AI doesn’t know what it doesn’t know.

The goal is to get enough evidence about the current workflow and its constraints to know whether a worthwhile problem exists. I should be able to explain what happens today, where it breaks down, who feels the pain, and what baseline I would use to judge improvement.

A real problem does not automatically justify an AI feature. I want to compare plausible interventions, make the tradeoffs explicit, and be willing to recommend a simpler approach or no build at all when the evidence points there.

I do not need an engineering curriculum, but I do need a working mental model of what the model sees, how information and tools reach it, where state lives, and what software enforces around it. That lets me separate model behavior from surrounding system behavior and have better conversations with engineers.

I want to define good behavior before launch, test representative and high-risk cases, and diagnose failures at the right layer. Reliability also means knowing what happens when the model, retrieval, or tools are wrong or unavailable.

For consequential work, authority should be visible: the model can interpret, software can enforce, and people can authorize. The system should make that boundary clear before something goes wrong and leave a path to inspect, challenge, and recover from decisions.

I want users to know what the AI is doing, what they still own, and what to do when the output is uncertain or wrong. Good AI UX should reduce cognitive load and make correction, approval, and recovery easier instead of hiding uncertainty behind a polished answer.

A working prototype does not create value if it does not fit the workflow or if people do not trust, understand, or own it. I want to diagnose adoption problems as product and operating signals, not assume more training is always the fix.

I want to define meaningful product outcomes, make sure the right data is captured, and use product analytics and experiments to understand what users are actually doing. I also want to use AI to accelerate analysis while being able to verify the data, queries, and conclusions. For AI products, that means connecting model and eval performance to normal product and business outcomes.

Client work starts with understanding the business problem, not selling a predetermined AI answer. I want to explain technical tradeoffs in plain language, surface missing evidence, manage expectations, and guide decisions without pretending uncertainty has disappeared.

The landscape matters when it changes what is feasible, reliable, economical, or easy to adopt. I want enough awareness of major shifts to update recommendations, while keeping the rest of the Library organized around durable product work.