AI has fundamentally shifted how people think about buy vs. build. The cost of writing code dropped so dramatically that building your own solution feels almost free now. Why pay for a SaaS when you can prompt your way to a working prototype in an afternoon?
For a certain class of problems, this instinct is correct. But for production systems - the ones your business actually runs on - I think we’re heading toward a massive and expensive overcorrection.
A personal example: this blog
When I started building this site, I heavily used AI (being an AWS Hero I obviously used Kiro) from the first line. “Build me a personal website with a blog section.” Simple enough. It worked.
But over the following days and weeks, I kept realizing what I’d missed. I wanted analytics to understand how many people visit the site. I’d forgotten to specify an error page. I needed clean, permanent URLs with a proper slug per blog post. I wanted a workflow that makes writing and publishing easy.
Each of these gaps required experience from past projects to even recognize. AI helped me implement every single one of them - but the vision, the knowing-what-to-ask-for, came from me. Not from the model.
And here’s the thing: this is just my blog. A system where I can do whatever I want, with zero dependencies on other people. No SLAs, no users waiting, no compliance requirements. If I break it, nobody cares except me. And even here, the gap between “working prototype” and “thoughtfully built system” was significant.
Where build is genuinely the right call
There’s a category of work where AI-assisted building is a clear win: the things that used to live in Excel spreadsheets with macros, Access databases, or internal tools held together with duct tape. “Citizen developers” have always built these. AI just makes them better - more capable, better UX, fewer guardrails of legacy tools holding you back.
I recently heard someone complain that their company is now flooded with AI-generated HTML pages containing data analyses. The data goes stale because it’s baked into a static file, and people keep sharing these pages around. The person was frustrated about it.
But here’s the thing: this is exactly what happened before with Excel sheets. People have always emailed around spreadsheets with outdated numbers. The work isn’t getting worse because of AI - it’s actually getting better, because the output is more tailored to its purpose. The stale data problem? That’s not new. AI didn’t introduce it.
The key insight: AI is a capable tool for making these things better. You just have to be careful not to introduce new problems with it. In this specific zone - internal tools, analyses, personal utilities - I see AI as a genuine positive.
For this class of problems: build. Use AI. Make something better than what existed before. This is great, and it opens up new possibilities.
The line most people don’t see
But there’s a hard boundary between “internal tool that helps a team” and “production system that runs business processes.” The moment other people depend on your system, everything changes.
You need uptime guarantees. Security audits. Data governance. Multi-user permissions. Scalability under unexpected load. Disaster recovery. Audit logging. Zero-downtime deployments. Data migrations.
These aren’t features you can bolt on later. They require upfront architectural decisions - the kind that come from experience, not from prompts. And this is exactly where I see the most dangerous naivety: people assuming AI can handle all of this because it can produce code that looks like it handles it.
The maintenance illusion
AI can write code. It cannot yet own the consequences of that code in production.
What happens when something breaks at 3am? When a security vulnerability needs patching across interdependent services? When a regulatory requirement changes and you need an audit trail you never built? When you need to migrate your data model without downtime?
These moments require experience, institutional knowledge, and operational maturity. None of which were planned for because “AI built it in a weekend.”
And here’s the pattern I keep seeing: when a requirement was missed and needs to be added retroactively, AI tends to rearchitect from scratch rather than evolve the system incrementally. Throw away and rebuild. Over and over. Each iteration wasting time, money, and resources - while a bought solution would have handled it out of the box from day one.
What AI actually does well
Looking back at my own experience building this blog - where AI truly helped was the concrete implementation. Writing the code. Setting up the infrastructure. Generating the CSS. Every step was faster because of AI.
But it wasn’t the vision. It wasn’t the product thinking. It wasn’t the “what should this become in six months” question. Those came from experience. From having built and operated systems before. From knowing what I’d regret not planning for.
The uncomfortable conclusion
For production systems - the ones with users, with business processes depending on them, with compliance requirements - “buy” is still the right answer more often than people want to admit right now.
The SaaS you pay for isn’t just code. It’s a team on-call at 3am. It’s compliance certifications maintained year after year. It’s thousands of edge cases handled over a decade. It’s a business model that keeps all of that running and evolving.
That’s hard to replicate with a prompt. And it’s expensive to learn the hard way.
AI is an extraordinary tool for execution. But it is not a replacement for experience, product thinking, or the hard-won knowledge of what production systems actually demand. Knowing when to build and when to buy still requires the one thing AI can’t provide: judgment shaped by years of getting it wrong first.
And yes, this post was written with the help of AI. It served as a sparring partner - someone to discuss and sharpen ideas with. The thoughts are mine, but AI helped me bring them across in a way that works for an audience beyond my own head. I hope that’s value added here without introducing a new problem.