The better an engineer is with AI, the more likely they say this: "With today's models, you just pick the few files you need and feed them in properly." They are right. For them. The problem is what comes next.
Objection 1. Can you feed it 100 million lines by hand?
A large enterprise codebase runs to hundreds of millions of lines. Hundreds of repositories, and most of the people who wrote them have left. "Picking the right context" assumes you already know what is relevant. That skill only works within what you know. But in a system of hundreds of millions of lines, even the most senior person knows a few percent of the whole at best. You cannot select code you do not know about, and the incidents happen in the parts you did not select.
Objection 2. Can everyone do what you do?
Grant, for the sake of argument, that developers will learn this. What about sales? What about marketing, support, and finance? They have to work with AI too, but will they have time to learn "how to pick the right code and paste it into a prompt"? A tool you have to study before you can use it is not an organizational tool. An organization's AI productivity is set by the average, or even the least skilled, not by the most skilled. Nobody has to learn how a generator works to use electricity.
Objection 3. Will you get the same result tomorrow, and at the next desk?
Hand-fed context ends with that session. Tomorrow you select it again, and the colleague next to you selects it from scratch. The same question gets different answers because each person feeds in different context. Repeat costs do not fall, and know-how does not accumulate as an asset. Individual skill is admirable, but individual skill is not infrastructure.
Objection 4. So where does every question end up?
If the dev team is the only group that knows how to pick the right context, every use of AI ends up routed through the dev team. The product manager's question, the CS check, the executive's curiosity all queue up in front of engineering. Sound familiar? That is why the bottleneck you had before adopting AI is still there after you adopt it.
You need layers, not more hands
The conclusion is simple. Context should not be selected by a person every time; the system should have it ready in advance. With an SSOT built from an analysis of the entire codebase, the AI answers on top of verified context whether or not the person asking knows how to select it. Expert hand-selection does not become unnecessary. It becomes the default for the entire organization.