AI features lose users when people cannot understand what they do, correct mistakes, or see useful results quickly.
2
Trust grows when applications expose sources, label generated content, show action plans, and let users approve or stop agentic work.
3
AI interfaces should guide people with examples, suggested prompts, undo and version history, permissions, cost estimates, and actions that connect results to the rest of their workflow.
Summary
Kathryn Grayson Nanz argues that AI-powered software has a user experience problem because developers understand concepts that most users have not learned yet. She compares the current state of AI interfaces with early Macintosh software, which used familiar metaphors and gradually introduced more technical language as users became comfortable. AI features should build on existing patterns while explaining where those patterns stop applying. Nanz groups the main design challenges into trust, clarity, control, transparency, and meaningful benefit. She recommends citations, visible action plans, AI labels, streaming responses, clear progress signals, prominent stop controls, undo and version history, granular permissions, memory controls, cost estimates, examples, templates, suggested prompts, and next-step integrations. Her argument is practical: users need to understand what the system is doing, retain authority over it, and get useful work done with its output. The model alone is not enough to make an application people will use.
AI creates a wider knowledge gap between developers and users
Nanz says AI combines unpredictable output with unfamiliar interaction patterns. Its quality, format, style, and accuracy can vary even when developers try to prevent problems. Terms such as prompt engineering, hallucinations, and retrieval-augmented generation are familiar to people working with AI, but most users do not share that vocabulary. When users try a poorly explained AI feature and get a result that differs from what they expected, they become negative about the feature and may stop using it. Nanz says creators have only so many chances to make the experience work before users disengage.
Early Macintosh design offers a model for introducing unfamiliar technology
Nanz compares AI interfaces with the Macintosh graphical user interface. Early Macintosh software borrowed language and flows from everyday life, used icons and visual representations, and taught people how to use personal computers. As users became more comfortable, later versions used more technical terms and fewer literal abstractions. She places current AI experiences around 'system 3' in this metaphor. Users have existing mental models from other software, but those models do not map perfectly to AI. Developers should reuse familiar patterns, combine them carefully, and introduce new behaviors gradually.
AI chat interfaces need familiar patterns with AI-specific controls
An AI chat can borrow the structure of direct messaging, including messages on opposite sides, an input field, a send button, and scrollable history. That gives users a starting point, but a two-person chat interface does not cover AI needs. Users may need source references, a way to stop an in-progress reply, and access to tools or integrations. Nanz contrasts Microsoft Copilot research agent chat with Microsoft Teams chat. Both include familiar elements, but the agent chat lowers the entry barrier with examples, large text buttons, labeled iconography, and voice options.
Nanz calls trust the biggest obstacle to AI adoption. Users often do not understand how the system works, and LLMs can still produce incorrect answers. She advises against claiming that a product is automatically safer or more reliable than other AI tools when much of the underlying model is outside the team's control. Applications should help users verify output instead. Citations and direct links let people inspect sources and reuse content while preserving a trail of accuracy. Depending on the task, teams can use tooltips with short quotes, inline links, or side panels that let users explore source material.
Agentic systems should show their plan before they act
When an AI agent can execute work on its own, users need to understand and approve the planned actions. Nanz points to agent flows in Claude and ChatGPT, where the system lists the steps it intends to take and asks for confirmation. Products can also let users turn approval prompts off or automatically allow repeated actions when that fits the workflow. If an agent takes action without showing why or how, users may distrust both the result and the system. Keeping a person involved is especially important when actions are difficult to reverse or have serious consequences.
Nanz says applications should explain what is happening and what comes next instead of presenting AI as mysterious magic. Streaming text gives users immediate feedback while content is generated, hides some of the wait, and lets them stop or redirect a response before it is complete. Showing the system's reasoning or processing steps can also help users understand how it reached a conclusion and identify where an answer went wrong. Other visual cues can mark new or changed content, move focus to updated areas, change the color of revised text, or outline modified lines of code.
Control requires a prominent stop button and suitable rollback
Users need to remain able to adjust, correct, and stop AI work. Nanz recommends a clear emergency brake that can halt generation, agentic workflows, scripts, or other actions without requiring a hidden menu or memorized command. After stopping a process, users often need to undo its changes. The right history depends on the task. A simple chat may need no versioning, while document tools may need undo and redo across roughly the last 10 steps. Advanced work may need checkpoints or save states, plus targeted edits that preserve good parts of a mixed-quality result.
Transparency covers permissions, memory, time, and cost
Users need to know what an AI tool can access and do. Nanz gives examples such as reading a database versus deleting tables, reading email versus sending from an address, and searching local files or running scripts. Permissions should support more than a permanent yes or no. Users may allow a single action, restrict access to one folder, or request notification each time. Systems that remember information should also let users see and control what is remembered, revoke access, and make the system forget. Approval screens should include expected time and cost when possible, while visible banners or outlines should show when an agent is controlling the browser.
Useful output needs guidance into the user's existing work
Nanz says a blank prompt box assumes users already know how to provide context, specify formats, break tasks down, and refine responses. Applications can teach these behaviors through examples, templates, suggested prompts, and guided workflows. They should also ask what users will do after generation. Next-step buttons can help people analyze results, share images, create documents, push code to a repository, or open a tracking ticket. Direct integrations make AI output useful inside the tools people already use. Without an easy next action, generated content risks remaining a novelty.
"After all, it doesn't actually matter what our software can do if our users hate using it so much that they will avoid it at all costs."02:02
Who should watch
You are adding generative AI or agentic actions to an existing product and need interaction patterns that users can understand.
Your AI feature produces uncertain results, accesses user data, or performs actions that need approval and rollback.
You have a working model integration but users do not know what to ask, what to verify, or what to do with the output.
