In which everyone keeps asking who dies, which is the wrong question.
The Drawbridge Theory of Software Survival
The dropdown has seventeen options. One says “Other (see Bob)” except Bob isn't a current employee; few people know who Bob even was, and from some timestamps you can deduce he left before the pandemic. Nobody has removed "Other (see Bob)" because nobody is entirely sure what would happen. Pebble thrown into a pond, or tsunami?
From inside a company, inherited software looks something like this: throw the pebble, or hold onto it. I've stared at structurally important flows and important-but-mysterious scripts that work, like wiring behind the walls, but then I choose to leave them for another day when there's another pressing matter, or I'm not quite sure how they got there, or if the outcome is close to what we need done today.
Say you're new to a company, and you've just been handed some problem that leads you to a permission group called “TEMP_FINANCE_ACCESS” (temporary since at least 2017), and that thing is hooked into two integrations: “Salesforce_Bridge_OLD” and “Salesforce_Bridge_NEW.” And the authors who knew OLD and NEW were consultants or a friend of someone or are now long lost to company history. Both integrations are running and highly active. You have no idea why, or how. What does AI do with this? Well, at least today, it asks you: "What is TEMP_FINANCE_ACCESS?" I don't know, bud, that's why we're talking.
And now, like that asteroid that wiped out the dinosaurs, we read AI is heading toward software, but it's really heading toward a lot of situations like the above. The asteroid might be a reach of an analogy, except: everyone is already talking about which companies are the dinosaurs of today, and which are the mammals that will flourish after the dust cloud and coming ice age. Which software companies are the walking dead? Which will survive? Why? How? What even are the switching costs anymore? Somebody produces a quadrant.
I will read the quadrants and often agree with them. But my mind inexorably returns to Bob.
People deciding whether to replace some piece of software are primarily concerned with whether the pond just ripples, and generally, AI will have to solve that in every company, in every TEMP and OLD and NEW situation. And it might be able to, sometime soon.
None of this means the old software is particularly good, or that the Bob dropdown is a moat in the Warren Buffett sense. TEMP_FINANCE_ACCESS is not company or product strategy; it's 100% a workaround that worked and kept working.
People accumulate systems that just “work” even if they don't "work well." Think about, let's say, payroll processing. Once paychecks are successfully deposited without issue, payroll period after payroll period, why rip it out and put in another one? Price, I guess. But payroll processors have a job to do and do it. There's a lot of that kind of stuff out there, software running quietly in the background because every person who might have replaced it with something else did some loose math that has about four repeating variables: finite human time, political capital, calendar space, and emotional bandwidth. And a lot of times the equation decides: do something else instead.
The dinosaurs continue to eat the grass or hunt or fly around. That payroll processor that's still working? Other systems have attached themselves to it like barnacles.
Okay, time for an abrupt shift away from dinosaurs and crustaceans to castles.
Let's look at what makes a castle work: ramparts, food, water, storage, horses, somebody who knows how to fix a drawbridge, supply lines and someone who keeps their head on how much food and weaponry are needed, a few different someones with a deep understanding of the surrounding countryside. Another someone who has an understanding of why the castle is there in the first place, why it had to be built. Also, there's an old but important tapestry in the corner, but no one really knows anything about its history anymore.
To me, all of that's a lot closer to software and what it does than the typical Warren Buffett-esque “moat” language. An attacker has to cross a castle moat to get to the place where people live, then figure out the real moat: what was even going on in that castle?
AI obviously introduces changes to a lot of what-we-thought-to-be-true about how people will buy and build software, and some of what it will change is fairly obvious already. Here's what we're all grappling with today: building some competitive thing gets a lot cheaper, because now everyone's a pseudo-software-engineer. And it's now inarguably easier to leave any given software platform: you can point your AI friend at some public API docs and it will fully ingest them faster than you, understand them more deeply in ways you could never. I need to move this data from here to over here. What was previously a daunting migration is now a few scripts helpfully written for you (or at least, the robot is holding your hand). All of this is true today.
The castle walls might be a lot lower to the ground now. And yet someone is still standing inside the walls of TEMP_FINANCE_ACCESS asking why both OLD and NEW are running.
Castles are fun to think about, but I'm not that interested in the walls after a few minutes of admiring the scale. Walls are up or they aren't, but let's imagine they are fine. You walk by the walls: completely fine. Nobody is currently assaulting it or getting over it and the people on the ramparts are secretly drinking something, barely watching what they're supposed to watch. And yet the castle can still go bad because people leave, or food stops arriving, or the roads around it stop mattering. Somebody inside opens a gate, or the reason anybody wanted to live there in the first place has quietly disappeared. None of those really require anybody to defeat the wall anyways. But, for some reason, we tend to talk about software disruption through this much more cinematic version: a new thing arrives, that new thing is somehow vastly better, and it easily defeats the castle walls.
Once you look for them, the boring versions of castle defeat are easier to recognize. A system can remain paid for and officially important while less work actually happens there. The best new tools start somewhere else. Your developers stop caring as much. A promised rebuild keeps everybody hanging on because help is apparently just over that ridge over there. Somebody brings in another product to solve one annoying problem, then a workflow moves, then another person asks for access. On their own, none of these "events" matters all that much.
Until all together, you see the stacking of one small state-change on the other and then another. One day, you look up from your desk, discover that a meaningful amount of life has moved somewhere else, even while the castle remains there, impressive and fully intact.
AI makes all of this a lot stranger because it can participate in almost every one of those processes, and I'm sure some new ones we haven't encountered yet. Increasingly the thing crossing the castle boundary isn't even a person; your drawbridge now has a lot of users that don't have feet.
Software has always had versions of a bits-and-bytes drawbridge, all some form of access to what's inside (API, partnerships, etc.). And now AI makes the question of "what even is a drawbridge anymore" a lot more interesting because a lot of the new products don't need to breach the castle or even interact with the castle in a way humans ever had to.
Which creates a sliding-scale decision for whoever thinks they own that particular castle: lower the drawbridge, or keep it raised, and decide how far in which direction. Either way, eventually everybody inside starts wondering why the castle is such a bad place to live. Castles do occasionally want people and things to move from the outside to the inside. People have to eat, and the turnip stores are getting a little moldy.
And meanwhile there is the postern gate, the little side entrance for who knows what essential medieval stuff that needs medieval doing.
We're dealing in mights here, but if I were to bet today, this is the most AI-ish path into the castle because a new LLM tool doesn't necessarily arrive at the gate asking to speak to the Lord or whoever. It does the small job you just stumbled on: summarize this something, clean this something in this way, answer this specific question I just discovered I'm battling (losing), and I don't want a wandering answer (need this fixed pronto, bud) or a million clarifying questions. Are these software decisions being made? Or is it one small job that begets another job. And another. And the decision is making itself.
The main drawbridge: up. That little door on the side? Wide open, so convenient, and the traffic going through it is regular enough to prompt you to think: who or what exactly is in the castle?
At this point in the castle metaphor, it's fraying and I know it.
There are other ways to think about the same question. Less tortured metaphors. Dan Hockenmaier's recent essay "The Software Shakeout" is in my opinion the best version of the "who's going to die?" genre. A clean question, presented well, but I keep getting hung up on the part where the model meets pre-pandemic Bob.
To be fair, Bob is already definitely in that essay somewhere, probably in the parts on workflow tax and business disruption risk, both of which are clearly and obviously real. It's just so hard, from outside any company, to get at the actual amount of either thing. Which makes me suspicious of any switching-cost number that claims much precision across the whole set of human organizations. The walls are measurable, but Versailles isn't Edinburgh Castle. I've a hard time at this point even knowing which party (company buying software, software company, AI company) is the castle.
Epic is an example I can't get away from. They've pissed off practically everyone, including the government, but are just seemingly never going anywhere.
Hospitals are about as castle-y as software gets. Huge amounts of data in weird shapes and formats (all 10+ years old probably), absolutely terrified of switching, regulatory constraints, workflows that have been Roman Army trained into thousands of people, and Bobs everywhere you look.
And yet somehow I've filled out the same medical intake information by hand, in 2026, for a provider whose health system already had all of it. Which is a strange experience if you're simultaneously wondering how this is a good system. The data is already floating around inside it, but somehow obscure to the intake form you're filling out once again.
Clearly my address accumulates somewhere here, and that is valuable?
I've been very confident about dying castles before, so I should introduce some caution. A lot of castles are still standing today. Sometime around 2015 I was pretty convinced Microsoft was on a long glide toward irrelevance. Everything about it seemed old and staid, from a different technological age. Windows, Office, machines that would endlessly fall into the blue screen of death. Everyone desperate to get on Apple devices.
Time has revealed this to be an extremely bad take. Microsoft found enough new, meaningful reasons for people to live there, and I think about that whenever I look at some enormous incumbent today. Sometimes the old castle has an absurd amount of land and it's just too hard to storm.
So my guess is that this all ends up probably somewhere messy, and we'll continue to make our 2x2s to explain the mess as best we can.
What will happen? Some castles: stormed. Some other castles: slowly emptied while everyone continues to pay for them, until they stop. Some postern gates: the main source of foot traffic in and out. And some Microsoft-esque things which everyone points at with a bit too much certainty will end up owning half the countryside.
Of course, people eventually stopped organizing much of their lives around castles at all.
Which is an irritating thought to arrive at after spending this much time thinking about them. Castles are obsolete in real life, and it's possible the castle metaphor is similarly obsolete, and I haven't admitted it yet.
Footnotes
Before LLMs, this wasn’t a hypothetical. It still maybe isn't, but perhaps that's changing before our eyes.
In every enterprise migration plan I’ve ever seen, at the outset there’s always a section labeled something like "Dependencies" that begins with so much confidence. Until it starts to lose steam and ends with a list of items that are blank, marked TBD, or maybe question marks, or maybe it's a person's name because something had to go there.
The "SaaSpocalypse." The "AI Scare Trade." All just framing that sounds scary as we all stare at bloodred charts. Before our eyes, an entire sector reprices overnight. But we also all know human institutions rarely change that fast. Human institutions are human. Slow.
We stare collectively at all those red charts like astrologers. Mercury is in retrograde, this month will not be good for you.
SaaS as a model clearly faces some impending structural pressures, and investors can just take their money and park it elsewhere. Wait it out and see who gets through the crucible. How do you price in the unknown?
You make the easiest available call: reprice the whole category and wait it out.
Or maybe attached barnacles is the wrong idea, and it's more like sediment mounding underwater. Sediment sounds negative, but it's the cost of moving around in the world. Stuff has to get done, and then little bits of organizational sand or silt or whatever accumulates as real work gets done under real constraints by people (usually) trying to do the right thing. It's not laziness.
Anyway, the incumbent (usually) benefits from sediment.
Writing about AI is everywhere, but the writing I can find on AI and organizations combines two questions that deserve their own treatment: What can AI actually do? And once an organization finds out those capabilities exist, what do organizations do with them?
These are obviously related, but they run on two different clocks.
For most of the modern software era, engineering talent has been the biggest constraint on product velocity. It's why every single advice podcast was learn to code for about ten years.
But now we have tools that can take a short written request and return something that works...close enough? Here are some good write-ups explaining we've crossed some threshold and institutions that don't engage will be blindsided. Yegge / Appleton / Shumer.
I find their arguments pretty good. Compelling. But I think the second question doesn't get enough attention, as it's the slower clock, the messier clock, and it gets abstracted or hand-waved away right when it starts to matter.
"You can do more with the software because AI makes it more productive." This is the margin expansion angle. Then there's the angle like catnip for everybody: category displacement. "AI Agents doing work don't buy seats." They don't sit in a chair. It's something like an alien workforce suddenly becoming available to all of us.
Transitions of any kind are always like catnip because "settled" questions are suddenly re-opened and all the definitions might shift or change completely. What even is a seat anymore?
Maybe the first story is bullish for today's incumbents, and maybe the second isn't. Or maybe a few of those incumbents might just like a little category displacement to turn into something new.
Due to "intelligence tools," Jack Dorsey and team reduced Block's headcount by 40% in February 2026. Is it a steely-eyed CEO reorganizing for the future instead of waiting for it, or a convenient narrative for a company that bet hard on crypto and didn't quite see that pan out? Uh... how can we know?
Choose your fighter! I don't know which it was, but the idea isn't laughable on its face. It's bold for sure. I think, though, it won't be how most places do it.
Organizations have their own metabolisms, and how many are willing to go on the Block emergency juice cleanse?
The frontal assault has a cousin: the internal build. Or: when the organization decides to stop buying altogether and rolls up its sleeves. The build-versus-buy debate is now into multiple decades, and even in those areas where it's been more or less settled on "buy," AI has kicked the door open.
If generating functional software gets cheap enough, why buy at all? Because there's a tradeoff to a custom tool someone spins up because vibes. That tool can become a thing nobody fully understands six months to two years later. The build option is often faster, definitely more tailored, but it (1) only relocates the maintenance problem, and usually somewhere with less visibility and fewer people paid to care deeply about it working, and (2) now you have something so custom you lose out on the vendor's expertise on how to generally do that thing.
You have to decide to trade vendor lock-in for something without a good name, like internal lock-in.
The thing about internal lock-in is that there's no transactional relationship with the person/team that built the thing; you have to see them tomorrow, no matter if your feelings are good, bad, or ugly. But you are mostly fine with this because you also get something tailored; those people know what's going on in your company way more than the vendor designing for the average or median user across all companies ever could.
And there's one other thing about the internal build. The larger the company, the larger the standardization pressure. A 10,000-person company will never let individual teams quickly choose their own tools because in that setting, governance moves slowly and conservatively. By design. The idea that Amazon would let each team independently build its own Workday replacement is comical. A joke at odds with how large organizations make software decisions.
AI agents don't have eyes, and if humans remain involved in the long run, the advantage likely goes to whoever can treat their product less like a traditional "humans only" application (pretty, well designed) and more like a place where agents, humans, both just pass through. The jobs get done.
It's a weird headstate to try to imagine what this means. But a lot of what today's companies think is valuable or defensible is probably not.
The Office of the National Coordinator for Health Information Technology. I used to be very interested in electronic medical records. Then I lost interest, in large part thanks to Epic.
Organizations are collections of people with motivated reasoning: careers, incentives, and cultures obsessed with their own importance.
So how, if you're one of those companies with all that human stuff swirling around, do you survive this coming platform shift? Even looking at everything that came before, most answers we have are ex post facto. Just spitballin here, but I think you have to have at least one of two things in abundance: steely-eyed, clear honesty about what your customers actually need today versus whatever it was your company has been doing all these years, or a practically insurmountable distribution advantage that gives you room to develop that honesty and figure out what to do with it before the market tells you you aren't doing it fast enough.
In 2015, I would have predicted Microsoft would have faded into irrelevance. Here we are.
| First published | February 2026 |
|---|---|
| Reading time | 17 min |
| Tags | ai, automation |
Reply
I’d welcome your thoughts on this essay. Send me a note →
