You’re not “AI Native”.

It’s an interesting new obsession of ours. Don’t get me wrong, disregarding context for a second, I think generative AI is potentially revolutionary technology, I just don’t believe you. I’m not trying to be mean, I’m trying to be helpful, and that means I have to be critical. I don’t like rhetorical questions either, nor do I think you’re helped by me beating the living daylights out of a strawman of my making, so let’s honestly grapple with this. I can’t promise it’s all going to be comfortable, but you’ll be better off for it.

Who’s afraid of a little competition?

Competing. It's what we're all here to do, right? Capture some market segment, defeat some other player. We've always competed, and technology has always been a competitive advantage. Gunpowder, the rifled barrel, the predator drone — entire wars have been settled by one side holding an edge the other couldn't answer. And the language of war imports cleanly into business: we talk about capturing share, defending position, killing the competition, going to market like it's a beachhead. It's not that I have naive ideas about history, I've just seen the pattern misapplied so many times that I always have to ask: well, what does that mean for you, specifically?

Because — and I need you to sit down for this — you're likely not at war. Not in the physical sense, and not in the metaphorical one either. The thing that makes a market actually winner-take-all isn't size or stakes or how it feels from inside; it's a specific combination of conditions that have to hold at once. You need:

When all three stack up, you get the dynamic everyone's afraid of: one player takes almost everything and holds it.

When they don't — when you've got one of the three, or none — you get what most markets actually look like. Several durable players, each holding a differentiated position, none of them dominating and none of them perishing. Oligopolies and duopolies, not conquests. So run the test on yourself before you accept the war footing: do all three conditions hold in your market, simultaneously? For most companies, they don't even hold in pairs. The concentration that does exist is clustered — it shows up heavily in platforms, search, marketplaces, social, a few corners of the knowledge economy — and it's largely absent from the services and product businesses where most software actually gets built. Even the people who preach speed-to-scale will, if you press them, admit the markets that reward it are the rare exception, not the rule.

Which means the urgency you're feeling — the one being sold to you under names like "AI native" — is, for most of us, second-hand smoke. And I do mean us. I’m inhaling it too. But no monster coming for your business doesn’t mean no one’s benefiting: plenty of people are happy to keep the air thick, because the thickness pays them. The part they'd rather you didn't notice is that nobody made you inhale. You chose your investors. You chose who to believe on LinkedIn. I chose mine. The urgency feels like ours because in a real sense it is — we reached for it, and reaching for it felt like nerve, like not getting left behind, right up until you notice it was mostly deference cosplaying as decisiveness. Which is good news, if you can stand it! A thing you chose is a thing you can stop choosing.

Take a beat. Think about the market you actually operate in. Think back to the trade shows of the last couple of years. Some competitors aren't there anymore — but most of them are, right? Now think about the product and sales meetings you sit in. There are companies in your market you've never framed as a real threat: they hold an adjacent segment with similar-but-different needs, or they come at the problem from an angle that was never going to be for everybody. You're not winning that fight or losing it. You're not in it. 

A lot of this pressure comes from the people we took money from, and their other investments selling their wares to us. Some truly awe-inspiring companies were founded with VC money, and we’re right to look up to them (well, some of them). But not everybody you look up to is someone you have to imitate, and not every pattern is worth copying. You should also know your investors are very likely also invested in the frontier model labs like Anthropic and OpenAI, or the infrastructure beneath. If not, count your blessings and cherish your freedom. 

Now, I owe you the honest exception (although there are likely a few more), and pretending it doesn’t exist, fingers in my ears, doesn’t help. There’s a specific kind of company whose threat is real, but badly diagnosed. Ask yourself what your moat is really made of. If the answer is data, regulatory position, switching costs, deep domain expertise, none of that got cheaper to replicate. If your honest answer is “building good software is expensive, and we paid”, your moat got genuinely shallower than it was just a few years ago. You’re allowed to feel this.

But before you let cortisol overwhelm your brain, look at who’s eyeing that lowered wall: it’s not Jeff & Greg in Jeff’s parents’ garage who are allegedly now coming for your crown. Cheaper construction benefits, more than anyone, those who already have the distribution, capital & customer list. This is likelier to be an adjacent platform that could have always come for you, but just never bothered. They probably thought your segment was too small to justify the engineering cost, which just dropped. So what do you do about this, what’s an effective defense here? Write more code? Cut your team? You likely won’t outship them, because shipping fast & safely over time is a rare capability, one most of us have never even experienced, let alone built. You certainly won’t build it by hollowing out the thing you need to deepen your actual moat, the thing the playbook is burning to the ground: domain depth, relationships, judgement about what your niche really needs. 

So real threat, wrong medicine, and the medicine you were prescribed has real contraindications.

The playbook is shattering the layers required for innovation

We talk a lot about innovation — what makes companies generate it instead of stagnating. We talk about it at conferences and meetups, and we've accumulated a real body of evidence about what conditions it actually needs. The least comfortable part of that evidence is that the conditions are human: it comes down to people, and the culture that emerges from how they work together. The AI-native playbook I’m about to describe isn’t happening at every company, but at enough of them to show up in research and conversation alike, so if this isn’t you, I suspect you’ll recognize a peer.

Let's start at the bottom, at the thing everything else rests on. The term is "psychological safety" — the thing someone gave a presentation about at that offsite or meetup you went to a couple of years ago. Being nice is fine, nice even, but that's not what it's about. It's about the cost of saying "I don't know," "I was wrong," "this is broken," or "I need help." If that cost is non-zero — or worse, if it's high — you've lost the most important feedback loop you have: the one that tells your organization what's actually true about itself. The only reason you or your leadership team ever knows what's wrong is that someone felt safe enough to tell you.

And I need us to drop this comforting lie that there's a clean line between the personal and the professional — that doubt, fear, "I'm underwater," "I'm lost" belong on the home side of it, and what shows up to work is the finished, competent surface. Keep it professional. We may never say the words out loud. We don't have to. We hold it as an expectation and people read it off us exactly, because reading what leadership rewards is a survival skill. And notice what it actually forbids: "I don't know," "I was wrong," "this is broken," "I need help" are the personal crossing into the professional — every one of them is an admission of an internal state. So a norm that keeps feelings at the door isn't neutral. It's an instruction not to say the four things.

The playbook walks into this room with all these messy humans, and its defining move is urgency. More! With less! Right now! And it’s not just talk: people watch their friends get laid off and then struggle to find anything, in a way that belies how good they actually are. Say "more, with less" out loud even once, and the four sentences go up in smoke. We've now taught everyone that visible output — ever more, ever faster — is what's keeping them here, and that the long-term plan is to make them redundant. This new calculus doesn't allow for "I don't understand the code we're generating." It turns it into a fatal confession. So they shut up. They don't quit — they shut up and code, as many lines as possible. And it's fine, until it isn't. Until production is on fire and everyone's too afraid to put it out, or to risk explaining how it really got this bad.

What broke is the flow of information. We lost more than just morale: we lost our instrumentation.

But just receiving signal was never enough. What happens next depends very much on your tolerance for “slack”: unbooked time that isn’t already spoken for. It’s always the first casualty, because it’s treated as a nicety for times of peace, and, you see, we’re at war, remember? It’s the easiest cut to defend, because it looks like waste. Why would your team leave 30% of their time free? In this economy? You imagine an engineer staring out the window, bobbing along to the music on their noise-canceling headphones, or a long lunch on Friday. It’s fat to be trimmed. So we trim it and feel like decisive leaders.

But that’s not what slack is. Slack is the only reason anything that comes up ever gets acted on. It’s what lets people take that signal that something is wrong, figure out why it’s happening, and then solve it. It lets them go and read the code, add instrumentation, reproduce the bug and form hypotheses. It lets them sit with a drowning junior, and teach them why they got the estimate so wrong, rather than just eating the difference. Without slack, the signal still arrives, everybody’s just too busy to pick up the phone or listen to the backlog of voicemails. 

And this is the part the playbook gets infuriatingly backwards: it treats slack as the thing AI lets you eliminate. If the tools make us faster, why do we still need buffer time? The buffer was never compensating for slowness. It holds the capacity to respond to change, you know, that agility our industry talks so much about. Getting rid of it at the time you’re flooding the system with machine-generated code you don’t fully understand removes the mechanism to absorb something we know will happen: it will surprise you. 

You didn’t cut fat. It was never fat. You cut your organization’s immune system.

One more layer up you hit the thing the last 15ish years were supposed to be about. DevOps. It wasn’t tooling fashion or your Ops team’s new name, it named a binding constraint — the wall between people who write software and the people who run it — and did something brave: it tore it down and dared to imagine a future without it. You build it, you run it, you understand it, 3am, PagerDuty screaming. The shared understanding was the whole point.

It breaks my heart to see the wall being rebuilt.

Not intentionally, not explicitly, and not even completely. But when you combine fear, “more with less” and generated code, you all but guarantee it. The code, a deluge of it, arrives finished. Compiles, tests are green, plausible sounding Pull Request. This doesn’t invite “please, investigate me, grapple with me”, it invites “accept me”. I don’t see how anybody would still allow themselves curiosity about this, this already done thing. And even if they did, there’s no more slack for them to chase it. So they swallow it. The understanding never forms, and we’ll find some reason to make it their fault. Except we removed every layer that allowed them to form understanding.

This wall is worse than the old one, in some ways. At least the old wall had humans on both sides of it, trying crudely to throw things over or smoke-signal. This one? Nobody on the other side. AI understands nothing. It has no model of your system, your customers, your failure modes. It didn’t steep in your culture for years, forming a habitus, and it’s removed from the consequences. The humans on this side who are supposed to own this never build the understanding either — because you can’t read your way to comprehension through hundreds of thousands of lines arriving faster than you can think about any of them. Understanding a complex system isn't something you assemble from reviewing its guts at volume. It's something that forms slowly, in the building.

I don’t even care much for the “AI writes bad code” argument. It does, sometimes, but it won’t forever. And when it stops writing bad code, the problems will still be here. The harm has been done. And the annoying thing about systems as complex as our software, or the teams that produce them, is that they don’t just go back. You can repair them, people do every day, but repair isn’t restoration. The system that comes out the other side is a different system. 

We’ve had problems like this for a long time, usually around legacy code. It’s a tempting analogy to reach for: the people who wrote it are usually long gone, leaving behind a trail of commit messages and incomplete documentation. You’re piecing together intent into something, into some comprehension that once existed. You might think that actually, the breadcrumbs are richer and more regular, making recovery easier, but recovery of what? The trail ends before you find its origin: at acceptance. Acceptance of a prompt's output — text shaped to look like something a person who understood the problem would write, produced by something that understood nothing. It might resemble comprehension, it might not. You’ll never know. You’re not inheriting a system someone once intended to be a certain way and comprehended, you’re manufacturing a system nobody ever did. 

And now — slack cut, mouths shut, wall rebuilt — you ask these same people to look at what the machine spits out at astonishing rates, and judge whether it's right. How? There’s no signal, and even if there was, no time to grapple with it, and no mental model of the system under change. And here you ask us to sit: on a pile of rubble, with the trash raining down, sorting it nugget by nugget into something fit for purpose — and judging, with what's left of us, whether any of it is.

So we reach for help. Why wouldn't we? There's a tool right there that'll also do the judging for us, and we're so goddamn tired. And it's so confident! Always confident. It gives us back an answer, shaped like something we'd say ourselves. Whole, ready, asking to be accepted. And we accept. Not because we're lazy, but because all the things we need to do otherwise are gone. We stop checking. We stop overriding wrong answers with our own reasoning, which was the only thing we were still there to do.

What’s scary, to me anyway, is that it doesn’t feel like loss or surrender. It feels more decisive, cleaner, faster. And that’s the trap, and where I think we need to stop: we don’t lose or surrender screaming, we lose relieved.

This is what we’re overseeing. Some of us anyway. But it doesn’t have to be this way. There are other paths available to us, and I think if we’re honest about organizations, innovation and the nature of generative AI, one almost naturally shakes out.

Rebuilding, brick by boring brick.

So what’s left, when you stop choosing urgency? It’s almost disappointing in how familiar it is. 

The AI native company is barely about AI at all. The companies putting one technology at their center are missing the point the winners realized: organize around capabilities, the conditions that let people do good work. They treat generative AI the way you’d treat any genuinely new capability: as something in its genesis, uncharted, almost guaranteed to shake out in unpredictable ways. You can’t mandate or schedule or leaderboard your way there, because you don’t know its shape yet. By definition, you require slack. Companies treating AI as a productivity dial they can crank to eleven have made a grave mistake.

So you rebuild. Brick by boring brick. Five layers of them, in order: take fear off the table, rebuild safety, protect time, build understanding and cultivate judgement. Skip a layer and the ones below come loose. 

You have to start by taking fear off the table, out loud. Because that fear, that economic anxiety, is what’s pricing “I don’t understand this” above zero. It doesn’t even matter where it came from, and it doesn’t always even implicate us. Your team reads the same LinkedIn posts & Altman interviews you did, and by the time everybody’s done reading, everybody’s absorbed this new calculus: visible output is what’s keeping me here, and the unspoken plan is that I too will ultimately be replaced. You can’t build on this. The good news is, the organizations seeing the most gains from AI are the ones using it to make their people more capable, not cut them loose. There’s going to be a dip in productivity — adopting anything substantial causes one — but whether it’s a temporary dip or your new floor depends on whether you learn your way out. Gut the foundation, and that J-curve everyone’s budgeting for loses its rebound. What’s left is an L. 

Only with economic anxiety off the table can you start bringing safety back, and it can’t be as a motivational poster. It comes back when “I don’t understand this” stops being a fatal confession, and becomes signal the organization is grateful for. This is older than any of us, and we already know how to do it. On the Toyota line, any worker can halt the entire line by pulling a cord, and when they do, it makes them the person who caught the defect before it moved downstream, not the alarmist who needs to be managed out. 

A manufacturing plant assembling cars has different failure modes though, and the problem isn’t necessarily bugs in software, or a bent panel on a car. Those are easy to detect. What you’re really looking for is an artifact of how a factory and a software team are different: gaps in understanding. This makes it incredibly challenging, because the artifact might be flawless, remember: compiles, tests pass, great PR. The defect is the missing relationship between the code and the minds that own it. Absence, negative space. 

This is the cord to pull: “I can’t vouch for this”, not “this is broken”. The response should be the same though: stop the machine, and swarm. Not async, and not one team member carrying the cost of honesty. Shared understanding is owned by the team, not any one individual, and the team should fix it together. 

But what’s this swarm made of? It’s three to four people, dropping what they’re doing right now. Not later that afternoon, not next sprint, right now. To do an unplanned thing. That’s slack, the stuff we cut, remember? Place yourself in the metaphor, and imagine what it’s like to stand in a factory, pull the cord because you see a defect, and watch… nothing happening. You only need to see this once to realize the cord is decorative and that honesty is not in fact appreciated here. The safety you just bought at a premium quietly unwinds. All the while, code piles up faster than anyone can understand. This is inventory, depreciating in the queue. Skip a layer and you lose the ones already laid

So you have to give the time back. As policy, defended against wherever the pressure usually comes from. This is the single most boring brick in the wall. There’s no clever practice here to present at a conference, no creative name to put on a slide, there’s just you deciding that time spent understanding is legitimate work, even when, especially when end-of-quarter is in sight, because it always gets tight right around then. The playbook says this is theft from throughput. I say it’s the job.

If you’re unmoved by principles, let pragmatism move you: remember what this technology is right now. It’s genesis-stage, novel, shape unknown, gains uneven and ideas being iterated on and discarded as its unintended consequences come into view. You can’t book this in advance, you can only give your team the time & space to discover, which is what the companies actually seeing gains did. Fund unscheduled time in which your engineers can work out what the tools are for, what tradeoffs they’re making, get it wrong a few times before actual gains shake out the other side.

If this sounds new, you should start internalizing an important lesson: at scale, understanding your system cannot be an ambient side-effect of writing code that maybe, fingers crossed, happens along the way. It’s something you can deliver, you can make part of everything you deliver. The practices to do so — swarming, walkthroughs, architectural decision records, refactoring systems nobody understands anymore — are documented, researched, sitting in decades of papers ready for you to cite. Most companies lack the conditions to run these processes, not the practices themselves, because every single one of them costs time and effort and admits ignorance, and doesn’t survive first contact with fear-based cultures. You can’t expect individual contributors to do this on heroics alone when the environment is rigged against it. The toolkit is real. It becomes available to you when your foundation is rebuilt.

The last brick, judgement, is not like the others, because you can’t just lay it. There’s no pledge, no cord, no swarm that installs this ability: to look at the machine’s output and know whether it’s right. Not just technically correct, but right for users too. Judgement grows back on its own, if, and only if, the layers below hold for long enough, and if, only if, it’s used. Judgement is more like a muscle than a brick maybe: stop exercising and it atrophies. 

We don’t have to theorize and come up with analogies of our own though, because the most automated high-stakes industry on earth ran the experiment for us and called it: automate the routine work, turn the operator into a monitor of a system they no longer operate, and watch their skills decay, right up until the day the automation fails and hands them a problem more demanding than the routine work ever was. Pilots in increasingly automated cockpits flew less and less by hand, and when the autopilot disconnected, like it did on AirFrance 447, usually because something had already gone wrong, some of them could no longer fly the plane, let alone recover it from stalling.

The industry’s answer could not have been clearer, and it ain’t “just trust the machine bro”: mandated manual flight time, deliberate, in calm conditions, so the skills exist on days conditions aren’t calm. 

If you’re looking for a brick, that’s probably it: scheduled disuse. A day a week with no agents, with just people writing and debugging code raw, like God intended. With their hands. Not as nostalgia or some sort of deranged spa day, but to maintain the skills that allow us to judge, and keep us experiencing the actual friction in our systems. Imagine what happens when someone pulls the cord, and who shows up is not a team of experts, but a flock of monitors. You can’t build these skills when PagerDuty pings you at 3am, they have to be constantly, consciously manufactured. 

The same experiment, run twice.

Now, I can hear your objection already. This is all very nice, very humane… Very #thoughtleader of you even, but it doesn’t survive contact with our P&L. I get it, but you’re wrong. 

IKEA put AI on its customer service, and let it handle ~50% of inquiries. Millions of conversations, real operational savings, exactly the bet the faulty playbook promises. Except they did something different: the remaining demand, they treated as signal: customers want help designing a room, problems made of taste, judgement, and that weird corner in your apartment you’re unsure what to do with. They didn’t fire anyone. They retrained thousands of call center workers into interior design consultants, a now highly profitable paid service. This is what taking fear off the table looks like: not a values slide, but retraining budget and a deep commitment to the humans under your watch.

Klarna also ran the experiment. Same technology, their customer support bot worked too. Better even! It did the work of ~700 customer-support agents, the human kind. Outsourced too, which was convenient, because it meant nobody at Klarna had to fire anybody. Where IKEA saw the residual demand as signal, Klarna read it as a rounding error in their IPO story. Hiring frozen, workforce shrunk by 2000, CEO on every stage that’d have him bragging, asking OpenAI to let Klarna be their guinea pig. You know how the story ends, because everybody loves a good cautionary tale: satisfaction & quality dropped, and they’re now hiring new gig workers, expendable, no knowledge formation required. 

At its lowest point, engineers and designers were asked to answer support tickets, the organization cannibalizing what remained to improvise their way back to a capability previously discarded. They climbed out, but not to where they started. You never do.

The “AI doesn’t even work” argument loses steam here. The AI worked fine, better than one might expect even. It’s beside the point. What matters is the judgement applied by leadership: what each company decided its people were for, and what else they might be capable of. One found a business, the other a cost. You decide which of these is a few bricks short of a building.

None of this is new. That’s the whole damn point. We’ve run this exact experiment before, and it’s worth reminding ourselves how it went: the order-of-magnitude gap DevOps opened up didn’t separate the companies that adopted the tools from the ones that didn’t. It separated the companies that dared to live the cultural change from the ones that installed Jenkins, renamed Ops to SRE and called it a day. The gap didn’t just stay open for a decade, it opened wider.

Then too, some got it backwards and said that speed of adoption would be the deciding factor. It wasn’t, and it isn’t. Depth wins. Depth is slow, because culture is slow, and can’t just be dialed up to eleven. The companies sprinting away hardest, gutting layers to pay for the tools, optimized the thing that didn’t matter last time — output per head, the measure research buried many times over — starving the thing that did.

So no, you’re not AI native, no matter how hard you scream it in your pitch deck and career site. Screaming gave away the game: you put technology at the center and broke every layer it needed. The real ones don’t need the label. They’re just going to be doing what they’ve been doing the whole time: stopping the conveyor belt when someone says “I don’t understand this”, solving the problems they find, building on the skills they never lost. 

So, compressed: if you run the place, take fear off the table, out loud, and mean it. Then lay the layers in order, and don’t generate faster than your people can comprehend. If you don’t run the place, you can’t take fear off the table, but you can refuse to pass it down, defend what slack you control, and hand your team its cord. 

You don’t need to be afraid. You need to protect your layers. Not just when it’s cheap. 

References