Designing for AI Features: Patterns for Uncertainty, Latency and Trust

Traditional interface design rests on a quiet assumption: the same input produces the same output, quickly. Click "Save" and the file saves. Every time, instantly.

AI features break that assumption three ways. The output varies between attempts. It often takes seconds rather than milliseconds. And it is sometimes confidently wrong. Designing well for AI means designing for all three, explicitly.

Pattern 1: Show work as it happens

A blank screen for eight seconds feels broken. The same eight seconds feels fine when the person can see progress.

• Stream text as it's generated rather than waiting for the full response.
• Narrate steps for longer tasks: "Reading 4 documents… Comparing prices… Drafting summary."
• Show partial results early — the first rows of a table, the outline of a document — so people can start evaluating.

Progress isn't decoration; it lets people decide early whether the output is heading somewhere useful, and stop it if not.

Pattern 2: Make stopping and redirecting easy

Because AI output can go off-track, the controls for steering it matter as much as the button that starts it.

• A visible Stop control while generation is running.
• Edit and resubmit on the original request, keeping the previous attempt for comparison.
• Quick refinement chips for common adjustments: shorter, more formal, add examples.

The best AI interfaces feel like collaborating with a fast assistant, not submitting a ticket and hoping.

Pattern 3: Put the source next to the claim

When an AI feature states facts, people need a way to check them without leaving their task.

• Inline citations that open the source passage, not just the source document.
• Highlighting which parts of a summary came from which input.
• Clear labels separating retrieved facts from generated suggestions.

Trust doesn't come from the model being right most of the time. It comes from people being able to verify cheaply when it matters.

Pattern 4: Draft, don't decide

For anything with consequences — sending an email, changing data, spending money — the AI should prepare and the person should commit.

A good default: AI output appears as a draft in an editable state, with the commit action clearly owned by the user. "Send," "Apply changes" and "Publish" stay human buttons. As confidence grows, you can offer automation as an explicit opt-in, per task type.

Pattern 5: Design the wrong answer

Every AI feature will produce bad output for some inputs. Decide in advance what happens then.

• Make regenerate and undo obvious and cheap.
• Offer a lightweight way to flag a bad result, ideally one click with an optional note.
• Write honest empty states for when the model can't help: say what it couldn't do and what the person can try instead.

Avoid apologetic, chatty error copy. "I couldn't find any invoices from March in this folder. Try another folder or upload the files." beats a paragraph of regret.

Pattern 6: Set expectations at the entry point

The input box shapes what people ask for. An empty field with "Ask anything" invites requests the feature can't handle.

• Use specific placeholder text that shows the feature's real strengths.
• Offer starter examples drawn from actual successful uses.
• State important limits plainly, close to the input: what data it can see, what it can't do.

A note on personality

It's tempting to give AI features a chatty persona. In most product contexts, restraint works better. The feature's voice should match the rest of your interface: clear, specific and calm. People are evaluating the output, not making a friend.

Checklist for your next AI feature

• What does the person see during the first two seconds?
• How do they stop, edit or retry?
• How can they verify a factual claim?
• Which actions require explicit human confirmation?
• What happens when the output is wrong or empty?
• Does the entry point set accurate expectations?

Answer those six questions before polishing anything else, and your AI feature will feel dependable even when the model isn't perfect.