Essay · October 1, 2025
Notes on AI design
A working set of notes on what I've learned designing for AI features inside enterprise products: what holds up, what falls apart, and what I'd push back on next time.
Placeholder content. Replace with the real essay when ready; frontmatter and routing already work.
Why I'm writing this down
Most "AI design" writing I read is either evangelism or pattern-library posting. Neither has been useful for the work I actually do, which is shipping AI features inside an existing enterprise product where the bar for trust is very high and the cost of being wrong is real money.
A few starting claims
- AI features are data products first, UI second. If the model isn't grounded in something the user already trusts (their data, their history, their entities), no amount of UI craft will rescue it.
- Confidence is not a UI element. Showing a percentage next to an answer is almost always worse than reshaping the interaction so the user only ever sees high-confidence outputs.
- Reversibility beats accuracy. A wrong answer the user can undo in two seconds is better than a right answer that they have to verify for two minutes.
- The design question is "where in the workflow does this earn its place?", not "what should the prompt look like?"
What I want to flesh out next
- Patterns for grounding AI in structured business data without leaking it.
- When to use a model and when to use a query.
- What it actually takes to design for review, audit, and rollback.
- The handoff between AI suggestions and the user's existing system of record.
(Drafting in public; expect this to change.)