The Upkeep After the Build

The Allure of Building Your Own

Eli: Nate Calloway runs his own business, and in 2026 he decided to stop paying for most of the software it runs on. Instead of a stack of monthly subscriptions, he switched to self-hosted open-source models, running the software on his own machines rather than through a cloud service.

Mara: Which is the question a listener sent in. As AI coding grows up, from quick prototypes into production-quality software, will businesses and consumers stop paying for expensive SaaS subscriptions, that's software rented by the month, and start building their own tools instead? Today's angle is the bespoke software paradox.

Eli: The bet rests on a new kind of tool. A coding agent doesn't just suggest the next line. It writes the code, runs it and fixes its own errors.

Mara: Which means someone who's never written code can now ship something that works.

Eli: And plenty are. Social Media Examiner profiled a woman who isn't a software engineer. In a couple of months, it reported, she replaced more than twelve hundred dollars a year in subscriptions. She does it on a hundred-dollar-a-month Claude Max plan.

Mara: A hundred a month is twelve hundred a year. Same figure, other column.

Eli: Fair. But a whole trade is growing around these swaps. The Staffless Business laid out a typical case: four subscriptions replaced by one focused build. That build costs about fourteen thousand dollars, used-car money. On their numbers, it pays for itself in under two years.

Mara: And that payback assumes four thousand a year for hosting, monitoring, AI usage and maintenance. Every year, with no end date. The software company ServiceTitan says it plainly: AI can generate a working app in minutes, but it can't maintain it or secure it. The build is a one-time bill, Eli. The upkeep never sends a last one.

The Productivity Trap

Eli: Those small jobs have bigger cousins. Greenleaf Management, an Atlanta real estate investment manager, replaced Salesforce, the big platform for customer records, with an app built on Replit and Claude Code. A partner there told The Information it saves them around a hundred thousand dollars a year.

Mara: And that's no point solution, a tool for one narrow job. That's the core.

Eli: Then a giant. Sanofi, the French drugmaker, is cutting eighty percent of its ServiceNow use, The Information reported. The work goes through AI agents built with Claude Code and Cursor instead. The aim is to save at least ten million dollars a year.

Mara: Eighty percent, not a hundred. Some of ServiceNow stays.

Eli: And months before those reports, investors had already panicked. In the first week of February 2026, they fled software stocks. Forrester summed up the fears. SaaS companies might not be the providers of choice for AI agents. Pricing per seat, per human user, could go obsolete. And vibe coding, building software just by describing it to an AI, might let startups copy complex platforms and erode their moat. One reading of cases like Greenleaf runs straight through that panic. Garry Tan is among those who see severe disruption coming, because companies can spin up software fitted to their exact workflows at a fraction of vendor prices.

Mara: The other reading looks at what comes after the build, Eli. Sebastian Siemiatkowski is among those who expect SaaS to stay dominant, because bespoke software needs compliance, security audits, hosting and continuous integration that AI alone can't handle cheaply. And the studies on developers are where that reading looks.

Eli: The pitch is always that AI makes coders faster.

Mara: It makes them feel faster. In a 2025 study, developers expected AI to speed them up, and afterward believed it had. The stopwatch said they were nineteen percent slower. And it isn't one study. A pair of researchers, in work TechRadar reported, found developers using AI turned out thirty percent more code, but no statistically significant gain in finished software. Their report notes humans are still heavily involved in code review, an engineer reading each change before it ships. One analysis argues reading is the step nobody has automated, so review, not generation, is the real bottleneck in 2026. That claim is disputed.

Eli: And the rework lands on senior people. In a 2026 survey Hanover Research ran for New Relic, tech leaders described heavy rework. About three in four said at least a quarter of AI-written code needed significant fixing after it went live. Nate Calloway had no senior engineer to call. He had himself.

The Reality of Custom Code

Mara: Now add the rules a business lives by. Many enterprise buyers won't pick a software vendor without a SOC 2 report. SOC 2 Type Two is an independent audit confirming a company really protects its customers' data. Building that compliance from scratch can top a hundred and fifty thousand dollars. And it can take up to eighteen months.

Eli: So a homemade tool needs its own audit before a big customer touches it.

Mara: And agents make the audit harder. A SOC 2 guide from Blaxel notes that audits assume a control, once tested, stays put, while AI agents drift as they take in new data. In a 2026 survey for Salt Security, nearly half of organizations said they'd confirmed or suspected an agent doing something nobody authorized. Its chief executive argues that security grants permission once and checks it at the door, which only works when the thing let in is predictable. An agent isn't.

Eli: Nate Calloway had none of that machinery. Just himself and Claude. And the first month went badly. By his own count, his productivity fell by roughly thirty percent. Claude sometimes hit rate limits, capping how many requests he could send before being told to wait. Its answers varied from one ask to the next, and now and then it misunderstood him. Every handoff between tools was manual. All of it took real mental effort.

Mara: Close to a third of his output, gone in a month.

Eli: It's the review problem from those studies, with exactly one reviewer. Then something deeper. He wrote, 'An AI assistant is, fundamentally, stateless between sessions.' It carries no memory from one conversation to the next. So it couldn't serve as his CRM, his client database. And his clients, he found, valued a real human connection. He couldn't replace Linear, HubSpot, Figma or his Loom walkthroughs with AI, and he switched back quickly. His line on it: 'Claude isn't a tech stack.'

Mara: Researchers haven't shown future agents can handle upkeep, compliance and security patches alone.

Eli: And Nate was only touching the light stuff. Big companies run on a system of record, the master file where a business keeps its history, the version every team trusts. A Forbes analysis says replacing one is still expensive and risky. It holds the historical data, sits inside a dense web of connections to other systems, and supports audit requirements. At the other end sits software meant to die young. TechCrunch found people building tiny apps for one household or one trip. One man built a web game for his family over the holidays and shut it down when the vacation ended.

Mara: But personal apps like these, it noted, may have bugs or critical security flaws, so they can't just be sold as-is. Platforms like Google's AI Studio now build in safeguards against harmful content.

Eli: A holiday game can vanish when the holiday ends, Mara. A client database has to survive every audit after it, and that's the weight any company going it alone would have to carry.

Why SaaS Endures

Mara: The February panic assumed any company could pull off what Klarna did.

Eli: The company held up as proof was Klarna. It replaced enterprise tools like Salesforce and Workday. But in 2025 its chief executive, Sebastian Siemiatkowski, said he was 'tremendously embarrassed' that his first comments were so widely misread. Klarna hadn't swapped Salesforce for an AI model. He said it built an internal tech stack on Neo4j, a graph database, to bring its data together, alongside lighter SaaS tools. And he doubted every company could copy that.

Mara: Most companies keep buying, because a custom build can't easily match a vendor's reliability, security, compliance and upkeep. SAP calls building an agent the tip of the iceberg. Maintaining, supervising and evolving it is where most of the effort lies.

Eli: That's Nate's month, written large. The quick wins are micro-apps: small, single-purpose programs made fast for one job, like that holiday game. The hard part was everything after. Memory. Checking. Upkeep. And since homemade apps can carry bugs and security holes, each one is another thing someone has to look after. That's the work a subscription pays someone else to do.

Mara: So, the answer. People will build narrow tools of their own, but the subscriptions underneath mostly stay. Evidence indicates that while AI drastically lowers initial build friction, recurring cognitive and operational maintenance costs continue to protect core collaborative SaaS platforms from total replacement. What grows is bespoke edge tools, custom pieces built around the rented core.

Eli: So the company rents the filing cabinet and builds its own gadgets around it.

Mara: And the gadgets can be thrown away. The cabinet can't. What would change that: multi-year renewal rates and defect logs on custom code, showing whether agents can patch and maintain it without a human. Whether companies churn through homemade apps faster than through SaaS contracts. And how many consumer-built apps on the app stores last past their first few weeks.

Eli: Until those numbers come in, Nate's experiment ends where it started: back on Linear, HubSpot and Figma.

Mara: Read up on coding agent.

Eli: And follow technology with us.

Mara: And that's the story. I'm Mara.

Eli: And I'm Eli. Thanks for listening, and until next time.