What the fuck am I supposed to build now?

In 2023, I wanted to put AI inside a VS Code extension. Someone told me nobody would give a random tool access to their private code, AI wasn't smart enough, and the idea was stupid. I listened. Calling it groundbreaking would be generous; Copilot had already been announced in 2021. Still, I had something I wanted to try, and I let someone else's certainty become my reason to stop. That part annoys me more than missing the idea.

Something similar happened while I was making an Instagram reel. Everyone was cutting videos to the same beats, so I started building a way to turn that pattern into a template. Then I saw Instagram doing it. I have no idea whether I was early or just unaware of what already existed. Either way, I kept collecting these little stories about things I could have built. Very convenient. An imaginary business never has to find a customer.

Now I have better tools, and every idea seems to come with five competitors and a GitHub repository. I can build more than I could back then. So can everyone else using these agents. That's the part I can't stop thinking about: what happens when the thing I've spent years learning becomes much easier to produce, but I still need someone to pay me for it?

There is evidence behind that anxiety.

Appfigures data published in RevenueCat's 2026 report puts new subscription-app launches at roughly 2,000 a month in January 2022 and over 14,700 in January 2026. In RevenueCat's observed new-app cohort, only 4.6% reached $10,000 in monthly revenue within two years. These are subscription-app figures, not the odds for every software business, and they don't tell us how much AI caused. But they do make “just ship something” feel like incomplete advice. Source: RevenueCat

Getting someone to try it isn't enough either. The same report found median twelve-month retention for monthly subscriptions was 6.1% for AI apps versus 9.5% for non-AI apps. That doesn't prove AI makes people leave. It does mean an impressive first experience and a subscription people keep paying for are different achievements. I can imagine building something that gets a lot of “holy shit” replies and quietly gets cancelled a month later. Source: RevenueCat

Tome is the uncomfortable example. Its co-founder Keith Peiris said they shut down a product with 25 million users because it wasn't “mission-critical.” It could make presentations, but it didn't understand customers and their businesses well enough to make the presentations they needed. Then there's Gamma, whose founder reported $100 million in annual recurring revenue and profitability in November 2025. Those are founder claims, not audited accounts, but I can't honestly use Tome as proof the category is dead while ignoring Gamma. Apparently, being able to generate slides doesn't settle whether you have a business. Tome founder's account · Gamma founder's account

I also keep asking why anyone would buy software when a free alternative exists. Which is slightly ridiculous coming from a developer who knows how much work can hide behind “just run it yourself.” Installation, hosting, updates, broken integrations, support. Someone has to deal with that shit. People can pay to hand it over. But if the free alternative becomes equally easy and reliable, that reason gets weaker. “People pay for convenience” is a possible answer. I still have to provide enough convenience to justify the price.

And then there's my actual job.

A client gives me a problem, I work with AI, give instructions, reject bad output, and deliver something usable. Some days I feel like a filter with an invoice. There is judgment involved, but I don't want to reassure myself that today's judgment will stay valuable forever. If a client can get the same result without me, explaining how carefully I prompted the model won't help.

Still, more generated code doesn't automatically mean more finished work. A 2026 working paper studying 718 firms estimated that agent adoption increased code production while lengthening the time between submitting and merging changes by 49%. That's elapsed time, not 49% more human review labor, and the study is observational. But it gives me a better question than counting how much code I personally write: can I take responsibility for getting something working in the customer's actual business? Source: Chen and Stratton

I think about selling to small businesses in India, then get stuck on language, onboarding, and explaining the technology. But why am I assuming the owner needs an AI education? Maybe they need fewer missed orders or less time retyping voice notes. CMR's 2025 survey of over 400 MSMEs found 62% sought digital advisory services. That isn't proof they'll buy my product. It does suggest the problem deserves more investigation than “there's a need but no market.” I'd have to show a useful result, in a language they're comfortable with, and find out whether they'd pay enough to cover the help they need. Source: CMR

I still feel late.

In college, I didn't have the money to keep trying tools and experimenting. Now I have access, and I keep looking at people who got further while I was catching up. That frustration is real. It also makes it very easy to turn every existing competitor into evidence that I missed my chance. I don't actually know what would have happened if I'd started earlier. I know I wanted to build something of my own, and I still do.

So maybe the next step is smaller than the question I've been torturing myself with. Find one recurring problem, get close enough to understand it, and see whether someone will pay for a solution and keep using it. Count what it costs to support them. I don't find that especially comforting. It means learning parts of running a business that I can't solve by opening my editor. But at least I could learn something from trying it. I've already spent enough time mourning companies I never built.