Not because AI doesn't work. It does. I use it every day to write code faster and save my clients money. But there's a difference between using AI and creating value with it, and most people building with AI right now haven't figured out where that line is.
Here's where AI genuinely earns its keep for me:
- Rapid code development. What used to take hours takes minutes.
- A second set of eyes. AI has flagged approaches I was about to take and explained why they'd cause problems later. That's real ROI: it's the bug you never had to fix, the rework that never happened.
But here's what AI can't do: understand the orchestration of a system built around a real customer's real needs.
Before I write a single line of Terraform, I have to know what the customer is actually trying to do. How will they use the application? Where will it break under pressure? How might their business, and their data, look different in two years?
Deciding whether a customer needs a data lake, a data warehouse, or an operational database isn't a prompt-engineering problem. That decision comes from listening to the customer and reading the cloud provider's documentation, not from asking a chatbot. Once I know what I'm building, that's when I bring AI in: to help me Terraform the resource, not to decide what the resource should be.
This is exactly where vibe coding falls apart
An inexperienced developer can prompt their way into an operational database when the business actually needed a data lake. It sounds like a small mistake. It isn't. Now you've got someone trying to manage an unmanageable volume of raw, unstructured data in a system that was never designed to hold it. At best, that's a painful migration. At worst, it's unrecoverable.
That's the risk of over-relying on AI: it will confidently give you an answer to the question you asked, even when you asked the wrong question. Human oversight isn't a nice-to-have here. It's the only thing standing between "it compiled" and "it actually solves the customer's problem."
Where AI still struggles
One area where I'll admit AI still struggles: complex frameworks.
The first time I implemented SSO across multiple Java web services (Microsoft login, AWS access, all wired through Spring Security), AI kept steering me deeper into the wrong hole. Eventually I went back to basics and pulled up Stack Overflow, because sometimes you need a human who's actually been in the trenches with a specific framework's quirks.
AI is a phenomenal accelerator. It is not a substitute for understanding your customer, your architecture, or your tools. The businesses that win with AI won't be the ones using it the most. They'll be the ones who know exactly when to trust it, and when to close the laptop and go read the docs.