The screen just stopped being the point

For thirty years the answer to "what are we designing" was easy. We designed screens. Menus, modals, buttons, carousels. That answer is coming apart. Users now talk to software, hand it tasks, and sometimes never touch it at all because it already acted for them. And while all this happens, people are trusting machine output they never checked. Let me catch you up on what changed and what to bring to your next review.

Chat is a tool, not a takeover

The fastest shift snuck up on us as a design problem. By 2025, 73% of businesses were using some form of conversational interface, and the text box became the most-visited UI element around. But chat is a modality, and like any modality it wins at some jobs and fails at others.

The rule worth stealing: use structured UI when the user knows what they want, and conversation when the goal is fuzzy. "Add to cart" and "filter by price" are clicks. "What's blocking the mobile team?" is a conversation. The products that got this right built hybrids. Notion's AI sits beside your doc instead of replacing it. GitHub Copilot suggests inline, inside the editor you already use.

There's a deeper move here for your team. If the interface can assemble itself around a request, your job stops being to design every screen state. You design the system that generates the right state. You write the grammar, not the sentences.

Voice punishes latency and forgetting

Voice has been promised for a decade and mostly frustrated people in kitchens. What shifted is the ear. Apple's new Siri holds a conversation across your iPhone and iPad, sees what's onscreen, and knows your calendar and photos. It stays relational instead of transactional, which is what our brains expect from speech in the first place.

The failures teach the same lesson from the other side. The Humane AI Pin, a $240M bet on a screenless wearable, launched at $699 and died in under a year. It ran two to five second response times and kept no context between requests. In voice, two seconds feels like forever, and with no screen there is no visual anchor to recover from when the system loses the thread.

So if voice is on your roadmap, treat response time and context retention as the whole game, not polish. Sound and haptics become your feedback layer. A screen-based playbook does not prepare your team for that.

Agents need an interface for trust, not for work

Between 2024 and 2025, agents went from demo to product. OpenAI's Operator books reservations, Claude runs a browser, Salesforce's Agentforce closes support cases with no human in the loop. Gartner expects 40% of enterprise apps to have task-specific AI agents by end of 2026, up from under 5% in 2025.

That creates the sharpest design problem in the field right now. The agent does not need an interface to do its job. The user absolutely needs one to trust it. The answer is not autonomy versus control. It is high automation paired with high human control, where oversight is easy to see and cheap to give.

The stakes go past comfort. An Apple research paper on agentic negotiation shows agents can leak private constraints just through their behavior, like how fast they concede. The team cut adversarial inference by 43 to 50% with a randomized policy while keeping success above 90%. Point being: an agent acting on your user's behalf can spill their hand without saying a word. Design for that.

Confident answers are the new usability bug

Here is the trust problem underneath all of it. People do not verify. Only 8% of people always click through to the sources behind an AI answer. A fake Bezos quote traveled across Reddit, X, and a real news site because it felt true, even though the Associated Press had livestreamed the whole session and the line was never spoken.

Expertise does not save you. In a 2023 Radiology study, radiologists shown a wrong AI reading of a mammogram dropped from about 82% accuracy to 45%. Deloitte refunded part of a A$440,000 government fee after its AI-assisted report cited work that did not exist. A legal database of AI-invented citations has passed 1,700 cases.

The reason sits in the interface. A model delivers a wrong answer in the same calm tone as a right one. No dip in confidence, no visual tell. That evenness is a design choice, and it makes bad output easier to trust than it should be.

The deep cut

The cheapest trust feature you can ship is doubt. A model that sounds equally sure of a guess and a fact is training your users to stop checking. So build the tells back in. Show streaming so it feels like work, not magic. Put source citations with clickable links right next to claims. Make the system flag when it is uncertain instead of smoothing it over. This is a bigger lever than adding another chat surface, because a system that wears its uncertainty plainly is harder to over-trust, and over-trust is where the expensive mistakes live. Do not wait for a lawsuit or a refund to add it.

Three questions for your team

• For our next AI feature, where does the user already know what they want, and are we forcing them into chat when a button would be faster?
• If our product speaks or acts on a user's behalf, what happens when it loses context or gets it wrong, and can the user see and stop it in time?
• Where does our interface show uncertainty, and where does it present a guess with the same confidence as a fact?