•5 min read
Share:

Agentic UI Series - Post 1: Why Chat Interfaces Will Stick Around For A While (and Where They Fall Short)

Chat was how GenAI went mainstream. Here’s what it’s good at, where it breaks, and why it’ll stick around until we invent better interfaces.

AIAgentic AIUX

Key takeaways

  • Chat sent GenAI mainstream - a simple text box, the first interface to gpt-3.5, was enough to introduce millions of people to LLMs & GenAI for the first time.
  • It’s flexible and forgiving - open‑ended chat invites creativity and rapid iteration. It’s like the command line for non‑coders.
  • But it's also overwhelming - blank pages and completely open-ended possibilities create decision paralysis. Good UX patterns (starter prompts, enhanced inputs) are essential.
  • Voice is now real - low‑latency, interruptible speech makes "chatting" to AI feel close to real conversation, but it comes with text chat’s strengths and limits.
  • It’s not the final form of AI interactions - chat will remain the default for a while (we already live in Slack/Teams), but I don’t believe we’ve found the best UI/UX for working with AI systems yet.

How we got here: next‑token prediction → chat windows

Large language models are trained to predict the next token. That simple objective turned out to be incredibly powerful: wrap it in a friendly chat UI and you get ChatGPT, the product that brought LLMs to the mainstream consciousness.

The chat box is the new command line for non-developers.

Why it worked:

  • Zero‑setup, zero‑knowledge. Users could  interact successfully with ChatGPT without learning how to use it or having deep technical skills.
  • It’s responsive and reactive. Ask → get an answer → repeat.
  • (Almost) Unlimited use cases. One interface worked for ideation, drafting, coding, planning and anything else people could come up with.

Where Chat Works Well (and how to build on that)

  • Creativity & exploration. Chat’s open‑endedness helps people discover uses. Support it with prompt suggestions that show example use cases in the empty state.
  • Rapid iteration. Keep turns quick and answers not overwhelmingly long, acknowledge the model won’t always get things right first time and  propose next steps.
  • Personalisation. Light‑weight memory (preferences, glossaries, goals) reduces re‑prompting, improves consistency and builds relationships.
  • Low onboarding effort. Anyone who can type a message can get started - keep it simple.

Design Choices To Improve The Experience

  • Give the user starter prompts (sample use cases, templates, builds on previous entries). These reduce the blank‑page problem and teach people what tools can be used for.
  • Offer structured outputs (tables, checklists, JSON) the user can easily copy into other tools. Nothing happens in one master platform (yet...) and we’re getting to the point where chat-based tools can output full Excel spreadsheets or Powerpoint presentations already.

Where Chat Falls Down (and how to improve it)

  • Decision paralysis. Completely open prompts can give users paralysis of choice or send them down rabbit holes. Counter with visible capability menus (using chips and suggestions) and clear next actions.
  • Messy state. Long replies and threads hide key facts and decisions. Solve this with named tasks/conversation summaries, and memory that persists across chats.
  • Reproducibility. LLMs are variable by design, but in a wider system consistency is more important. Add system prompts, retrieval, typed tools, and evaluation to raise floor consistency. Experiment with parameters like temperature to find the ideal level for each use case.
  • Break down tasks Chat isn’t great at co-ordination. For workflows, pivot to structured flows/inputs (forms, planners) and hand tasks to an agent that expects and returns fixed formats.

Voice Chat Is Now Good(ish)

Low‑latency, interruptible speech changes the feel of chat, makes it more “human” and opens more use cases:

  • Pros of voice chat: hands‑free, faster than typing, more natural turn‑taking, better for brainstorming and “getting stuff down on a page"
  • Trade‑offs: potential for transcription errors and model input being less structured (a spoken “stream of consciousness" rather than an optimised prompt)

Why Chat Will Stick Around (for now)

Chat fits nicely with how we already work. Teams and Slack are already our first point of contact with many human colleagues today. This built-in comfort level, plus chat’s flexibility, make it the default interface for GenAI apps for the foreseeable future.

The future is agentic, but the entry point is still chat.

Chat Is Not The Final Form Of AI Interaction, So What’s Next?

We’re already seeing interfaces that go beyond the humble text box:

  • Agents that act. Models with computer control can literally use your PC for you.
  • Multi-modal assistants. Systems with access to camera and audio input that can “see”, explain, and guide users (see the latest Samsung Galaxy ad).
  • Native "AI-Based" Tools. "AI Spreadsheets”, “AI Browsers”, “AI Whatever-You-Can-Think-Of"

I’ll cover these patterns in more detail in this series of blogs.

Summary: A Checklist For Building Better AI Chat UX

  1. Show 3-5  starter prompts in the empty state (DON’T just offer “Ask me anything”)
  2. Add use‑case chips that reveal capabilities without overwhelming users
  3. Offer structured output modes (bullet points, tables/CSVs, JSON) and easy Export functionality
  4. Include memories (but make sure they are transparent and editable)
  5. Use web/context retrieval for facts and prompt to avoid confident hallucinations where possible
  6. Add evaluation & retries under the hood to raise reliability
  7. Let users name and manage threads
  8. Don’t use chat where it’s not the best interface. Eg use structured form inputs for set workflows

Sources & further reading

All Posts