A conventional product can often be described as a set of screens connected to a database. An AI product is more like a living system: a user gives context, a model interprets it, tools change the world, and feedback determines what the system should do next.
The interface is the visible tip
The most important decisions sit beneath the interface. What context does the model receive? Which sources can it trust? What actions is it allowed to take? How do we know when the answer is good enough? What happens when the model is uncertain?
These are architecture questions, not prompt polish. They shape the experience just as much as typography, navigation, or motion.
Choose the simplest useful pattern
Anthropic’s research on effective agents makes a useful distinction between workflows and agents. A workflow follows a defined path. An agent decides its own path. Both can be valuable, but they carry different costs in latency, reliability, observability, and failure recovery.
- Use a single model call when a well-scoped instruction is enough.
- Use retrieval when the product needs grounded, current context.
- Use a workflow when the steps are known and consistency matters.
- Use an agent only when flexible planning creates enough value to justify the uncertainty.
Design the trust loop
Every AI product needs a way to observe its own behavior. We define representative tasks, expected outcomes, failure categories, and a review loop before we ship. Evaluation is not a final QA gate. It is part of the product architecture.
The system should also show its boundaries. Users need to know what the product knows, what it is doing, when it needs confirmation, and how to correct it. Trust grows from legibility, not from pretending the model is infallible.
The new product brief
For AI products, a useful brief includes the user job, the context contract, the tool boundary, the evaluation plan, and the fallback behavior. The result is a product that can learn without becoming mysterious.