Hand holding phoneHand interacting with VUI

Exploring fluid interfaces.

In my opinion, conversational design is similar to choreography. You're designing a dance between a human and a machine, where neither one fully controls the music. My job is to make the machine a good enough dance partner that the human never feels stepped on.

Every conversational experience, regardless of modality, has the same underlying structure in simple terms:

  • The user has intent — they want something done, answered
  • The system must recognize that intent — through voice, text, gesture, or context.
  • The system responds — with words, actions, visuals, or silence.
  • The conversation advances or repairs — the dance continues, or someone stumbles and recovers.

My job is to juxtapose all four of these moments across every path a conversation can take, the happy paths, the confused paths, the frustrated paths, and the paths nobody predicted.

AI conversational design

AI Assistant

My approach is to design for error states first. In conversation, misunderstanding is the norm. Around 30-40% of conversational turns involve some form of repair, clarification, or renegotiation. Designing "The happy path" dialogue is roughly a third of the work. The rest, and most important, is figuring out what happens when the system doesn't understand, when the user changes their mind mid-task, when context shifts, or when confidence is low.

AI conversational design: a practical framework for building systems that are good conversational partners, not just accurate ones.
The Cooperation Model: be relevant, be clear, don't overshare, be truthful. The core insight: a system that breaks these maxims feels broken even when the information is correct.
Turn design vs. flow design: start with individual turns, then zoom out to flows, then stress-test the chain.
Error and repair as a first-class concern: 30 to 40 percent of conversational turns involve repair, clarification or renegotiation, so the work lives in what happens when the system doesn't understand, the user changes their mind, context shifts or confidence is low.
Persona and register: a casual, friendly tone implies flexibility and forgiveness; a clinical, precise tone implies reliability and narrow scope. Trust breaks down when tone and capability don't match.
Diagram: the conversation layer takes the user's turn, understands it (intent, entities, sentiment), decides what to do (answer, clarify, confirm, repair) and responds, shaped by Gricean maxims. Context feeds every step: the conversation so far, who the user is, and the situation (time, place, device). Each reply becomes context for the next turn.
Diagram: the visual interface is what people see and touch: the message stream, system status, shortcuts and the composer. The conversation layer decides what happens, sending system output to the interface and receiving user input back from it.
A practical starting framework, seven steps in order: inventory the user intents, map the system's domain knowledge, identify the gap, design individual turns before flows, define the persona as a functional spec, build your state model and recovery paths, and test with real language, not idealized scripts.
1 / 8

Click a slide to enlarge

Session View DashboardSystem Actions Table

Multimodal channel architecture

Intent mapping diagram

Glossary

Entity

A recognized real-world object or concept within user input (e.g., dates, locations, names).

Confidence Score

A probabilistic measure indicating how certain a model is about its classification or generated output.

Latency

The time delay between a user's input and the system's response. Critical for conversational interfaces.

Guardrails

Hardcoded rules or secondary models designed to constrain AI behavior, preventing unsafe or off-topic outputs.

RAG (Retrieval-Augmented Generation)

Enhancing LLM responses by dynamically retrieving relevant facts from an external knowledge base before generating text.

Recall

The fraction of relevant instances that were actually retrieved or identified by the model.

Precision / Recall Tradeoff

The balance between being strictly accurate (precision) and capturing all possible positive matches (recall).

F1 Score

The harmonic mean of precision and recall, often used as a single metric to evaluate overall classification accuracy.

Multimodal

Systems capable of processing and understanding multiple types of input concurrently, such as text, voice, vision, and touch.