Choosing an AI model can feel like the central decision in a new system. Teams compare model names, study benchmark results, measure context windows, and review pricing sheets before deciding what to build. But the real challenge often begins after that choice.
A New Stack article by Son Nguyen examines how human oversight is shifting from writing code to defining requirements in AI development. Nguyen is CTO of Orient Software, co-founded Neurond AI, and is a Forbes Councils Member. His article was published on September 17, 2026, at 08:00am EDT.
The model question can take over the conversation
Nguyen describes a question teams often ask when they begin planning an AI system: “Should we use Claude Fable 5, Gemini 3.7 Flash or DeepSeek-V4-Flash-Vision-Exp?” That question leads teams into debates over benchmarks, context windows, and pricing sheets.
Those comparisons matter, but they can pull attention toward the model itself. The discussion becomes focused on which option looks strongest on paper, even when the larger requirements for the AI system have not been defined. The model becomes the center of the plan before the team has settled what the system needs to do.
Nguyen puts that experience plainly: “We chased that same question early on. But after years of building AI systems in the real world, the model is rarely the whole story.”
That point changes the role of human oversight. Instead of focusing only on writing code or selecting a model, teams must define the requirements that shape the entire AI system. The work is not limited to identifying a powerful model. It also depends on knowing what the system is meant to accomplish.
Requirements shape what gets built
The shift from code to requirements does not remove technical work. It changes where human judgment matters most. Teams still face choices about models, benchmarks, context windows, and pricing, but those choices need to serve a clearly defined system rather than stand alone.
When requirements remain unclear, a team can choose what looks like the best model and still fail to move forward. Nguyen has seen teams pick the best model on paper and then stall for six months. The stalled project shows why a model decision cannot replace a clear direction for the AI system.
The same contrast works in the other direction. Nguyen has also seen teams build something transformative. The difference, as presented in his article, is not described through one winning model name. It rests on the broader work of building AI systems in the real world and understanding that the model is rarely the whole story.
This puts requirements at the heart of human oversight. People must decide what the system should deliver, what problem it is meant to address, and how the selected model fits that purpose. Those decisions guide the project before code turns the plan into a working system.
Human judgment moves upstream
As AI development changes, human involvement moves upstream in the process. The most important contribution may happen before implementation begins, when teams decide what they are building and what success means for that system.
That does not make model comparisons irrelevant. Claude Fable 5, Gemini 3.7 Flash, and DeepSeek-V4-Flash-Vision-Exp remain part of the conversation Nguyen describes. Benchmarks, context windows, and pricing sheets still form part of the decision. But those details cannot answer every question about an AI system.
The larger lesson is simple: a model is one part of a real-world AI project. Teams that treat it as the whole story can spend six months stalled after choosing the best option on paper. Teams that define their requirements and understand how the system is designed can build something transformative.
For human oversight, that means the key task is no longer limited to producing code. It includes setting the requirements that give the code and the model a clear purpose. The model may start the discussion, but the requirements determine what the team is trying to build.
Based on
