It’s 11 PM on a Wednesday and I’m sitting in front of two monitors. The left one has a terminal open, three agent waves deep into a Coverage Tower prototype, quality gates green across the board. The right one has the Insurity website, where I’m clicking through their new Pro Suite platform. The one they rebranded from Sure MGA last year with $50 million in R&D behind it. Trusted by seven of the top ten MGAs in the country.
It looks pretty good. Better than I expected, honestly. The UI is cleaner. The workflows make more sense. They’re clearly investing in the right places, and I find myself wondering what else they have in the pipeline that hasn’t been announced yet. Fifty million buys a lot of roadmap. Have they cracked AI yet? Their Borealis release from last month has “AI-enabled self-service” and a no-code ML workflow for predictive models. It’s the right direction. But I also know that MIT published a study last year finding that 95% of enterprise generative AI pilots fail to deliver measurable business impact. Forty-two percent of companies abandoned most of their AI initiatives in 2025 entirely. The gap between “we’re investing in AI” and “AI is changing how we operate” is a graveyard of press releases. I wonder which side of that line Insurity’s fifty million lands on.
I click through a few more pages, taking notes. Not anxious, exactly. More like the feeling you get when a sparring partner shows up in better shape than last time. You’re not scared. But you pay closer attention. You tighten up. The products that get better fastest are the ones with someone gaining on them, and good competition has a way of burning off everything that isn’t essential.
But the competition isn’t what keeps me up. It’s the question underneath it, the one that surfaces at 11 PM when the code is working and my brain has nothing left to distract itself with:
How do I get on their radar?
The engineers are easy. Engineers look at the architecture, the test coverage, the Rust proxy layer, and they get it. Tech people speak tech. That conversation I can have in my sleep.
The harder conversation is with the executive buying class. The CIOs, the COOs, the operations VPs who sign the contracts and approve the migrations. These are the people who decide which platform their company bets the next decade on, and they are not evaluating my Rust proxy. They’re evaluating whether this migration will make their life easier or become the project that gets them fired. Those are very different criteria.
And right now, those people are already using something. It might be running on a database architecture from 2003. But it’s theirs. They’ve spent years customizing it, building around it, integrating it so deeply that ripping it out would be like renovating the foundation of a building while everyone’s still working on the 40th floor. I used that exact metaphor two posts ago because I’ve literally watched it happen.
So the question isn’t whether I can build something better. I’m pretty sure I can. The question is whether I can walk into a room full of people who’ve never heard of me and make them believe the juice is worth the squeeze.
The Coming War
There’s something I keep circling back to, and I don’t love what it implies: software is becoming worthless.
Not worthless as in “nobody needs it.” Worthless as in “the barrier to creating it just collapsed.” Anyone with a terminal and an AI subscription can spin up a functional application in an afternoon. The thing that used to take a team of ten engineers six months to build can now be prototyped by one person before lunch. I know this because I’m doing it. Right now. In the terminal on my left monitor.
This will set off a tech war in every vertical. Insurance. Healthcare. Logistics. Finance. Every industry where incumbents were protected by the sheer difficulty of building alternatives is about to discover that the difficulty was the moat, and the moat is draining. New entrants will flood every market with platforms that are 80% as good, built in a fraction of the time, at a fraction of the cost.
Most buyers will stay with the devil they know. They’ll stay because switching is terrifying when your business depends on it, because the last migration was a nightmare, and because the incumbent’s sales rep takes you to a steak dinner and tells you what you want to hear. That’s easier than betting on someone new.
Meanwhile, actual wars are happening. The situation with Iran is escalating. The economy feels like it’s held together with duct tape and optimism. And here I am, at 11 PM, considering whether to throw away a career I spent 27 years building to chase a tech war that might not reward the best technology anyway.
Is building software even a moat anymore?
The Graveyard of Better Ideas
I’ve watched the answer play out enough times to know it sucks.
I was a Novell Certified NetWare Engineer. I ran those servers. I knew eDirectory inside and out: cross-platform, hierarchical, technically superior to anything Microsoft had by a margin that wasn’t even close. eDirectory ran on NetWare, Linux, Windows, and Solaris simultaneously. Active Directory ran on Windows. Period. And NDS shipped six years before Active Directory even existed.
Didn’t matter. Microsoft bypassed the technical staff entirely. They went straight to the executives. Sold Windows Server as a bundle: you get the OS, you get the directory service, you get Group Policy, you get the GUI that your CIO can demo at the board meeting. Novell charged separately for everything and marketed like they were still selling to sysadmins in server rooms. The resellers saw which way the wind was blowing, dropped their Novell certifications, re-trained as Microsoft MCSEs, and were encouraged to position NetWare as legacy technology. The people who understood the technology best were overruled by the people who signed the checks.
I watched it happen. I was in those server rooms.
Delphi was the same story, and I took it just as personally. I taught myself Object Pascal because AIM, the policy admin system I spent 15 years building around, was written in it. Delphi compiled to native executables, had real object-oriented programming, and was fast. Visual Basic was slow, interpreted, and limited. But VB lived inside the Microsoft ecosystem: it talked to Office, it talked to Access, it talked to SQL Server. Businesses didn’t pick VB because it was better. They picked it because it was already there.
Then Anders Hejlsberg, the architect who created Delphi, left Borland for Microsoft and built C#. The best compiler engineer in the world walked across the street because Borland couldn’t figure out where the money was. They chased enterprise middleware and CORBA while their core product withered. Delphi is technically still alive, maintained by Embarcadero. Good luck finding someone who knows it.
Sony’s Betamax had better video quality than VHS. Sony kept it proprietary. JVC licensed VHS to anyone who asked. Distribution beat quality. Xerox PARC invented the graphical user interface, the mouse, Ethernet, laser printing, and object-oriented programming. Then they failed to commercialize any of it. Apple saw it, shipped the Macintosh. Microsoft saw the Macintosh, shipped Windows. Xerox got credit in the history books and nothing else.
(Edison’s DC losing to AC is the one exception, and it’s instructive: AC won because it was cheaper at scale. When money and better technology are on the same side, better technology wins. That almost never happens.)
I’ve watched this pattern play out three times in my own career. The technology I was most excited about, the technology I was most proud of understanding, was the technology the market stopped caring about. The people writing the checks don’t understand what they’re buying. The people who do understand don’t get a vote. That gap is a graveyard.
So why bother?
The Math That Changed My Mind
Because the alternative is worse.
Jobs are already being eliminated by AI — not in some theoretical future, but right now, in companies I talk to. Every one of them is figuring out how to do more with fewer people, and the people getting cut first are the ones whose work can be reproduced by a prompt.
If I sit in a director’s chair and wait, someone younger, hungrier, and AI-native is going to figure out what I already know. They’ll build the platform I should have built. They’ll sell it to the executives I was too cautious to approach. And I’ll be the guy who had 27 years of domain expertise and did nothing with it while the industry shifted under his feet.
It’s not really a choice between safe and risky. The safe option stopped existing when AI started eating jobs. It’s more like: try to build something while you still can, or wait until someone else does. I looked at my LinkedIn and tried to think of a job title an AI couldn’t eventually do. I couldn’t think of one. That was a fun exercise at midnight.
And when I frame it that way, the math changes. A model can generate code. It can’t tell you that the last time someone tried this migration pattern at scale, it took down the quoting system for three days during renewal season. I can, because I was there. The time we migrated off OMNI’s flat files and I spent three weeks writing a data bridge that handled every edge case the vendor said was impossible — that’s what AI can’t do. Twenty-seven years of watching the same mistakes play out across enterprise software, of knowing which shortcuts will bite you in eighteen months and which ones are actually fine. That’s the part that’s getting more valuable, not less.
What JARVIS Actually Does
If experience is the thing that can’t be commoditized, the question becomes: what do you do with it?
There’s a scene in the first Iron Man movie. Tony is in his garage, working on a motorcycle engine. Not the suit. Not a weapon. Just a guy with grease on his hands doing detailed mechanical work. He asks JARVIS to pull up a holographic exploded view of the part he’s working on, and suddenly every component is floating in front of him in three dimensions. He can see how the pieces relate, rotate them, zoom into tolerances. But Tony is the one making the decisions. He knows which part is worn, which gasket needs replacing, how the assembly fits together. JARVIS gives him a better view. The engineering judgment is his.
That scene is not “AI fixes a motorcycle.” That scene is “an engineer, enhanced by AI, sees what he already understands in a way he couldn’t see before.”
Later, Tony changes the alloy composition of the suit and tells JARVIS to fabricate while he goes to a party. One parameter adjustment. JARVIS handles all the repercussions: recalculates structural tolerances, updates the fabrication specs, adjusts the power distribution, runs quality checks. Tony doesn’t micromanage the cascade. He designed the protocols ahead of time. The constraints, the feedback loops, the self-checking systems. All of that architecture existed before the parameter change. JARVIS executes within it.
That’s what I described in How We Build Software. The thousand-line spec. The quality gates. The constitutional constraints. The parallel agent waves with automated verification. I built the protocols. AI executes within them. When I change a requirement, the entire system adapts: new tasks generate, agents spin up, quality gates verify, and the output is constrained by rules I wrote before the first line of code was generated.
Tony designed systems that activate other systems. Feedback loops that trigger self-improvement analysis. Diagnostics that catch anomalies before they cascade. Protocols that activate protocols. That’s not science fiction. That’s /check-yourself. That’s the orchestration pipeline. That’s the quality gate that won’t let bad code through even when the AI insists it’s fine.
JARVIS didn’t replace Tony Stark. JARVIS made Tony’s decades of engineering worth more, not less. That’s the part everyone gets wrong about AI. The suit doesn’t work without the person inside it. And the person inside it doesn’t get superpowers without the suit.
Twenty-seven years of insurance domain knowledge. Fifteen years of building on AIM’s architecture. Every migration, every integration, every time I traced a bug through six layers of legacy systems to find the one stored procedure that someone modified in 2009 and forgot to document. That’s the person inside the suit.
And that’s when it hit me: this isn’t just how I build the platform. It’s what the platform should be.
We have an AI assistant called Iris. And I realized I want Iris to be the JARVIS of plcy.io. Not the chatbot that replaces the underwriter. The AI that makes the underwriter dangerous. An underwriter with 20 years of experience who can ask Iris to pull up a risk profile, cross-reference loss history across three carriers, and flag coverage gaps in a layered tower — that’s the motorcycle scene. The professional is still doing the work. Iris gives them the holographic exploded view. The same underwriter who can tell Iris to adjust a rate factor and have the entire quote cascade recalculate across layers, taxes, and commission splits — that’s the alloy change. The setup and protocols are already in place. Iris handles the repercussions.
The goal isn’t to replace the professionals who run this industry. It’s to make their experience worth more. To give them superpowers they didn’t have yesterday, so the 20-year underwriter can do in an afternoon what used to take a week, and still have the judgment that no model can replicate.
The Suit Fits
I think the executive buying class will come around eventually. I won’t out-market the incumbents or out-spend them — I’ll lose both of those fights. But the gap between what insurance operations need and what current platforms provide keeps getting wider every year. And I’ve spent 27 years standing on both sides of it, which is either my biggest advantage or the kind of thing you tell yourself to justify staying up until midnight.
The graveyard of better ideas is full of brilliant engineers who couldn’t explain to a CIO why any of it mattered. I’ve sat in those meetings. I’ve watched a COO’s eyes glaze over at “microservice architecture” and light up at “your team stops re-keying the same data into three systems.” That’s the translation work. That’s the part that isn’t becoming a commodity either.
I might be wrong about all of this. I might end up in the graveyard next to Novell and Betamax, another cautionary tale about a guy who built a better mousetrap and couldn’t figure out where to put the cheese. That possibility keeps me honest.
But I don’t think I’m wrong. And I’d rather ship it and find out than spend the next ten years wondering what would have happened if I’d closed the terminal and gone to bed.
The terminal is still open. The quality gates are still green. The suit fits.
I am Iron Man.
Subscribe via RSS, follow along at plcy.io, or bookmark this and come back when the war gets interesting. We’ll be here, building.