How to design an AI feature
A good AI feature starts with a clear user job, honest limits, and a plan for when the model is wrong. Work through the six steps below as a design review or PRD checklist.
Start here
How to spec it
Define the user task
Name the job in one sentence. What does the user need to finish, and what would success look like without AI?
Clarify capabilities and limits
Say what the model can do, what it cannot do, and what data it sees. Show limits before people bump into them, not in a buried FAQ.
Design input and output affordances
Decide composer defaults, attachments, streaming, and how answers show up as text, artifacts, or cards. Match the UI to how risky the task is.
Add fallback and failure states
Design for wrong answers: degrade gracefully, hand off to a human, let people retry, and say when something failed. Happy path only is not enough.
Test with real users
Test mental models and trust calibration, not only task success. Watch where people over-trust the answer or give up.
Iterate on trust signals
Track correction rates, undos, and return visits. Add citations, checkpoints, or less autonomy where people keep guessing wrong.
Patterns to ship first
Streaming
Show replies token-by-token as they generate
Citations
Attach verifiable sources, quality signals, and claim previews
Human in the loop
Require human approval before AI acts
Failure Disclosure
Say clearly when the system cannot answer or a tool failed
Prompt Starters
Example prompts for empty states, templates, and libraries
Regeneration Carousel
Swipe or step between regenerated answers
Patterns by design step
Deeper build playbooks
Frequently asked questions
What is an AI feature?
Any product capability where a model generates, transforms, or acts on user input: chat replies, summaries, agents, copilots, or generative UI. Plain static screens without a model do not count.
What should I spec first?
Start with the user task and failure modes. Capabilities, composer UI, and trust patterns follow once you know what goes wrong when the model is wrong.
How do trust patterns fit in?
Citations, confidence, human checkpoints, and failure disclosure belong in the same spec as the core UX. They are not a polish pass after launch.
How do I test AI UX?
Test whether people calibrate trust correctly. Do they verify before acting? Do they know when to escalate? Task time alone misses over-reliance on wrong answers.
What are common AI feature design mistakes?
Hiding limits, auto-applying consequential actions, chat-only UI for structured tasks, and shipping without empty, loading, error, and low-confidence states.