Chat earned its moment. It's the fastest way to demo a model and the cheapest way to ship one. But a chat box quietly hands the user three jobs the interface should be doing: figure out what to ask, figure out how to ask it, and figure out whether to believe what comes back.
The AI products people actually keep don't make users do all three. They use a small set of interaction patterns that do the work on-screen. None of them are exotic. All of them are stealable.
Seven patterns worth stealing
| Pattern | What it does | You've seen it in |
|---|---|---|
| Receipts on everything | Sources, reasoning, and an honest confidence signal attached to output — believing doesn't require faith | Perplexity's citations |
| Ghost text | The AI suggests inline, dismissible with one keystroke; the user stays the author | GitHub Copilot |
| Preview, then commit | Show what will change before it changes; keep one-step undo after | Photoshop's Generative Fill on a new layer |
| Visible work | Streaming and progress that names what's happening ("reading your March invoices…") instead of a silent spinner | Every agent product worth trusting |
| Co-creation loops | Treat the first output as a draft to refine together, not a verdict — most users edit the AI's first attempt anyway | AI writing tools people actually reuse |
| The ledger | Background work leaves a trail: what the agent did, why, and how to reverse it | Activity feeds and audit trails in agent tools |
| Earned autonomy | Start with suggestions the user approves; let the AI do more alone as wins accumulate | "Review and approve" flows that graduate to automation |
Notice what they share: every one makes the AI's work easier to judge. That's not a coincidence — judging the work is the user's real job now. (The Gulfs Have Flipped makes the full argument.)
The anti-patterns — naming them helps
- Mystery AI. Features that pop up unpredictably, look different on every screen, and can't be invoked on purpose. Users can't build a habit around a ghost.
- Confidence theater. A confidence badge that changes nothing. If "low confidence" doesn't change what the interface asks of the user, it's decoration.
- Sparkle-washing. An ✨ on every feature, value on none. Users learned this tell fast, and it now reads as a warning label.
- The 3,000-word answer. Dumping everything the model knows on someone who asked a small question. You don't look smart; you look noisy.
Try this — the pattern audit
List every AI surface in your product. Next to each, write the pattern it uses. Two findings are common, and both are fixable:
- If every line says "chat," you shipped a demo. Take the highest-traffic surface and rebuild it around receipts, previews, or ghost text.
- If the patterns differ screen to screen — approval here, silent overwrite there — users have to relearn trust in every corner of your product. Turn the patterns into shared components with one behavior. (Trust Is Not a Feeling covers how.)
Go deeper: why patterns beat one-off cleverness
Patterns turn trust into a component problem, and component problems scale. When confidence, preview, and undo are reusable pieces with consistent rules, every new feature inherits trustworthy behavior instead of reinventing it — and the coding agents helping you build can be instructed once, in writing, instead of corrected feature by feature. One-off clever interactions demo well and drift immediately.
Chat will keep its place — for the jobs where conversation is the point. For everything else, the interface is finally growing up.