I keep running into the same pattern in AI products: you ask a question, the model answers the literal question, and nothing else happens.
No follow-up. No retained context. No sense that the system actually understood what you were after.
The Problem With Feature-First AI
When a team adds AI as a feature — a button that calls an LLM, a sidebar that summarizes things — they're solving the wrong problem. The question isn't "can we put AI in here?" It's "what would this product understand if it were truly intelligent?"
There's a difference between:
- A search bar backed by GPT-4
- A system that remembers what you were looking for last time, infers what you actually need, and retrieves the right context before you even finish typing
The first is a feature. The second is a product.
What "Woven In" Actually Means
When I build with RAG and LLMs, the goal is to make intelligence part of the architecture — not a layer you can peel off.
That means:
- Context is first-class. Every user action is an opportunity to build a richer model of what they need.
- Retrieval is intentional. Not "search the database," but "find the three documents most relevant to this user's current task."
- The LLM is a reasoning layer, not the whole product. It interprets and generates. The system around it does the heavy lifting.
The Signal That Something Is Different
You know AI is woven in — not bolted on — when removing it would break the product, not just reduce it.
That's the bar I build toward.