Beta World

Copilot as the UI

Published
September 30, 2025
Share article

When the AI is the interface, menus fade and the experience becomes an ongoing conversation between human and agent.

Introduction: A cockpit without dials

Nearly 30 years ago, computing pioneer Mark Weiser criticized the idea of building a “copilot,” a chatty assistant who lived alongside the pilot and handled tasks. Weiser argued that the better approach was to redesign the cockpit itself so that the pilot always knew what was happening and could act instinctively. So, is “copilot” even the right metaphor? Sometimes, yes, but often a better one is the heads-up display that blends into your view. His vision anticipated today’s hidden UIs, ambient interfaces that make information part of our perception, like heads-up displays in fighter jets (“Enough AI Copilots, We Need AI HUDs”). Spell-checkers, GPS overlays and auto-complete are everyday examples of this approach (Litt’s essay).

Fast-forward to 2025, over 700 million people interact weekly with ChatGPT alone, quadruple the number a year ago (TechCrunch coverage). What does this actually look like in practice? People ask for outcomes, and the software figures out the steps.

Microsoft’s redesign of Copilot for 365 positions users as “agent bosses” directing intelligent systems (Design team write-up), and Figma’s AI report found that one in three designers shipped an AI-powered product in the past year, a 50 percent increase over 2024 (Figma’s 2025 AI report).

In this article we explore how copilot-style conversations are transforming UIs, when they are not enough, and what designers and business leaders need to know.

The conversational turn

Large language models enable chat interfaces to infer context and fill gaps. Instead of asking users to pick from rigid menus, a modern assistant can deduce that “Park” likely means “find parking near my destination” (LLM design patterns). Do menus go away? Not entirely. They shrink as intention takes the lead. Copilot Chat in Microsoft 365 is built around conversation as the core interaction model, shifting work from app-centric to intention-centric workflows (Copilot design principles). Redesigning to support this shift meant collapsing redundant menus, simplifying the palette and using collapsible navigation to reduce cognitive load (visual and structural changes).

“Design isn’t just polish anymore. It is how these products win users, drive loyalty, and stand apart.” (Andrew Hogan)

From a business perspective, the adoption numbers are staggering. OpenAI reported over 400 million weekly active users in February 2025, with usage still climbing (news reports). Sam Altman later remarked that roughly 10 percent of the world now uses their systems (interview recap), and the service was on track to hit 700 million weekly active users by August 2025 (audience update). Earlier, 92 percent of Fortune 500 companies were said to be using OpenAI products (enterprise adoption note). How do we measure the “turn”? Look at adoption, task completion and release velocity, not just time-on-page.

Why conversation works

  • Lowering barriers: LLMs can infer context and ask clarifying questions, reducing the need for explicit navigation. What does that buy us? Fewer clicks, faster intent capture.
  • Intention-centric flows: Complex tasks can start with a simple request (“summarize my sales pipeline”), and the system orchestrates multiple services.
  • Personalization: Conversational AI can adjust tone and complexity based on the user’s language, skill level and past interactions.

When conversation falls short

James Zein’s 2025 account of building a medical research platform reveals the limitations of chat-only interfaces. User testing showed that chat dominated the experience, causing people to miss critical context (case study). Why not stick to chat anyway? Because people need visible state, clear affordances and a shared understanding of what the system is doing. The team abandoned the chat-first approach and introduced dynamic blocks, UI elements that appear and populate based on AI analysis of user needs (dynamic blocks pattern). This shift let users see what the AI understood and what it was missing, creating a living interface (transformation notes).

Other patterns emerged:

  • Governor mechanisms: new content appears at reduced opacity until the user reviews and approves it, creating a human-in-the-loop feedback loop that builds trust (governor mechanism).
  • Milestone markers: visual indicators show progress and suggest next steps, guiding users without forcing them down a rigid path (milestone markers).

These patterns illustrate an important principle, not every interaction should be a conversation. Many tasks are better served through adaptive UIs that respond to context while keeping the user in control.

Diagram: the evolution of UIs

flowchart LR
  A["Traditional UI\n(menus & buttons)"] --> B["Conversational UI\n(chat & voice)"]
  B --> C["Generative UI\n(dynamically generated layouts)"]
  C --> D["Agentic UX\n(AI acts on your behalf)"]
  D -. Human oversight .-> C
  B --> A

The diagram above positions conversation as a milestone on a longer journey. Generative UIs dynamically assemble layouts and content to suit the context (definition and implications). Agentic UX goes further, AI agents act autonomously across systems with minimal human input (agentic UX overview). Designers increasingly craft semantic architectures, APIs, schemas and workflows, rather than pixels (design shift discussion). So, where are we headed? Toward software that understands goals first and renders interface second.

Beyond chat: generative and agentic interfaces

Generative UI: designing outcomes, not screens

Nielsen Norman Group defines a generative UI as a user interface generated by AI in real time to fit user needs and context (NN/g explainer). Instead of crafting static pages, designers set goals and constraints, then the system builds the interface that gets the user to the desired outcome (outcome orientation). What makes this different from dynamic forms? The interface is assembled on demand around the goal, not the other way around.

An example might be booking a flight. Rather than a step-by-step wizard, the system could ask, “Where to?” then build a custom interface showing recommended flights, dynamic filters and alerts based on your frequent-flyer status, calendar conflicts and weather. NN/g calls this outcome-oriented design and urges teams to orchestrate multi-step workflows while automating interface details (practical guidance).

Agentic UX: designing for machines

Agentic UX pushes further. Sometimes the “user” is an AI agent, not a person. Interfaces become invisible and live in API payloads, schemas and structured workflows (foundational view). Who is the audience then? Often a machine, which means labels, contracts and states must be unambiguous. This shift means designers must ensure that data is machine-legible and actions are predictable, because mislabelled data can block an AI from completing a task (practical constraints).

The economic stakes are large. A Cognizant–Oxford Economics study cited in that overview estimates that generative AI could inject up to one trillion dollars into the U.S. economy and affect 90 percent of jobs (impact estimate). The fastest-growing category in Figma’s survey is agentic AI, with roughly twice as many teams building agentic products as last year (survey highlight). How much autonomy should an AI have? Enough to create leverage, not so much that it erodes user agency.

Building trust: transparency and hallucinations

Powerful as they are, LLMs often hallucinate. A hallucination occurs when a generative AI produces output that seems plausible but is untrue (what hallucinations are). In one roundup, misattribution showed up in 76 percent of 200 quotations with uncertainty signaled only seven times, and legal tools produced incorrect information in at least one of six queries (findings and examples). Can we trust outputs by default? No. Trust comes from clarity about what the AI saw and what you can undo.

Designers cannot eliminate hallucinations, but they can mitigate harm through UI patterns:

  • Re-stating: repeating what the AI understood helps catch misinterpretations before an action is taken (re-stating pattern).
  • Confidence indicators: show a confidence score or highlight tentative sections so users know when to verify information.
  • Human-in-the-loop controls: governor mechanisms and milestone markers let users review AI-generated content (review controls).
  • Error recovery: provide clear undo paths and documentation when the AI makes a mistake.

A 2025 NN/g article stresses that LLMs cannot distinguish truth from fiction because they are statistical fill-in-the-blank machines (core limitation). What does “transparency” look like in the UI? Show inputs, assumptions and links, and always provide an undo.

The designer’s role in the copilot era

Figma’s report offers a clear signal, 95 percent of designers and developers said design is at least as important in AI products as in traditional ones, and over half said it is more important (report summary). The most successful teams embrace rapid prototyping, struggle and frequent customer research, treating best practices as starting points. What new skills matter now? Conversation design, semantic modeling and service choreography.

Meanwhile, Microsoft’s Copilot team talks about designers orchestrating coherence across complex systems (orchestration lens). This shift demands new skills:

  • Prompt engineering and conversation design: crafting prompts that elicit useful responses, anticipating misunderstandings and guiding users.
  • Semantic architecture: defining data schemas, taxonomies and APIs that agents can navigate (semantic focus).
  • Service choreography: orchestrating multiple services and managing their state transitions.
  • Ethical foresight: considering fairness, bias, privacy and user agency when delegating tasks to AI.

Designers must also act as translators between human values and machine logic. What keeps us grounded? Remembering that technology is a useful servant and a dangerous master, and designing so people always retain control (perspective and patterns).

Quick comparison of emerging patterns

| Pattern | How it works | When to use | | | |:-------------------:|:---------------------------------------------------------------------------------:|:----------------------------------------------------------------------------------------------:|---|---| | Conversational UI | Chat or voice interface that infers intent. | Use for exploratory tasks, complex queries or when users cannot articulate a precise sequence. | | | | Dynamic blocks | AI-generated UI components that appear based on context (pattern detail). | Use when chat would hide context, great for data-heavy tasks or dashboards. | | | | Governor mechanisms | Provisional content appears at partial opacity until approved (trust technique). | Use to build trust and give users control over AI outputs. | | | | Milestone markers | Visual indicators show progress and suggest next steps (guidance pattern). | Use for workflows where users need orientation and gentle guidance. | | | | Generative UI | AI generates the interface in real time to fit user needs (concept and guidance). | Use for highly personalized or outcome-driven tasks, such as travel planning or complex forms. | | | | Agentic UX | AI agents act autonomously via APIs and schemas (approach overview). | Use when tasks can be fully delegated, but always maintain human oversight. | | |

Frequently Asked Questions (FAQ)

Do I need to redesign everything to adopt AI?

No. Start by identifying tasks where conversation or dynamic generation will genuinely improve the experience. For repetitive or data-heavy flows, dynamic blocks may suffice. For novel capabilities like summarizing across apps, conversational interfaces can unlock value. Use agentic UX for tasks that can be safely delegated.

How do I handle errors or hallucinations?

Implement transparency measures, re-state the AI’s interpretation, display confidence levels, and allow users to review or undo actions. Monitor high-stakes outputs and involve subject matter experts. If the AI serves regulated industries, include human review.

Will AI replace designers?

Figma’s report suggests the opposite, design is more important than ever (key takeaway). AI accelerates mundane work but cannot replace the human ability to define goals, weigh tradeoffs and empathize. So, what is the designer’s edge? Orchestrating systems and defending user agency.

How much autonomy should AI agents have?

Decide based on risk and context. Low-risk tasks can be automated. High-risk tasks require strict controls and audit trails. Agentic AI is rising fast, but users still want final say (survey context).

What about accessibility?

Conversational UIs can be accessible for users with limited motor skills or literacy. Dynamic interfaces must be tested with assistive tech. Provide alt text and ensure voice and chat features support screen readers and captioning.

Conclusion: Designing with, not just for, AI

When AI becomes the interface, menus fade and the UI becomes an ongoing conversation. But conversation alone is not enough. Great experiences blend intention-centric chat, context-aware dynamic components, generative layouts and agentic workflows. They respect human agency, mitigate hallucinations and embrace iteration. Where should you start Monday? Pick one flow where intention beats navigation, ship a small pattern, learn and expand.

As Andrew Hogan reminds us, the next chapter will not be written by AI, it will be written with it. Business leaders who invest in design and ethical foresight today will be the ones whose copilot-driven products soar tomorrow.