AI Can Build the App. The Hard Part Is Knowing What to Ask

Last night, I went down two different rabbit holes at two different ends of the AI-assisted development spectrum.

The first was Microsoft's new managed app building in Copilot Cowork. This is a brand-new Frontier capability, and it didn't go well at first. Three failed attempts to get an app running, $15 burned in Copilot Credits, eventually I gave up. By this morning whatever error I'd run into had cleared (and a video will be incoming on this new feature soon), but it gave me the evening back to work on something else.

Local AI is something I've barely touched, but I do have a backburner list of things I'd like to try. With an evening free, I picked one from the list: create a local AI pipeline to generate good quality transcripts from my video files, including speaker attribution for those with more than one voice. Within two hours I had a working solution. Point it at a folder of video files, get high-quality transcripts out the other end. This was well beyond my own knowledge of local implementation and Python, and it worked.

But it working wasn't the interesting part. On LinkedIn, "look what I built" posts are a dime a dozen. The interesting aspect was what made it work.

Generating a plan and code was easy for ChatGPT, and where its initial plan didn't quite work it was quick to give me better instructions. This was the easy bit. The harder part was to do with judgement. Choosing to push back to ask for stronger isolation from other services and data. Understanding that the first path ChatGPT took might work in isolation but needed to be improved for a holistic context. And knowing enough to read through the code and be reasonably confident there were no horrible red flags smuggled inside. None of that came from ChatGPT. It came from years of asking those questions on different projects in entirely unrelated contexts.

But this is the uncomfortable lesson of this type of development. On this project and with these tools, I feel like a novice. But I have to accept that compared to most business users, even out of my depth on something like this, I'm still an expert. I know what questions to ask, I know that "it runs" and "it's right" aren't the same thing, and I know when an approach just doesn't feel right. I could give the same tools to someone with none of this knowledge, and they can build equally impressive technical solutions that work incredibly, right up until they don't.

This is exactly the reason managed platforms like app building in Copilot Cowork need to exist and are probably the right solution for most businesses, even if they feel less powerful and end up a little more expensive. Our investment isn't in apps. It's in abstraction. Allowing users to worry about what they need to get done without feeling the pressure to learn all the things AI is supposed to just take care of, but so often just doesn't.

A future of software on demand is real. The question is which guardrails are the right ones to deliver it.

First posted on Linkedin on 09/13/2026 → View Linkedin Post Here