AI UX WEEKLY
Week of July 14, 2026
7 stories · curated for designers
As agents get more capable and more embedded in daily tools, the most important design decisions are not about what the AI can do, but about where it pauses, what it shows before it acts, and how much control the user feels at every step.
The stories
This Week in AI Products
 Vercel Agent |
Jul 8 |
Vercel puts an autonomous agent directly inside the production dashboard
Vercel Agent now lives in your project dashboard, where it can investigate production issues, answer questions about your deployments, and propose fixes for your approval before acting. It pulls from logs, metrics, and deployment history to find root causes autonomously, but requires a human sign-off before taking action. The explicit approval gate is doing real design work here.
| “ |
Borrow Vercel's 'propose before act' pattern for any AI feature in your product that touches irreversible or high-stakes actions. Sketch the approval gate screen specifically: what does the agent say it wants to do, what context does it show to justify that action, and what does the user actually confirm before anything changes? — Designer's Takeaway |
 Nielsen Norman Group |
Jul 10 |
NNG names 5 qualities that separate trustworthy chatbots from frustrating ones
Nielsen Norman Group published research identifying five qualities that define effective site-specific AI chatbots: handoff willingness, flexibility, proactivity, emotional responsiveness, and transparency. The framework is based on observed user behavior across chatbot interactions and is intended to help teams evaluate and prioritize chatbot design decisions. Each quality maps to a specific failure mode that erodes user trust when missing.
| “ |
Run your current chatbot against all five qualities and score each one honestly. Whichever quality scores lowest is your next sprint priority. If emotional responsiveness is the gap, that probably means designing specific response variants for frustrated or confused users, not just tweaking copy. — Designer's Takeaway |
| ChatGPT |
Jul 9 |
OpenAI ships ChatGPT Work, a long-horizon agent that takes action across apps and files
OpenAI launched ChatGPT Work, an agent mode that can take actions across a user's apps and files, stay engaged on a single project for hours, and turn a stated goal into finished output. Unlike previous ChatGPT features, it is explicitly framed around multi-step, extended work sessions rather than single-turn responses. This represents a meaningful shift in how OpenAI is positioning ChatGPT: less as a chat interface and more as an autonomous collaborator for complex tasks.
| “ |
Audit your own product's AI features for long-running tasks. Map out three specific states: what the UI shows while the agent is working, how a user gracefully interrupts mid-task, and what a readable summary of completed actions looks like when the agent finishes. If any of those three states are undefined, design them before shipping. — Designer's Takeaway |
| CX Dive |
Jul 10 |
Third-party AI tools are outperforming brand-built chatbots in customer service
A new report from CX Dive finds that general-purpose AI tools like Claude and ChatGPT are beating company-built chatbots at resolving customer service issues. The gap comes down to language quality, flexibility, and the ability to handle unexpected requests without breaking. Brand chatbots built on rigid decision trees are losing users to tools that feel more like conversations.
| “ |
Run 10 real user queries through your chatbot this month and note every deflection, loop, or canned response. Then run the same queries through a general-purpose model and compare the results. Where the general model wins, the design problem is almost always over-constrained intent matching, and that is a flow change you can propose, not just an engineering ask. — Designer's Takeaway |
| Smashing Magazine |
Jul 9 |
A new framework argues mental health app designers should resist common UI trends
Smashing Magazine published a piece introducing an evaluation framework specifically for mental health app UI. The argument is that patterns designed to capture attention or signal modernity often work against the core needs of distressed users: reducing cognitive load, building trust, and providing a sense of calm and refuge. The framework gives designers a concrete lens for deciding when a popular pattern is inappropriate for their specific user context.
| “ |
Before shipping a trendy interaction pattern in any health, wellness, or safety-critical product, ask two questions: does this reduce cognitive strain, and does it build trust? If the honest answer to either is 'not really,' that is a strong signal to reach for a simpler, more familiar pattern instead, even if it feels less interesting to ship. — Designer's Takeaway |
 Figma |
Jul 9 |
Figma built a hidden accessibility layer to bring screen readers to its canvas
Figma's canvas renderer bypasses the browser's native accessibility tree entirely, which meant screen reader and keyboard users were locked out. To fix this, the team built a synthetic Mirror DOM: an internal accessibility tree that generates per-layer summaries and mirrors them into hidden but spatially-positioned DOM elements that stay in sync with canvas selection. The system now covers Prototype viewer and is expanding across the product.
| “ |
Figma's team had to invent a parallel structure just to meet baseline accessibility needs on a custom canvas. Audit your own product for any custom rendering surfaces where the browser's native accessibility tree gets skipped entirely. Even if you are not building the fix yourself, naming these gaps clearly in a ticket is design work that moves things forward. — Designer's Takeaway |
| Apple Siri |
Jul 6 |
iOS 27 lets users tune Siri's pace and expressivity for the first time
Apple's latest iOS 27 beta adds controls that let users adjust how fast Siri speaks and how emotionally expressive it sounds. It's part of Apple's larger push to rebuild Siri on generative AI and make the assistant feel less robotic and more personal. Users can now dial in a voice character that suits them rather than accepting whatever Apple ships as the default.
| “ |
If you are designing any AI voice feature, add pace and expressivity controls to your next design review. These are no longer nice-to-haves. As Apple ships them at OS level, users will start expecting to tune these settings everywhere. Sketch where those controls live in your settings flow and what the default state should be for your specific user context. — Designer's Takeaway |
Steal this week
Vercel Agent's Propose-before-act approval gate
Vercel's agent doesn't just act, it surfaces a plain-language summary of what it wants to do and waits for explicit sign-off before touching anything. This single pattern does more trust-building work than any amount of animation or reassuring copy could. Any product shipping an AI feature that writes, deletes, sends, or modifies something on a user's behalf should copy this immediately.
Pattern deep-dive
Human-in-the-Loop
Multiple stories this week, from Vercel's approval gate to ChatGPT Work's long-running agent sessions to the CX Dive report on chatbot failure, all pointed to the same underlying tension: AI can act, but users need a clear moment where they stay in control. The pattern kept surfacing because agentic AI is moving fast and the products that feel trustworthy are the ones that have deliberately designed the handoff point, not just the action itself.
When to use it: Apply human-in-the-loop any time your AI feature takes an action that is difficult or impossible to reverse, touches data or systems outside your product, runs for more than a few seconds without visible progress, or operates on behalf of the user without a clear confirmation step. The more autonomous the agent, the more deliberate the approval moment needs to be.
Stop shipping AI slop
Audit your AI design against 36 patterns
Drop a screenshot, get specific gaps and a Claude Code prompt to fix them. Free, no signup for the first audit.
Audit your design →