Here is the pattern we keep seeing, across industries that have nothing to do with each other: AI lands hard where the problem is already well-defined, and it flatters everyone everywhere else. The gap between those two situations is where most founders are losing money right now.
Look at what just happened in wastewater treatment. Researchers studying AI-controlled aeration systems deployed across French treatment plants found meaningful reductions in energy consumption without degrading environmental outcomes. This is not a glamorous use case. Nobody's putting it on a pitch deck. But it worked precisely because the problem was narrow, the variables were measurable, and the feedback loop was fast. The system optimized something specific. The gains were real.
Now look at what's happening on the creative side. A rapper appears to have used an AI music generation tool to produce a song, then went quiet when the producer community started asking questions. The debate that followed wasn't about quality. It was about authorship, disclosure, and what "making" something even means anymore. That's what happens when you use AI in a domain where the problem is ambiguous by design. Music isn't a system to be optimized. It's an argument about identity, emotion and intent. AI doesn't clarify that argument. It intensifies it.
The enterprise software world is catching up to this distinction, slowly. Recent research on what's being called "agentic nesting" describes a framework for using AI agents to orchestrate the fragmented mess of enterprise systems: ERP here, CRM there, a pile of middleware nobody fully understands, data trapped in silos nobody wants to admit exist. The pitch is that agents can thread through all of it without the pain of traditional integration. And it's credible, but only to the degree that someone first mapped what the actual process should be. Agents don't invent the workflow. They execute it. If your process is unclear, you're not automating it. You're automating the confusion.
Tools like parallel AI coding agents are surfacing the same pressure in development work. The promise is real: multiple AI workers running simultaneously, checking each other's output, moving faster than any solo developer. We've experimented with setups like this. When the task is well-scoped, the speed gains are genuine. When it's not, you get five agents confidently building five different interpretations of a vague requirement. Faster output, worse results. The bottleneck was never the code. It was the clarity of thought upstream.
Even the food innovation story fits this frame. A company developing a cocoa alternative from faba bean hulls isn't chasing AI for the sake of it. They're solving a supply chain problem that has exact coordinates: cocoa is getting scarcer, prices are volatile, and a specific ingredient profile needs to be matched. That's a defined problem. AI can help optimize the formulation. It can't tell you whether people want to eat a faba-based chocolate bar. That's still a human bet.
So here's the thread: every one of these stories is really about the same mistake or the same success. The successes use AI where there's a closed loop. A measurable input. A measurable output. A clear definition of "better." The struggles happen where AI is deployed as a proxy for thinking that hasn't happened yet.
We watch founders do this constantly. They come to us wanting to "add AI" to their operations. When we ask what outcome they're optimizing for, the answer is usually something like "just make it smarter" or "reduce manual work." That's not a brief. That's a feeling. And while the feeling is usually correct, there's an actual problem underneath it that needs to be named before anything gets built.
The question is never "can AI do this?" It almost always can do something. The question is whether you've defined "this" precisely enough that the output can be evaluated. If you can't describe what success looks like in specific terms, you're not ready for AI. You're ready for a process conversation. And that's okay. That's actually where most of the money is hiding.
The wastewater system worked because engineers had decades of data on how aeration affects treatment outcomes. They knew what variables mattered. They built a model around a real, measured relationship between inputs and results. Nobody eyeballed it. The AI didn't add vision. It added speed and consistency to a framework that already existed.
That's what we help people build. Not the AI layer. The layer under it: the decisions that have to be made, the data that has to be captured, the process that has to be clear before automation adds anything except speed to the wrong direction. Sometimes that takes a week. Sometimes it takes three. But it's the work that actually determines whether your AI investment returns something or just looks good in a demo.
If you're sitting on a business that's working, hitting a wall, and considering an AI project, do this one thing first. Write down the exact decision you want AI to make, or the exact task you want it to handle, in one sentence. If the sentence is fuzzy, the project will be too. If the sentence is crisp, you're ready. Call us. We'll take it from there.
Defined beats dazzled. Every time. In sewage treatment, in music, in code, in supply chains. If the problem is clear, AI is a multiplier. If the problem is vague, AI is just a very expensive way to feel busy.