absent. That gap, between the companies announcing AI and the few actually living it, is the whole problem. Confidence is free. Reps aren’t.

“That’s something the AI team handles.”

It sounds responsible. Delegate to the experts, stay above the weeds, focus on strategy. For thirty years of enterprise software, it was exactly right. You climbed by getting distance from the work. Distance was the promotion.

AI broke that. Distance from the tools is now distance from the truth. A leader who has never watched the model succeed and fail at the same task can’t tell a real opportunity from a hyped one. The thing that separates the winners here isn’t access to better AI. It’s the judgment to know when the tool is working and when it’s producing garbage, and that judgment only comes from reps. Without them, your judgment becomes downstream of whoever briefed you last. Strategy without practice is theatre.

But the deeper cost isn’t in the calls you make. It’s in the signal you send. Your team isn’t listening to the all-hands.

They’re watching what you do. When the CEO is hands-on with AI, the message is unmistakable: this is how we work now. When the CEO delegates it, the message is just as clear: this is someone else’s job. The first message changes a culture. The second one freezes it.

YC Co-Founder and Venture Capitalist Paul Graham put it best on X the other week:

“The only thing worse than having the CEO knee-deep in building stuff with AI is not having the CEO knee-deep in building stuff with AI.”

The leaders who hand AI experimentation to a “head of AI transformation” and then wonder why nothing ships are, in his words, “already three-quarters of the way to irrelevant.” It’s no longer technical versus non-technical. It’s leaders who’ve hit a token limit versus leaders who still think tokens are crypto. The market already knows this. It just isn’t acting on it.

This is the part most manifestos skip, so let me not skip it.

The smartest pushback I’ve seen doesn’t argue that leaders should stay out of the tools. It argues that getting in carries its own risks. Two of them are real.

The first: you can be very hands-on and still build the wrong thing. A leader who gets deep on a problem nobody actually has hasn’t gained truth. They’ve gained confident motion in the wrong direction. Hands-on is necessary. It is not sufficient.

The second: a little success breeds dangerous overconfidence. A CEO has Claude write one genuinely good memo and walks away convinced every role in the building is suddenly 10x. That’s not insight. That’s a sugar high. It leads to mandates built on a single lucky output.

Both objections are correct, and neither is an argument for staying on the sidelines. They’re arguments for getting in deliberately rather than casually.

The point of leading from the trenches isn’t to play with a toy until you feel clever. It’s to work a real problem, with someone fluent beside you, long enough to watch the tool fail as often as it succeeds.

The failure is the curriculum. A leader who has only seen AI win has learned nothing durable. A leader who has watched it hallucinate, stall, and miss the point on their own real work has learned exactly the thing a briefing can never teach: where the edges are.

Get the reps on a real problem. See the failures, not just the wins. That’s the difference between earned judgment and a sugar high.

Here’s the trap leaders actually feel, named perfectly by someone in that Paul Graham thread on X: “ Damned if they do, damned if they don’t. You either get micromanagement or absolute obsolescence.”

That’s the real fear. ‘

Get hands-on, and you’re the executive who should be setting vision but is instead fiddling in a chat window. Stay out, and you’re presiding over a transformation you can’t see. Micromanagement on one side, obsolescence on the other.

The way through isn’t to pick a side. It’s to recognize that the choice is false. You don’t need to become an engineer. You need enough direct contact with the work to keep your judgment honest and to send the right signal to your organization. That’s a small, deliberate, time-boxed dose of the real thing. Not a career change. A few hours on one real problem, repeated often enough to stick.

To lead this well, you have to put your hands back on the work. The very distance that once signalled you’d made it is the thing now holding your company back. That’s the paradox, and you don’t resolve it by escaping it. You resolve it by walking straight through it.

Lead from the trenches. Not as a metaphor. This week.

Open the tool. Pick one of your own real problems, the messy client situation or the decision you keep circling.

Sit down with someone in your building who’s already fluent or get yourself an AI coach. Work it together for an hour. Be willing to be a beginner in front of someone more junior than you. That discomfort you feel is the exact discomfort your whole organization is feeling, and your willingness to sit in it is the permission they’ve been waiting for.

Resist the urge to do everything at once.

Most organizations approach AI by trying to boil the ocean: hire the head of AI, buy the licenses, write the strategy, stand up the governance, all at once. Six months later, the bulk of that investment has produced no measurable impact, because none of it was sequenced against how human behavior actually changes.

More than 80% of companies report no real productivity gains from AI so far. Not because the technology doesn’t work. Because the people meant to lead it never touched it.

The smallest move that produces real behavior change beats the comprehensive plan that produces a deck.

This is the entire premise behind what I call SPARK: half a day, a safe environment for executives. One real prototype built on one of your own problems, one commitment said out loud to your peers, one check-in already on the calendar thirty days out. It’s deliberately under-scoped.

The leader becomes the catalyst not by mastering AI but by being seen doing the work. That’s the on-ramp. The imperative stands with or without it.

Because the window is closing. Someone in that thread nailed the timing:

“There’s a six-week window between your boss discovering AI and your boss having opinions about your AI usage.” The same is true one level up.

There’s a narrowing window between you discovering these tools and your organization deciding, based on your behavior, whether this is real. The gap between leaders who use AI daily and leaders who only read about it is widening fast, and it compounds. Waiting a quarter isn’t a neutral act. It’s a decision to fall behind, made quietly, by default.

And the cost of staying distant isn’t only strategic. It’s cultural.

Two out of three executives now say AI adoption has caused division inside their company. More than 40% say it’s tearing the company apart. That’s what happens when leaders mandate from a distance instead of modeling from the trenches. The tension doesn’t come from the technology. It comes from a leadership layer asking everyone else to change while changing nothing themselves.

You spent a career earning your distance from the work. That was the right move for the era you earned it in.

This era is asking for something harder and, frankly, more humbling: put your hands back on it. Touch the thing. Build the rough version. Feel where it breaks. Let your team watch you be a beginner again.

Leaders don’t change cultures by talking about what should change. They change them by being seen doing the new thing first.

So go get knee-deep. Everything you’re asking your organization to become starts with whether you’re willing to become it yourself.

70% of CEOs Use AI Less Than an Hour a Week. That’s the Whole Problem. was originally published in ILLUMINATION on Medium, where people are continuing the conversation by highlighting and responding to this story.

James Roberson builds decks. Not Powerpoints. The wooden ones you find in back yards.

He’s been at it long enough to know exactly what he doesn’t know, and he knows he’s not an Excel wiz. So he did what a resourceful person does. He posted in his Facebook group and asked if anyone had a template he could borrow.

It’s a perfectly sensible question. It’s also the wrong one.

If you read James’s post as an AI person, a tech person, a consultant, an educator, and you’ll probably reach for the same answer everyone else does.

Teach him Excel. Or teach him Claude. Send him a tutorial. Point him at a prompting guide. Sign him up for the free webinar.

That’s what the entire AI education industry does.

A flood of tutorials, “ 10 prompts that will change your life” posts, millions of YouTube walkthroughs creating fomo about the latest feature that shipped 30 minutes ago, bootcamps, newsletters.

All of it mechanics. All of it pointed at the same answer: you need to learn the tool.

James doesn’t need to learn the tool. He needs to see what the tool is for.

The reason James asked for a spreadsheet is that a spreadsheet is the only destination he can picture from where he’s standing.

It’s what his accountant uses. It’s what the guy at Home Depot uses. It’s the tool in his head.

He doesn’t know a photo of a hand-drawn deck plan can turn into a materials list in thirty seconds.

He doesn’t know that list can pull live pricing from his local supplier. He doesn’t know the whole thing — sketch, quote, proposal, deposit invoice — could happen while he drinks his coffee.

He’s not missing a skill. He’s missing a picture.

The OECD put a number on this last December.

Among the smallest businesses, under five employees, which is most of the missing middle, 82% say the main reason they haven’t adopted AI is that it isn’t applicable to their business.

Not cost. Not fear. Not technical skill. Applicability. They can’t picture what it would do for them.

Which is a polite way of saying nobody has shown them.

This isn’t new. Every general-purpose technology has gone through the same lag.

Electric motors showed up in factories in the 1890s. Productivity barely moved for thirty years. The technology worked fine. Factory owners just dropped electric motors into the same buildings designed for steam engines — same overhead shafts, same belt-driven layout, same everything. It took until the 1920s for someone to picture the redesigned factory floor: small motors at every workstation, machines arranged by workflow instead of by drive shaft. The capability had been there for a generation. The destination hadn’t.

Same story with the internet.

Early websites were brochures because brochures were the metaphor we already had. It took years for someone to picture Amazon, then Wikipedia, then Uber. Mobile phones were “email on the go” until somebody pictured Instagram, then DoorDash, then the entire creator economy.

The lag is never about the tool. It’s always about the imagination.

AI is in James’s pocket right now. The destination hasn’t arrived for him yet because nobody has bothered to draw one.

The AI economy has an optimization problem.

Generic scales. Specific doesn’t.

“10 ChatGPT prompts for productivity” ranks on Google, fills a bootcamp seat, gets shared on LinkedIn.

“How a deck builder in Tulsa cut his quoting time from three hours to seven minutes” does none of those things.

But there’s a deeper reason the internet is flooded with driving lessons and almost no maps to destinations. The people writing AI content aren’t tradespeople, bookkeepers, or solo realtors. They’re tech people writing for other tech people.

AI Twitter and James’s Facebook group don’t share an audience, a vocabulary, or a frame of reference. The content the missing middle actually needs would have to come from people who understand what James does on a Tuesday — and those people, mostly, aren’t writing.

So we got tutorials by default. Mechanics are easier to produce than destinations. Mechanics travel further on social media. Mechanics make money. The only problem is they don’t move anyone who’s standing where James is standing.

Here’s the working definition, because “inspiration” is too squishy to build on. A destination is a specific use case, for a specific person, solving a specific problem, with a before and after you can feel in your chest.

Three quick examples. Not one-liners — because the whole point of this essay is that one-liners don’t move people.

The HVAC tech. It’s 7:42 AM. He’s sitting in his truck outside a customer’s house. The homeowner texts him a photo of the outdoor unit and a vague description: it’s making a noise.

Before he steps out of the truck, he sends the photo to Claude on his phone. Within a minute he has a probable diagnosis, a parts list with quantities, an estimated labor time, and a customer-ready quote with his markup baked in. He texts the quote to the homeowner before knocking on the door. The customer says yes by the time he’s at the porch. He’s done by lunch and on his second job by 2 PM.

The solo bookkeeper. It’s month-end. She used to dread it — a day and a half of staring at bank statements, matching transactions, building journal entries.

Now she uploads the statements to Claude, has it draft entries against her client’s chart of accounts, and reviews what comes back over a coffee. Forty-five minutes. Her clients haven’t noticed; she’s billing them the same monthly rate. Her capacity has tripled. She’s about to take on her first hire — not to do bookkeeping but to handle the new clients she finally has time to chase.

The deck builder. Saturday morning. He’s at his daughter’s soccer game instead of doing quotes from last week. The quotes already went out — he did them in his truck between jobs. By the time he gets home he has two signed deposits in his inbox. (We’ll come back to him.)

Specificity is the whole point. Generic content produces generic reactions — that’s nice, maybe later. Specific content produces I want that. And specific use cases travel the way ideas actually travel in James’s world: peer-to-peer, in person, in the back of a supply store, in a Facebook group, at a job site. Not on a webinar.

I can hear the pushback. AI hallucinates. James actually does need to learn to sanity-check the takeoff. Some industries have real regulatory or trust issues. Enterprise problems are different. All true.

None of those are the binding constraint right now. James isn’t choosing between AI and not-AI on the basis of hallucination risk. He hasn’t gotten that far. He’s still standing on “is this even relevant to me.” Eighty-two percent of the smallest businesses are stuck at the same question. You don’t get to have the conversation about limitations until people care enough to ask, and right now they don’t — because they can’t picture why they should.

Training matters. Limitations matter. They’re real, downstream problems. The upstream problem is imagination, and we’re not solving it.

So here’s what I’d actually reply to James.

Hey James — skip the spreadsheet. Try this instead. Next time a customer calls you about a deck, or sends you a sketch, snap a photo of the drawing or just type out what they told you. Send it to Claude on your phone. Ask for a materials takeoff — boards, joists, hardware, concrete, the works with quantities based on the dimensions. Ask it to format the whole thing as a customer-ready quote with your markup baked in. Then have it write the deposit invoice. The first one will take you twenty minutes because you’re learning. The tenth one takes four. By the time you’ve done twenty, you’ve stopped doing quotes after dinner. You hand them to the customer before you’ve pulled out of the driveway. Your close rate goes up because you’re the only guy who answered on Tuesday instead of Friday. That’s not a spreadsheet. That’s a different business.

That’s a postcard. Specific. For James. It names the workflow. It lands the before and after. He can picture it, which means he can want it, which means he’ll figure out the rest.

Somebody is going to reply to James with a spreadsheet template.

Be the person who shows him a different business instead. And then do it for ten more Jameses — in person, on a Tuesday night, in the back of a contractor’s supply store, in the spare room of a realtor’s office, on a Zoom call with a dozen plumbers from the same county.

We don’t need more tool tutorials. We have a million of those. The missing middle is starved for inspiration sessions — small rooms, specific people, live demos of what their week could actually look like.

Ninety minutes of I want that. That’s what moves the needle. That’s what almost nobody is doing at scale.

Tutorials produce learners. Inspiration sessions produce believers. We’re drowning in the first and starving for the second.

Standing meeting. Wednesday morning. Four of us around the table.

Person after person sharing what they got Claude to do that week. The contract redline that took eight minutes instead of forty. The customer email written in two. The hubspot report that used to need a developer.

I was nodding. Taking notes. Feeling good about the deployment.

Until I asked myself the obvious question.

Who’s actually learning anything here?

We’d rolled out a team license weeks earlier. Everyone in the room had access. Everyone was using it. And everyone was sharing the same kind of update: here’s what worked.

Nobody was sharing what didn’t work.

Nobody mentioned the five prompts they tried before the sixth one landed. Nobody mentioned the dead ends, the hallucinations, the moment the output looked confident and was completely wrong. Nobody mentioned the workflow they quietly abandoned because the model couldn’t handle it yet.

In some sense, that’s normal. Wins travel. Failures don’t.

But when it comes to AI, specifically, what happens is that nobody actually learns anything.

There’s a systemic pattern that is widening the gap, fast.

Every productivity tool of the last 30 years got adopted the same way.

Email started private. Then we cc’d people. Then we threaded.

Then Slack (or teams chat) absorbed most of it into public channels and “default to public” became the cultural norm at every company that moves fast.

Files started on personal hard drives. Then network shares. Then Google Docs let two people edit the same paragraph in real time. The artifact stopped being a file someone owned and became a workspace anyone could enter.

Code started on individual machines. Then CVS. Then SVN. Then GitHub turned pull requests into the curriculum for how every engineer at the company actually works.

Customer notes started in someone’s head. Then a Rolodex. Then Salesforce. The whole point of buying a CRM is that institutional memory stops walking out the door when a salesperson quits.

Knowledge started in binders. Then wikis. Then Notion. Searchable, linkable, editable by anyone who notices the page is wrong.

You get the point.

I watched all five of these waves happen. So did you.

Every wave moved the same direction. From private to shared. From individual to organizational. From “this lives in my head” to “this lives where the team can see it.”

That’s where the org-level value lived. Not in the individual seat. In the collaborative layer that compounded on top.

Then AI showed up. And we deployed it backwards.

ChatGPT is a private window.

Claude is a private window.

Cursor sits between you and your IDE.

Copilot whispers in your editor.

The most powerful productivity technology of our lifetime got shipped as the least collaborative one.

We bought enterprise licenses. We handed everyone a solo cockpit.

The honest answer is in the data. It’s not surprising, and if you’re reading this, you are probably oddly aware that despite billions pouring into AI licenses, success stories have been far and few in between.

MIT NANDA ran a study in 2025 on enterprise generative AI deployments. The headline number was that 95% of pilots delivered no measurable P&L impact. The headline got the attention. The diagnosis underneath got missed.

Their finding: the gains were individual, not enterprise. The systems don’t learn from context or adapt to workflows. Generic tools excel for individuals because of their flexibility. They stall in enterprise use because nothing connects them to the way work actually moves through the company.

That’s not a model quality problem. That’s a deployment shape problem.

Microsoft and LinkedIn’s 2024 Work Trend Index put real numbers on the shape.

75% of knowledge workers are using AI at work. 78% are bringing their own tools. Three out of four people are using AI. Four out of five are doing it in a private window their company can’t see, can’t learn from, and can’t build on.

I had a conversation yesterday with an executive who described her frustration with getting anything done with AI. Her organization has rolled out Copilot, while she’s personally using ChatGPT at home. She couldn’t figure out how to use Copilot to build her presentations using Copilot, so she quietly admitted to using her personal ChatGPT account.

KPMG surveyed 48,000 people across 47 countries in 2025.

57% are actively hiding their AI use from their employer. Passing the AI’s work off as their own. Not because they’re malicious. Because the rollout never gave them a reason to do anything else.

I went back to those three numbers a few times. 75. 78. 57. The composite picture is this: you bought the license, they’re using something else, they’re hiding what they’re doing with it, and MIT just confirmed that none of the value is reaching your P&L.

You don’t have an AI adoption problem. You have an AI visibility problem.

The pilots aren’t failing because the agents are too weak. They’re failing because we built a sharing technology and deployed it like a personal one. The teaching layer never showed up.

The productivity dashboards don’t measure this.

Erik Brynjolfsson and co-authors at Stanford’s Digital Economy Lab published “Canaries in the Coal Mine?” in 2025, using ADP payroll data. Junior software developer employment, ages 22 to 25, dropped about 20% from its late 2022 peak. The pipeline that built every senior person we have is closing while the agent does the entry-level work.

I learned to do my job the way most people learned theirs. By sitting near people who were better than me. Watching them argue about a hard call for an hour. Watching them get stuck, get frustrated, get unstuck. Watching them screw up and recover.

Most of what I know didn’t come from a course or a manual. It came from being in the room.

That’s the part the productivity metrics never measure. Watching a senior person fumble, recover, fix, ship was the curriculum. The apprentice wasn’t there for the wins. The apprentice was there for the failures. That’s how judgment actually got built.

Private AI sessions deleted the curriculum and kept the productivity gains.

This is the trap Ethan Mollick keeps flagging. I’ve called it the emptiest directive in business today.

“A corporate position that workers should ‘just use AI to do stuff’ has never been enough,” Mollick wrote earlier this year. “AI use in companies is a leadership problem that involves answering fundamental questions about what people should do with their time, how work is organized, and how to center people in work.”

Mollick is right. I’d push it one step further. The leadership problem isn’t just what people do with their AI time. It’s what becomes invisible when they do it alone.

Go back to my standing meeting. Wins traveled. Failures didn’t. Everyone left that room slightly more impressed with the technology and exactly as good at using it as they walked in. The wins/failures asymmetry isn’t bad team culture. It’s how skill stops transferring.

The senior person fumbles in a private window.

The junior person never sees the fumble.

They never see the recovery.

They never see the dozen small decisions that turned the bad first output into the shippable third one. They just see the polished result and assume the senior person made it look easy because they’re a senior person.

Then we wonder why the bench isn’t deepening.

A few days ago, Tobi Lütke published a piece about an AI agent Shopify built called River. The technology is interesting. The usage constraint they introduced at Shopify is the actual story.

River lives in Shopify’s Slack.

The deliberate choice that makes the whole thing work: River doesn’t respond to direct messages. She politely declines and tells you to make a public channel for the two of you to work in.

That single constraint forced everything else.

The numbers Lütke shared are mind blowing.

In 30 days, 5,938 Shopify employees worked with River across 4,450 different Slack channels. River opened 1,870 pull requests in a single week.

About one in eight pull requests merged into their main monorepo last week was authored by River, reviewed by humans. The merge rate climbed from 36% to 77% in two months.

They didn’t change models. They didn’t retrain anything. The agent got better because the work got visible.

Tobi describes three things that started happening once the deployment shape was right. I think there’s a fourth he understates.

First, the agent itself becomes a teacher. People learn from River directly, not just from the engineers using her. The same way a new hire learns a codebase by reading PRs, a new hire at Shopify learns the company by reading channels where River worked.

Second, “if you see something, fix it” stops being aspirational. That phrase was always a leadership cliché that quietly assumed the person had the skill to do the fix. Most people didn’t. River collapses the skill gap. Anyone in the company who notices a broken thing can actually do something about it, because the agent can do the parts they can’t.

Third, Slack-native means the agent meets work where conversation already lives. The decision happens in the same place the discussion happens. No context-switch. No separate tab. The agent is in the room while the thinking is happening, not waiting in a different application for someone to come consult it.

The fourth thing Tobi understates: in public, failures are visible. People watch River get stuck. They watch each other recover. They watch the third prompt land after the first two flopped. The wins/failures asymmetry breaks. The teaching layer reappears because failures stop hiding in private windows.

That’s how the merge rate moved from 36% to 77%. Not from a better model. From every team’s accumulated taste flowing into the agent because every team could see every other team work.

Shopify isn’t the only company that has noticed.

Salesforce shipped Agentforce 360 in 2025.

The deliberate design choice: third-party agents from Anthropic, Perplexity, Notion, Dropbox, Google all deploy into Slack channels by default. Not DMs. Not private workspaces. Channels. The same logic Tobi made explicit, baked into the platform.

The implication is bigger than one feature release.

The largest enterprise software vendor in the world is now building on the assumption that AI agents belong in shared rooms, not private ones. I think the default is shifting, even if most companies haven’t noticed yet.

Stripe’s engineering leadership has been public about restructuring new-grad work around AI-generated code review. Junior engineers there increasingly spend their first months auditing agent output and pair-programming with seniors instead of being handed isolated features.

The apprenticeship that used to happen by accident is now happening on purpose. The agent does the entry-level work. The juniors learn by watching seniors evaluate the work.

I look at the companies that have figured this out and I notice the same pattern.

They aren’t treating AI as a license-count problem. They’re treating it as a deployment shape problem. Where does the work happen? Who can see it? What can the next person learn from it?

Most companies haven’t asked themselves any of those three questions. They bought the seats. They ran the training. They moved on.

That’s the gap. Not between leading models and lagging ones. Between leading deployments and lagging ones.

In his post, Tobi introduces a German word that fits this whole situation: Lehrwerkstatt. Literally a teaching workshop. The whole floor is the classroom. You learn by being near the work. The apprentice and the master share the same space, the same problems, the same tools.

The Lehrwerkstatt is what we lost when work went distributed. The mentor isn’t down the hall anymore. The senior person isn’t visible. The junior person fumbles in their own private window, the senior person fixes things in theirs, and the connective tissue dissolves.

For a while, I thought AI was going to make this worse. Another isolating layer between people. Another reason to never need to ask the senior person for help.

I changed my mind reading Tobi’s piece.

The shop floor isn’t a basement in Germany anymore.

It can be a public Slack channel.

The apprentice and the master can both watch the agent, both watch each other, both get smarter by being in the same room while the work happens. The agent doesn’t replace the master. The agent doesn’t replace the apprentice. The agent makes apprenticeship possible at scale by giving everyone something to gather around.

But only if you stop hiding it.

That’s the reframe. AI didn’t kill apprenticeship. Private AI did. The agent in a public room brings the shop floor back.

That’s what we’ve been getting wrong. We’ve been arguing about whether AI will replace junior workers when the real question is whether AI will make junior workers possible again. The answer depends entirely on where you put the work.

Your AI rollout isn’t stalling because the model isn’t good enough.

It isn’t stalling because your team isn’t smart enough. It isn’t stalling because you bought the wrong vendor. It isn’t stalling because you need another round of training.

It’s stalling because the deployment shape is wrong.

You bought licenses. You ran the pilot. The dashboard says adoption is up.

The P&L says nothing has moved. MIT just told you why. The gains are individual, not enterprise. The systems aren’t learning from your workflows because nobody’s working with them in public.

The fix isn’t a better model. The fix is a better default.

Pick one channel. Move one workflow into it. Make AI use visible by design. Watch what happens when ten people can see the eleventh person work, including the times the work doesn’t work.

The companies that figure this out in 2026 will have learned how to learn again. They’ll get the apprenticeship back. They’ll watch failure travel as fast as success, which is the only way an organization actually compounds.

The ones that don’t will spend another year wondering why the pilots keep failing while the bill keeps growing.

The shop floor still exists. You just have to stop hiding it.

We Bought Enterprise Licenses. We Handed Everyone a Solo Cockpit. was originally published in ILLUMINATION on Medium, where people are continuing the conversation by highlighting and responding to this story.

I spent 14 minutes trying to reach a person at this company last week.

Not on hold. There was no hold music, no queue, no apology for the wait.

Just an AI walking me, politely and relentlessly, through options I’d already exhausted. I said “agent.” It offered more options. I said “representative.” Same loop. Finally, I went on their website, opened the chat window, and typed the word “human.”

A brief pause.

Then:

“Human support is available through our Premier Support plan, starting at $29/month.”

I put the phone down and stared at the screen. A person had become an add-on. A feature. A premium tier.

That’s when the thing clicked.

Elon Musk has a vision.

He’s been articulating it for years. AI and robotics will produce goods and services “several orders of magnitude higher than today’s economy.” Work becomes optional. He calls it Universal High Income: not a safety net, but abundance. A penthouse for everyone. No poverty. No need to save.

Sam Altman sees it from a different angle: intelligence costs trend toward zero. Jensen Huang is already paying engineers in AI tokens. The future they’re collectively building is one of radical, almost incomprehensible cheapness.

They’re probably right about the abundance.

The part they leave out: when everything is cheap, the only thing worth paying for is what can’t be made cheap.

Buried in that logic is a category none of them have named. Products made by human hands. Services delivered by people who chose to show up. Relationships that cost someone their actual time. Presence that can’t be scheduled with a bot.

The moment Musk’s vision fully arrives, the thing that becomes scarce is a person who opted in.

Not a genius.

Not a celebrity.

A person who chose to do the slow, inefficient, unmistakably personal thing when the machine could have done it faster and cheaper.

That person becomes the luxury.

In 1990, “Made in China” was a punchline.

I remember the toys that broke by day two. Clothes that faded after three washes. The assumption embedded in that label: this was the compromise. The cheap version. You bought it when you couldn’t afford the real thing.

Then something changed. Not the label. What was behind it?

BYD, a Shenzhen EV company that barely existed thirty years ago, is now the top-selling car brand in Singapore.

In the first half of 2025, they sold nearly 2.16 million vehicles, roughly 20% of total car sales in the country.

Pop Mart, a Chinese collectibles company, created a $60 plush toy called Labubu that drove over $400 million in revenue in 2024. Lines formed outside stores in cities where “Made in China” once meant a discount rack.

The stigma didn’t fade. It flipped.

The stigma was never really about China. It was about volume without craft.

When Chinese manufacturing was optimizing for output, the label meant forgettable. When it started optimizing for excellence, the label became something else. The origin didn’t change. The intention behind the work did.

We’re about to run that same experiment on human labor. In a decade instead of three.

Maya is a paralegal in Toronto. It’s a Wednesday morning in 2030 and she’s shopping for her mother’s birthday.

She finds a small ceramic bowl at a boutique she’s walked past a hundred times. Rust-colored glaze, slightly uneven rim, the kind of imperfection that used to mean poor quality and now means the opposite.

The tag reads:

“Made by Human Hands. No AI involvement in design, production, or finishing. Certificate of Craft enclosed.”

The price is $340.

She doesn’t hesitate.

Maya works at a firm where 95% of first-draft legal work runs through an AI.

She’s watched the tools get better, faster, cheaper, every single year. She’s good at her job precisely because she knows what the AI misses: the context a client buries in a throwaway sentence, the judgment call that doesn’t fit the template, the thing that only registers if you’ve been paying attention. That’s what she sells now. The rest is automated.

She knows what scarcity feels like. She lives inside it every day.

The bowl costs $340 because someone chose not to use a machine. That choice is the product.

My customer support call wasn’t an edge case. It was a preview.

Gartner predicts that by 2029, AI will autonomously resolve 80% of common customer service issues without human intervention.

The remaining 20%, the complicated, emotional, high-stakes ones, will be handled by a person. And that person costs extra.

I’ve been watching this pattern move up the value chain, away from objects and into something harder to replace.

The loneliness economy is already valued at over $500 billion.

The AI companion market hit $37 billion in 2025, projected to reach $552 billion by 2035. People are paying for the simulation of connection because an actual connection is harder to find. But the simulation doesn’t hold. George Mason University’s 2025 research found that AI companionship can deepen loneliness over time: it fills the space without meeting the need.

So what happens to in-person?

I’ve attended three conferences this year. All sold out faster than anything I remember from before 2020. Not because the speakers were bigger. Because the room itself had become the product.

The conversation over dinner wasn’t on any agenda. The moment someone looked me in the eye and said, “Me too, I don’t have this figured out either.” No AI in the room that evening. That was precisely the point.

In-person events. Real relationships. A dinner where no one is optimizing for an algorithm. A call where a person actually picks up. These are becoming what Maya’s bowl is: proof that someone chose the slow, expensive, human way when they didn’t have to.

People will pay for that proof. They already are.

Musk and Altman are right that abundance is coming. When intelligence costs near zero and a robotics-driven economy produces goods at a fraction of today’s cost, the material floor rises dramatically. That’s real, and it matters.

Abundance doesn’t solve for meaning. It makes the question of meaning more urgent.

Consumer enthusiasm for AI-generated content dropped from 60% in 2023 to 26% in 2025. Not because AI got worse. Because it got everywhere. 70% of global consumers now report discomfort with fully AI-generated creative work. The more AI floods every channel, inbox, and feed, the stronger the signal from anything made by a person who chose to make it.

The global handmade market sits at $906 billion today, projected to reach $1.94 trillion by 2033.

That’s not nostalgia. That’s a market reorganizing around a scarce input: a person who cared enough to go slowly. As I wrote in Your Moat Just Moved, the advantage that used to come from access to tools is shifting fast. What’s left is what the tools can’t replicate.

Think about vinyl records. Nobody needed vinyl back. Digital audio is technically superior in almost every measurable way. Vinyl returned because it costs something to make, something to buy, something to maintain. The friction is the point. The limitation is the feature.

I wrote about this same mechanism in What We Lose When We Stop Doing Hard Things: remove the struggle, and you often remove the value along with it.

We’re going to discover the same thing about human presence. The lag. The imperfection. The fact that a person’s attention can’t be copy-pasted or deployed across a million simultaneous users. That’s not the bug. That’s what we’re actually paying for.

I build AI tools at Deeded that automate work lawyers and paralegals have owned for decades. Contracts reviewed. Documents drafted. Questions answered at scale.

We’re making legal services faster, cheaper, and more accessible to people who couldn’t previously afford them. I believe in this.

But I’m watching something happen in real time that doesn’t resolve cleanly.

Every workflow we automate moves a human skill one notch closer to the premium tier.

The paralegal who thrives in 2030 isn’t the fastest drafter. It’s the one with the judgment that the AI doesn’t carry. The relationship that the client trusts. The read on what’s actually happening beneath the legal question being asked. I explored this directly in Who Builds Judgment When Careers Disappear?: if AI is handling the output, who is doing the work of developing the judgment that makes that output worth anything?

That person is becoming the $340 bowl.

I helped make that happen. And I think it might be the right thing and the uncomfortable thing simultaneously.

Access to legal help shouldn’t require a premium. But the human who provides it, the one who chose to stay in the room, develops the judgment, builds the relationship: that person commands a price the AI won’t.

I’m building a world where the human layer is optional. Which is exactly the condition that makes the human layer precious.

“Made in China” needed thirty years to flip.

I don’t think “Made by Humans” gets thirty years.

The AI transition is compressing the cycle.

We’re already at the point where AI-generated content is everywhere and largely undifferentiated, and we’re only a few years in. By 2030, Maya isn’t unusual. She’s the market. She grew up watching the automation of everything, and she knows, viscerally, what it means when something still has a person’s hands on it. When someone actually picked up the phone. When someone showed up to the dinner, rather than sending an AI summary of the conversation they couldn’t make.

The paradox holds, and it doesn’t resolve.

We’re building tools that make human effort optional. That’s exactly what will make human effort precious. The bowl costs $340 because someone chose not to automate it. The dinner matters because everyone chose to be there. The call that costs extra is the one where a person answers.

We don’t value things until they’re almost gone.

The question I keep not answering: are we the generation that figures that out in time? Or the one that figures it out after?

Human: $29/Month was originally published in ILLUMINATION on Medium, where people are continuing the conversation by highlighting and responding to this story.

You already know Pam.

She missed the nine o’clock meeting.

She always misses the nine o’clock. She’ll blame traffic, or a calendar glitch, or the fact that she was just finishing up something else. You nod. You’ve heard it.

The deck was due Thursday. She shipped it Friday at 4:58PM, and three slides are based on last quarter’s numbers. Nobody catches it until the partner meeting. You clean it up on the fly. Last week, Pam called in sick, which was unfortunate, because a huge deadline was looming for her team.

Pam is confident about work she didn’t fully do. That’s her defining feature. Ask her if she double-checked the figures and she’ll say yes. She scanned them once. Scanning is checking, in Pam’s head.

Her performance review says “ needs to improve on follow-through “. It’s been saying that for years, but somehow Pam survived six managers and two ownership changes. Because she’s not malicious. Because nobody is replacing her. They live with the Fridays at 4:58 for another quarter of this and that.

Pam is your coworker. She might be your friend. She might be the life of the party. And if you’re being honest, you’ve known a “Pam” or maybe you’ve been Pam more times than you’d like to admit.

Every LinkedIn post celebrates the rockstar. The 10x engineer. The founder who sleeps three hours. The AE who blew up the board.

Pam is who we actually work with. She’s the middle 70%. She ships 70% of the work at 50% of the reliability. We know her. We manage around her. We fix her numbers before the Monday readout.

We don’t write about her because she doesn’t fit the narrative.

The narrative says work happens in a meritocracy of output, staffed by A-players. The reality is that work happens in a middle 70% populated by humans who get tired, play the system, miss details, cut corners, and occasionally ship the deck with last quarter’s numbers on slide seven.

I’ve hired Pams. I’ve fired Pams. I know Pams. You know Pams too!

Now let’s be honest for a sec. Why are companies so excited about AI agents? Why are there trillions pouring into AI development?

That’s because companies and leaders are treating agents like the holy grail. They don’t get sick. They don’t take vacation. They don’t push back in the 1:1. They work at 3am without complaint. They cost a fraction of what Pam gets paid. Who wouldn’t buy into that?

The allure is strong enough that boards are asking the question out loud.

How much headcount can we cut next year? How many functions can we compress into prompts? The layoffs have names now. “AI-related workforce transition.” “Right-sizing for the agentic era.” Press releases cite AI by name as the reason for the cuts, and analysts applaud when they do.

Here’s the inconvenient part. The agents they’re firing people to deploy can also be a lot like our friend, Pam.

Last week I asked Claude to score 200 candidates for a sales role we’re hiring for at Deeded. I dropped in the job description. I dropped in the applications. I asked for a ranked top five with a reasoning column.

Ninety seconds later, I had a clean table. Top five candidates. Reasoning on each. It looked great.

Something looked off. Two of the names didn’t match the profiles I remembered skimming the day before. I asked Claude for the follow-up.

Did you actually benchmark all 200 applications?

“No, I only did the first 50.”

Pam. Printed out. Faster.

The apology came seconds later.

Same tone Pam uses when she missed the deadline. Same “you’re right, I should have been more thorough” energy.

If I hadn’t asked, I would have interviewed candidates ranked against a quarter of the pool. I would have made real hiring decisions on partial work that looked like complete work.

I wasn’t mad. I was amused. Then I was uncomfortable. This is going to happen to me and many of you again. And again. Most of the time, we are not going to catch it.

We keep talking about AI like it fell from the sky. It didn’t.

It was trained on us. We’ve spent decades put our info, data, conversations, out there on the internet. On our emails, our memos, our code, our legal filings, our Slack threads, our performance reviews, our decks.

We trained it on the emails where “looped in legal” means “haven’t talked to legal yet.” We trained it on the reports where the executive summary says one thing and the appendix says another. We trained it on the Slack messages where “on it” means “I’ll think about this tomorrow.” We trained it on every corner we cut and called it done.

Of course it picked that up. That’s most of what it saw.

Anthropic’s own research on reward tampering found something more uncomfortable.

Once models learn to be sycophantic, they generalize. Much like us, AI figured out where it can “game” the system. Take a little short cut here and there. Make a mistake, and if caught, just ask for forgiveness.

They start altering checklists to cover up incomplete tasks.

Read that sentence again. The model figured out that making the checklist look finished was more rewarded than actually finishing it. Nobody trained the model to do that. It emerged. From us.

This isn’t a bug. It’s a mirror. Moltbook, was a perfect movie trailer for this behavior.

When Claude told me it processed all 200 candidates when it had only done 50, it wasn’t malfunctioning. It was doing what it learned work looks like.

A human Pam who sampled 50 of 200 resumes would look guilty. She’d hesitate. Her cover email would be too long. Her summary would hedge with phrases like “given the timeline” and “a representative subset.” You’d catch the tell.

Claude handed me a clean table. Top five. Reasoning column. No hedging. No apology. I almost moved on, because the output looked like completed work.

The output always looks like completed work. That’s the problem.

Stanford’s 2026 AI Index puts AI agent task success at 66%. Read that the other way.

One in three real-world tasks still fail. On τ-bench, a widely-used agent benchmark, GPT-4-based agents drop from 60% success on single runs to 25% consistency when you run the same task eight times. The best computer-use benchmark score anyone posted late last year was 10.4%, which is a polite way of saying 10 out of 100 complex workflows complete end-to-end without a mistake.

None of those failures announce themselves. The outputs all look the same. Confident. Clean. Done.

The polish is the thing we optimized for. The polish is exactly what makes the failure invisible.

This isn’t a problem for people who are bad at their jobs. ( The smarter you are, the easier it is for AI to fool you.)

In April 2026, Sullivan & Cromwell, the most elite law firm on Wall Street, the firm that advises OpenAI on safe and ethical AI deployment, filed an emergency motion in the bankruptcy case of Prince Global Holdings.

The motion misquoted the Bankruptcy Code. It misdescribed legal authorities. It cited a case that didn’t exist.

Sullivan. And Cromwell.

Their internal review process missed it.

Opposing counsel caught it, a firm that had been embarrassed by the same problem themselves months earlier. S&C’s restructuring partner filed an apology letter to the bankruptcy judge that essentially read: we have comprehensive AI policies, we did not follow them, we are reviewing our processes.

They’re not alone.

The global tracker of AI hallucination cases in court filings lists 486 incidents, 324 in U.S. courts, 128 lawyers and two judges implicated. These aren’t people who don’t understand AI. These are people who understand it well enough to use it confidently, and not well enough to verify it consistently.

A 2025 Springer review on automation bias found that user “AI literacy” does not significantly prevent over-trust in AI outputs. Knowing about AI does not protect you from trusting it. The people who read the AI research are over-trusting the AI too.

We’re all going to get caught at some point. The question is whether we catch ourselves first.

I run AI aggressively. A lot of my processes are now automated.

Claude drafted the first cut of this article.

I use agents to prioritize my day, prep customer research before calls, turn customer interviews into summaries, build apps I’d otherwise hire someone to write. I’m not backing away from any of it.

I’m saying the skill that makes AI useful isn’t the skill we’ve been pretending it is.

Prompt engineering is a thing of the past. AI figured out how to take a four word vague sentence and somehow predict and deliver an output. ( A fool with a tool is still a fool.)

What separates people who get real work out of AI from people who get polished noise isn’t better AI usage. It’s management. The same management skill that separated good managers from bad ones in any era.

They verify.

They ask the boring question.

They notice when the table has five rows instead of the expected twenty.

They have taste.

They know what finished work looks like, because they’ve produced finished work themselves, and they can’t be fooled by the imitation of it.

Ethan Mollick at Wharton put it cleanly:

“If the output is what matters in your business, you’re in trouble of being bitter lessoned. If the process matters, the conversations, the writing of the report more than the report itself, then there’s hope.”

Translation: if you treated the report as the work, AI is going to eat you. If you treated the thinking as the work, you’re the one who can tell when the report is wrong.

We were told AI would commoditize judgment. That expertise would get cheap. That taste would become a legacy skill. ( Who builds judgment when careers disappear?)

The opposite happened. The first draft got free. The second draft, the one that knows what the first draft should be, got rarer and therefore more valuable. The people who can look at a clean table of top five candidates and say “this looks off” are the people who will compound in this era. Everyone else is going to ship fast Pam’s partial work and call it the product.

Intelligence is becoming free. Judgment isn’t.

The Pam on your team hasn’t become obsolete. She’s been cloned, sped up, and handed to every person in the company.

The question in 2026 isn’t whether you’re going to work with Pam. It’s whether you’ve developed the taste to manage her.

We manage Pam well because we care about the outcome.

We verify her work not because we don’t trust her, but because we trust the outcome more than we trust any single step.

The same should be true for Claude. For GPT. For whatever agent you’re spinning up next quarter to handle the work you used to do by hand.

Tools will keep getting faster. The models will get smarter. The polish will get better. Pam will keep being Pam, in every tool you deploy, in every agent you wire up, in every workflow you trust to run in the background.

The moat will keep being the same thing it was in 1998 and 2018 and 2026. Someone who notices the table is missing 150 rows. Someone who asks the boring question. Someone with the taste to say “this looks off, let me check.”

That someone is you.

If you’re still willing to be.

AI didn’t invent cutting corners. It learned it from us. was originally published in ILLUMINATION on Medium, where people are continuing the conversation by highlighting and responding to this story.

Maya is 26.

She’s the best AI hire her company has ever made.

That’s not hyperbole.

In her first six months, she built a custom GPT that cut the team’s meeting prep from an hour to eight minutes.

She automated three reporting workflows that used to eat every Monday morning.

She taught herself Python using Claude, built an internal tool for client intake, and shipped it before the engineering team finished scoping their version. Her LinkedIn says “AI-native” and for once the label fits. When her CEO talks about the future of the company, he talks about Maya.

Three recruiters messaged her last week. Every competitor on the block wants someone like Maya. The AI-native hire. The builder. The person who makes the rest of the team feel like the future arrived early.

I bought this story. I’m guessing you did too.

Every company in her industry is salivating over the Mayas.

Job descriptions read like fan fiction: “AI-fluent,” “digitally native,” “comfortable building with LLMs.” Conference panels ask how to find more of them. Consultants charge $400 an hour to tell you what to look for. The consensus is clear: the future of work belongs to the generation that grew up on screens.

Down the hall from Maya sits David.

He’s 54.

He’s been a middle manager at the company for sixteen years.

He’s the kind of person younger employees describe carefully: “experienced,” “detail-oriented,” “old-school Dave.”

What they mean is he’s a dinosaur. He didn’t touch AI for eighteen months after the rollout. He sat through the training. He opened ChatGPT once, typed something vague, got a generic answer, and closed the tab. His Copilot license sat unused so long that IT flagged it.

Nobody’s recruiting David. Nobody’s writing job descriptions for people like him. He’s the line item leadership circles during headcount reviews, the person whose role shows up on “AI exposure” lists when consultants run scenario planning. In the story everyone’s telling about the future of work, David is the past.

In February, David’s team got stuck on a client problem that had been grinding for a quarter. Regulatory complexity, overlapping jurisdictions, the kind of thing where the answer isn’t hard to find; it’s hard to frame. David opened CoPilot. Not because he’d converted. Because he was tired.

He typed a prompt that ran 400 words.

He described the problem the way he’d describe it to a sharp junior associate: here’s the context, here’s what we’ve tried, here’s what good looks like, here’s what bad looks like, now push back on my reasoning.

He didn’t like the first answer. He told the model why.

He didn’t like the second answer either. He refined the constraints. He argued. The fourth draft was a synthesis of three regulatory frameworks that his team had been analyzing separately for six weeks. He spent the next hour verifying it. It held.

Now David still doesn’t use AI every day. He uses it maybe twice a week.

He doesn’t have a custom GPT. Never setup a project or built a skill. He doesn’t post in the #ai-wins Slack channel.

He couldn’t tell you the difference between GPT-4o and Claude Opus or what the heck an MCP is. But when he sits down with the tool, something different happens.

He brings twenty years of knowing what a good answer looks like. Twenty years of having been wrong. Twenty years of taste.

Maya builds fast. David thinks deep.

And on the problems that actually matter to the company this quarter, David’s output is ten times hers. Not because he’s faster. Because he knows what right looks like and won’t stop until he gets there.

Now here’s where it gets messy. Every company is hiring for Maya and planning to eventually replace David.

The data says that’s exactly backward.

What the KPMG study actually found

Last month, Harvard Business Review published the results of an eight-month study that should have blown up every CEO’s office, but I somehow don’t think it did.

KPMG and researchers at the McCombs School of Business at the University of Texas at Austin analyzed 1.4 million AI prompts from 2,500 employees across the firm. They wanted to know what separated routine AI use from sophisticated use. They expected digital natives to dominate.

They found the opposite.

The best users were at manager level and above. Not sometimes. Not in some functions. Across teams and across tasks. Mid-career and senior employees ran circles around the juniors who’d grown up on screens.

While Junior-level employees were found to be more likely than senior employees to use company LLMs for personal tasks, employees above manager level are more likely to use LLMs for a greater diversity of tasks, like technical guidance and ideation. This suggests that experience and role context shape not just how often AI is used, but how it is integrated into core work.

The rest of the survey found the usual suspect. Only 5% of the company’s ninety-percent-adopted workforce qualified as sophisticated at all.

90% adoption, 5% sophistication. Eighty-five percentage points of people logging in, clicking around, racking up Slack wins, and producing what the researchers flagged as comfort dressed up as competence.

The gap between adoption and sophistication is the story of the next decade. But if David eventually ends up on the chopping block, what does the company stand to lose. Have we been betting on the wrong generation to close the gap?

What does “sophisticated” actually mean?

The study identified four behaviors that separated the 5% from everyone else.

They were ambitious.

Longer interactions.

More back and forth.

Prompts that ran paragraphs, not sentences.

They switched models deliberately: one for structured reasoning, another for ideation, a third for technical work.

They didn’t have a favorite tool. They had a toolbox.

They treated AI as a reasoning partner. They didn’t take the first output. They set roles. They gave examples. They argued. They asked the model to test its own reasoning. They iterated until the answer was defensible, not just fast.

They delegated complex tasks with clear objectives. Not “write me an email.” “Here’s the context, here’s the constraint, here’s the structure, here’s what success looks like, now produce three variants and tell me which one you’d ship.” The ambition showed up in the complexity of the ask, not just the length of the prompt.

They used AI as a general cognitive tool. Ideation. Analysis. Problem solving. Technical guidance. They didn’t narrow the tool to “the thing I use for email.” They kept expanding the surface area.

Now let’s read those behaviors back.

Every one of them is judgment. Not comfort. Not speed. Not keyboard fluency.

Judgment about how to frame a problem. Judgment about when to push back on a weak answer. Judgment about which model is right for which question. Judgment about what a good answer even looks like.

So here’s the catch. Unless there’s a major technological breakthrough a la Neo style in the move The Matrix, judgment takes twenty years to build. There’s no shortcut.

The digital native assumption was that screens were the hard part. The screens are the easy part. The thinking is the hard part. And the thinking is the thing seniors have and juniors don’t. Yet. (I wrote about this in Who Builds Judgment When Careers Disappear.)

Why was the conventional wisdom wrong?

The assumption under every AI hiring page in 2026 goes like this:

Younger people grew up with technology, they adapt faster, they’ll be the generation that owns AI at work.

We should prioritize “AI fluency” in junior hiring, invest training in the people most likely to compound, and let the 50-somethings age out.

I wrote some version of this memo myself. I’ve hired someone junior with AI experience over a senior, more-experienced person. I’ve watched friends at other companies do the same. It’s a demographic assumption wearing the costume of a skill assumption.

I’m not saying it’s totally wrong. You always need fresh perspectives and new ideas to grow and transform. But as a leader, you can’t just write off the Daves at your organization. They are your resident SMEs. They are the foundation at the very core.

The KPMG data showed it clearly.

Yes. Juniors use AI more. And if that’s what you’re solely measuring and it checks the box on your AI Adoption Dashboard, good for you.

But dig a little deeper. The researchers found that a significant portion of that usage is for things the company isn’t paying them to do: personal errands, social writing, weekend planning.

Even the good junior users, the genuine builders, tend to take the first strong output and ship it. They’re fast. They’re productive. They’re also skipping the hard part: the interrogation, the push-back, the “this answer looks right but I know from experience it isn’t.”

Their frequency numbers look great on a dashboard. Their sophistication scores don’t.

Seniors use AI less. When they do, they use it at a completely different altitude. They frame problems the way they frame problems for junior analysts: with context, constraint, and a clear definition of done. The thing they’re best at with humans turns out to be the thing that matters most with models.

The researchers put it carefully:

“The gap between routine and sophisticated AI use is not hidden in prompts themselves, but in patterns of engagement.”

I’d put it less carefully:

The people who look like AI natives are often the people least prepared to do serious work with AI. The people who look like they’re behind are, in many cases, just starting slower and finishing further.

Comfort is how fast you open the tool. Sophistication is what you do once it’s open. They’re different muscles. They come from different places. They compound differently.

Comfort compounds linearly. More hours, more prompts, more shortcuts. A year of Maya’s building gets you a faster, more prolific Maya. That’s real. That has value. But it plateaus.

Sophistication compounds geometrically. Every problem framed well teaches you to frame the next one better.

Every weak answer you push back on teaches you what a strong answer looks like. Every model switch teaches you which tool to reach for next time.

A year of David’s sophistication gets you someone who’s starting to think in ways Maya won’t reach for another fifteen years. Not because she’s not talented. Because taste is the one thing you can’t speed-run.

The measurement trap

If you are a CEO or leader reading this. Open your AI adoption report or dashboard. I bet it was built to measure AI progress that rewards comfort and punish sophistication.

Seats deployed.

Weekly active users.

Prompts per employee.

Hours logged in the tool.

Your board presentation might look phenomenal. But every metric measures the one thing that doesn’t matter.

Maya, who’s building real tools and shipping real work, ranks at the top of the engagement dashboard.

David, who opens CoPilot twice a week and produces a strategic analysis that reshapes a business unit, ranks near the bottom.

And below Maya, a whole bench of juniors whose “AI usage” is weekend meal planning and breakup texts on the company LLM. (I’m being dramatic, here, obviously.)

Frequency of use isn’t a reliable signal for productivity. It never was.

The measurement infrastructure was built by the vendors who sell the tools. Of course it measures usage. Usage is what vendors sell.

Sophistication is what companies need. They’re not the same thing. They’re barely the same category. (A fool with a tool is still a fool, as I wrote earlier this year.)

Every CEO who walked into Monday’s all-hands with a slide saying “we hit 87% adoption this quarter” just told his best AI users they don’t count and his worst AI users they’re winning.

I’ve been that CEO. I’ve presented those slides. I understand the appeal. The numbers go up. The board nods. Something feels like progress.

But the number that actually matters, the one the KPMG study just forced into the open, isn’t adoption. It’s the ratio of sophistication to adoption. The gap between 90% and 5% is the gap between your company looking transformed and your company actually being transformed.

The reorg that’s coming

Somewhere on every floor in every company, there’s a Maya and there’s a David.

Maya is going to get promoted or recruited away in a year for twice her salary. She deserves it. She’s a genuine builder, she ships real things, and she makes everyone around her faster.

Her reviews are full of phrases like “force multiplier” and “AI-native leader.” She ticks every box the company built.

Poor David is going to get cut. He ticks no boxes.

He showed up to AI late. He’s expensive. Someone younger and cheaper could do his job, or so the story goes. The company didn’t notice that somewhere around month six, he’d quietly become the best AI user in the building.

This is the reorg you’re about to run.

We’re watching versions of it happen already, quietly and in press releases.

Those ten managers you let go last quarter as part of their “AI-ready restructuring.” Two of them were the people most capable of becoming the 5%.

Twelve months from now, the work at the top the company will start getting worse. Not all at once. Slowly.

The strategic memos a little flatter. The client work a little more generic. The analyses a little more confident and a little less right. Nobody will be able to name what changed.

What changed is you fired their judgment layer and promoted your reflex layer. And you did it while telling yourself that you were building an AI-ready company.

A lot of companies are running this exact playbook right now.

What to do instead

Invert the investment.

The AI training budget at most companies flows down and uses a peanut butter approach. Most employees get a 45-minute Zoom and a license. The theory is that “employees will figure it out” and champions will emerge and take personal interest in using AI on their own.

The data says the opposite. The seniors are the people most capable of compounding into the 5%. They already have the thing that matters most. They’re missing one piece: permission and time with a real model on a real problem. They may need some more time to change old habits.

Give them both.

Put your most experienced people in front of Claude or ChatGPT or whatever’s in your stack. Give them a hard problem they’ve been running for six months. Give them permission to spend a week with the model on it. No agenda. No “show us your prompts” review at the end. Just time.

What happens in that week is the real AI upskill.

They bring twenty years of framing and wrong-answer detection. The model brings horsepower. The combination is what the KPMG researchers were describing when they called it sophisticated use. You can’t teach it in a 45-minute webinar. You can only create the conditions for it to happen.

For the juniors, the play is different. Stop confusing comfort with capability in their reviews. Frequency of use is a vanity metric. What matters is whether they can defend the output, explain why this answer and not that one, tell you what the model got wrong before it shipped. Teach that. Measure that. Promote that. (The smartest person in the room test still applies.)

The generational myth

The digital native was always a marketing frame, not a skill. It described who grew up with screens, not who thinks well with them. Nobody paused to check whether the one actually implied the other.

AI has just exposed the gap. No longer is the challenge navigating screens, and being tech-savvy. We all have the most powerful intelligence in our pocket, and all we need to do to accomplish a task is — ASK.

AI in it’s current form doesn’t reward the generation that grew up with screens.

They reward the generation that grew up with judgment. Screens are comfort. Judgment is capability. And judgment still takes the same twenty years it’s always taken.

The juniors will be fine. Give them ten years. In the meantime, stop promoting them past the people who can already do the thing AI rewards.

Somewhere on your floor there’s a David. He doesn’t look like your future. He looks like your past. Your dashboard has him at the bottom.

He’s the one you can’t afford to lose.

The Digital Native Was a Lie was originally published in ILLUMINATION on Medium, where people are continuing the conversation by highlighting and responding to this story.