Why AI Features Should Not Be Owned by Engineering Alone

When a team decides to build an AI feature, the default move is to hand it to engineering. Please, engineering team, implement this prompt. Engineers are very good at the technical part. But they might be missing a lot of the nuances of context that actually make the AI intelligent.

That context is domain knowledge: the specifics of your product, what a good answer looks like, what you are expecting, what tone fits, whether it should be more playful or more serious. The person who holds that knowledge is usually not the engineer.

The complexity moved into the model

Building an AI feature means designing the behind-the-scenes instructions, the system prompt, so the model solves the problem the same way you would. You give the AI the same context you have in your head as an expert.

What separates a generic answer from an expert answer is this context, and it belongs to the people who understand the business and the problem. That is why the work of designing AI, both the prompts and the evaluation of the results, should include that expertise rather than defaulting straight to engineering.

You are the expert on your problem

To experiment with AI, you do not need technical expertise. You need a problem you are trying to solve, and an understanding of what good results look like for that problem. For the problem you are trying to solve, you are the expert.

This shows up clearly when non-technical people run their own experiments. In one cohort, a participant who studies knitting evaluated an AI that checks knitting patterns for mistakes, going row by row to judge whether each flagged issue was really an error. She could see things in the answers that a non-knitter could not, because she is the one who understands knitting patterns. Other participants worked on a fitness feature suggesting alternative exercises, a feature extracting structured information from voice notes, and an AI explainer for an NGO's course. None of them needed to be engineers to judge whether the output was meaningful.

What the domain expert actually does

The tool can add the structure of a prompt, the role, the task, the sections. What the expert brings is reading those instructions and assessing whether they are aligned with what they expect. The expert is the one who knows that a particular answer is wrong even when it looks confident, who notices the model misread a term, who decides what counts as a good response.

Lowering the technical barrier so domain experts can contribute to AI directly, instead of depending only on engineering teams, is the point. The rules of who can build AI changed when the complexity moved into the model. The expertise that makes it work has been sitting with product people, legal teams, and domain specialists the whole time.