He Called It 90% Done
I want to tell you about two projects. One is mine. One belongs to a client who called me back a week later, not to hire me, but to explain, very politely, why my work wouldn’t be necessary. He had discovered something better. A new AI tool that supposedly solved everything the others couldn’t. Something of a technological messiah.
That call told me more about the current moment in AI than anything I’ve read about it.
But let me start at the beginning.
My plugin
A while back I built a plugin using AI assistance. Vibe coded the whole thing. It worked. Actually, it worked well, clean functionality, solid output, exactly what I had envisioned.
My first thought was something close to “finally, this changes everything.”
And in some ways it did. The speed was real. The capability was real. I had built something genuinely functional in a fraction of the time traditional development would have required.
But then someone needed support.
I went back to the AI that helped me build it. Asked it to help me fix a specific behavior. It could address the symptom but had lost the thread of why certain portions of the code were written the way they were. The original decisions, the architectural logic, the purpose behind specific choices, that context had evaporated. A few weeks of conversation history and the project’s internal reasoning was gone.
The plugin works. It runs internally at Ei and does its job. But it will never be on a store shelf. It will never be sold to a customer who expects consistent support.
Here’s why that matters.
When you buy a Coca Cola, you know exactly what you’re getting. Same taste, same experience, same product every single time. That consistency is the product. It’s what you’re actually paying for.
A pizza I make at home is always good in some way. But it’s never exactly the same twice. Different factors, different results. I enjoy that. But I can’t sell it. The moment I charge someone for that pizza, they have a right to expect the same pizza next time.
AI-assisted development, in its current form, produces homemade pizza. Often excellent. Genuinely impressive. But not yet industry ready, always same quality Coca Cola.
My client’s project
A client came to me with an ambitious idea. Something like a segmented YouTube, a platform that would host his own videos and community content, allow community members to build courses, generate leads, and earn him commission from those courses.
Good idea. Real business logic behind it.
He told me Hostinger Horizons had already taken it 90% of the way there. He just needed me to handle the final tweaks.
I asked to see what 90% done looked like.
What I found was a beautiful homepage. Clean design, convincing layout, buttons for login and navigation in all the right places. It looked exactly like a product that was almost finished. There was zero functional code behind it. Not 90% done. Not anywhere near it. What existed was a visual surface, a Far West movie set with a convincing storefront facade and nothing behind the doors.
He was certain he could finish it himself. He started making adjustments. The AI that produced the first version couldn’t maintain coherence with the changes. What had been working stopped working. What he thought was a tweak broke the surface he’d started with.
I sent him a proper quotation for building the actual product. He thanked me and said he would analyze it.
Then a few days later he met Emergent.
The call I wasn’t surprised by
A week later he scheduled a meeting. He was gracious about it. Thanked me for my attention to his project. Then explained, carefully and kindly, why he would not be contracting Ei.
He had found Emergent, a tool that connects to Claude’s API and, according to him, solves the coherence problems that other AI tools apparently can’t handle. He was placing his faith and his money there instead.
I wished him well. I meant it genuinely.
But based on everything I know from building AI-assisted projects, from the plugin that works internally but can never be sold, from systems that look complete until they need to evolve, I can tell you what I expect will happen. At some point the project will reach a stage where what’s been built needs to be maintained, extended, or corrected in a way that requires understanding why specific decisions were made. And that understanding, that architectural memory, that coherence, will not be reliably there.
Not because the AI is bad. Because no AI tool available today can substitute for the structural thinking that happens before and during a real development process. The kind of thinking that asks not just “does this work now” but “will this behave consistently six months from now when something changes.”
What this is actually about
AI is not the problem. I use it every day. It has made me a more capable professional, faster and more productive than I could be without it. Business Sessions itself is built with AI as a genuine collaborative partner.
The problem is a specific kind of blindness that new technology tends to produce in its early adopters. The lights are bright and impressive and it becomes difficult to see clearly what they’re illuminating and what they’re leaving in the dark.
What AI does exceptionally well right now is generating functional outputs quickly, producing convincing surfaces, and accelerating early phases of almost any digital project. What it does not yet do reliably is maintain structural coherence as a system evolves, support a product through the inevitable changes that real-world use produces, or substitute for the planning and architectural judgment that determines whether something is actually buildable and supportable at scale.
The visual coherence of modern AI output is genuinely impressive and genuinely misleading at the same time. A homepage that looks like a product is not a product. A plugin that works today but can’t be reliably supported is not a commercial product. A platform built from a prompt with no architectural planning is not a platform.
The question worth asking
Before you invest time, money or confidence in an AI-generated digital product, ask one question.
Who is responsible for this when something needs to change?
Not who built it. Not which tool produced it. Who understands it well enough to maintain it, support it, extend it, and fix it when real-world conditions require something different from what was originally built?
If the answer is “the AI will handle it,” keep asking.
AI is a powerful tool in experienced hands. In those hands it genuinely accelerates what’s possible. But a tool is not a plan. And impressive output at 70 percent is not the same as a coherent, supportable product at 100.
The lights are bright. Look carefully at what they’re showing you.
Have you encountered this pattern in your own projects? I’d genuinely like to hear about it in the comments.
If you’re evaluating an AI-built product or considering an AI-assisted development project and want an honest assessment of where it actually stands, that’s a conversation I’m happy to have.