Software Industry Revitalization

In response to: “It’s Over. AI Has Utterly Conquered Software Dev” — Medium, and “Starbucks Building AI-Powered Internal Software to Replace Microsoft and IBM Tools” — Bloomberg via MSN

Two articles crossed my desk this week that are really the same article. One is a Medium piece declaring that AI has “utterly conquered” software development — websites for $12 instead of $20K, apps for $1,000 instead of $250K, and, in the author’s words, “No tech teams needed. No AI engineer. No DB guy. No solutions architect. No UX/UI.” The other is a news story about Starbucks building its own AI-assisted software to replace inventory tracking it buys from Microsoft and maintenance management it buys from IBM. Starbucks spends about $400 million a year on software, and it has decided some of that money could be doing something better.

The Medium piece says the software industry is over. I think it’s looking at a demolition site and missing the construction crane. The software industry isn’t ending. It’s being taken down and rebuilt, and what’s being unlocked in the process is bigger than what’s being torn down.

Start with the part nobody should find surprising: people are going to build their own software. Companies are going to build their own software. And it will no longer be irresponsible to do so, no matter how many people say otherwise. This outcome was never really in doubt. It was baked in the moment APIs emerged, became securable, and became affordable to host. Once the interfaces to capability were open, priced by the sip, and defensible, “build your own” stopped being a question of whether and became a question of when. AI just answered the when.

For decades the responsible thing was to buy. Nobody got fired for buying IBM. Now a coffee company is building its own replacement for IBM’s software, and the remarkable part is how unremarkable it is — no moonshot lab, no acquisition, just a company deciding its own doing is back within reach. That’s not an industry dying. That’s a threshold being crossed, and thresholds like this one open more doors than they close.

There’s a lesson in the timing, though, and it’s worth taking cheerfully rather than bitterly. Companies had the chance, when cloud arrived, to rebuild — to shed legacy, decompose their operations into services, and reconstitute themselves around APIs. Mostly they lifted and shifted instead, and the ones that did the real rebuilding are positioned beautifully right now. Because none of what’s coming can be net new. A company is not a blank page. It has inventory that exists, suppliers with contracts, stores with histories. AI adoption in a real enterprise isn’t an act of creation; it’s an act of connection. Tools consuming well-defined APIs are precisely how you connect gradually — one workflow at a time, each new tool a small bet against a stable interface, instead of a bet-the-company rewrite. The API layer everyone was told to build fifteen years ago turns out to be the on-ramp to AI. If you built it, congratulations: your road is paved. If you didn’t, the good news is that the cost of paving it just collapsed. The rebuild that was too expensive to justify in 2015 is now the cheapest it has ever been, and AI is the reason. Late is no longer locked out. Late just means starting now.

And here’s the part I find genuinely exciting, the thing I hope people finally pay attention to. For thirty years the software industry has been in the business of making things that do. Vendors made things that do inventory, things that do maintenance, things that do customer service. Customers bought the things and arranged their companies around them — often contorting how they actually work to fit how the software insisted they work. The ability to do was industrialized, packaged, and rented back to the people who needed the doing done.

What’s changing is that the doing is coming home. People can now pay attention to the ability to really do — to shape software around how their business actually works — instead of buying yet another thing that does, built for a statistical average of a thousand companies, none of which is theirs. Every company that takes this seriously gets something it has never had: software that fits. Software as a means, not a landlord. That’s not the software industry shrinking. That’s the software industry multiplying — from a few thousand vendors to potentially every organization with something worth doing. I’d call that a revitalization.

Now, about “No tech teams needed. No AI engineer. No DB guy. No solutions architect. No UX/UI.”

There’s no doubt AI can do a lot of what those people do. Anyone who has watched a current model design a schema or lay out an interface knows the mechanical portion of those jobs is genuinely automatable. But it isn’t cliché to say that human judgment, imagination, and creativity are not forces easily replicated by technology alone — it’s just unfashionable this quarter. The DB guy was never valuable because he could write DDL. He was valuable because he knew which of your data actually mattered, which of it lied, and what would break three years from now if you modeled the world wrong today. That knowledge doesn’t evaporate when the typing gets automated. It gets promoted. The people who used to spend their days on the mechanical parts are exactly the people who can now spend their days on the parts that were always the real job.

This matters because connecting software bits naively — no judgment, no one who understands the whole — leads not just to screw-ups but, I think, to existential problems at least at a corporate level. Systems nobody understands, wired together by generated code nobody read, carrying assumptions nobody chose. The answer isn’t to slow down. The answer is to bring your people with you.

Which is why the right frame is AI as amplifier, not AI as automator that removes people. The automator story says: same work, fewer people, pocket the difference. It’s the small story, and companies that follow it will get exactly what they optimized for — a cheaper version of their current selves. The amplifier story says: same people, vastly more capability. Under that frame the limit stops being headcount or budget and becomes imagination — what can these people now attempt that was unthinkable before? A company that could never justify custom software for its supply quirks, its regional oddities, its one weird process that’s actually its competitive advantage — now it can. The backlog of things worth building was never small. It was just unaffordable. It isn’t anymore.

Imagination, plus one other thing: an Elon Musk-like focus on costs. Not cost-cutting in the usual sense, which is across-the-board and indiscriminate, but the more interesting discipline of asking why each cost exists at all. Most enterprise costs are archaeological — the residue of some constraint that stopped being true years ago. A license bought because building was once impossible. A process staffed because software was once rigid. Interrogate them and you find that a surprising number need not be accepted. The Starbucks CTO saying there are “clear opportunities to reduce the spend in software” is this discipline applied to one line item, and the beautiful part is what it funds: every dollar recovered from renting things-that-do is a dollar available for building the ability to do. The savings aren’t the prize. They’re the seed money.

One more thing, because the triumphalists deserve a precise observation rather than a dismissive one. The ability of AI to replicate human ability is, as far as I can tell, inversely proportional to the complexity of the thing being implemented. An application — a form, a report, a workflow, one team’s tool — sits well within reach, and the $12 website is real. But a system is not a big application, and a platform is not a big system. Complexity compounds. A platform is thousands of decisions that must cohere: failure modes, security boundaries, data lineage, the seams between domains, the exceptional cases that make up half of any real business. AI is unlikely to get all the nuanced facets of such a thing on its own, because many of those facets aren’t in any training data. They’re in the heads of people who’ve been burned.

Will that always be true? Maybe not. But notice what “it’s only a matter of time” actually asks you to do. It asks you to bet on the machine’s trajectory while shorting human ingenuity — to assume the people improve at nothing while the models improve at everything. Every previous wave of automation lost that bet, and lost it in the best possible way: the tools got better, the people got more ambitious, and the frontier moved instead of closing. I don’t consider shorting human ingenuity a wise bet. I’d rather hold both positions — better machines and bolder people — because that’s the portfolio that has paid out every single time.

So no, it’s not over. What’s over is the arrangement where the ability to do was something you could only rent. What’s beginning is better: a period when the cost of acting on an idea has fallen so far that the binding constraint is having ideas worth acting on. The demolition makes the headlines. The construction is the story. Software is going back to being a means rather than a toll booth, the doing is returning to the people who have things to do, and the only limit left — for those who take the amplifier view and bring their people along — is imagination.

That’s not the end of an industry. That’s the best thing that has happened to it in thirty years.