AI conversation designer · Author · 20+ years in enterprise CX
Generative AI makes conversations sound more natural. It doesn't make them work better. The difference between AI that builds trust and AI that quietly destroys it comes down to conversation design — and whether anyone in the room has actually practised it.
Most people describe conversation design as writing for chatbots. That framing misses almost everything important about the discipline.
Conversation design is the discipline of deciding how AI systems behave in language — not just what they say, but how they structure dialogue, handle misunderstanding, manage uncertainty, and determine when to escalate to a human.
Writing is the medium. Judgement is the work. The core responsibility of a conversation designer is deciding how the system behaves when understanding is partial, confidence is low, and consequences matter. That is a risk discipline as much as a language discipline.
Every generation of language technology has produced the same impression: natural conversation is now solved. Each time, the same coordination failures appeared again, in slightly more fluent form. Generative AI is no different. The models are more capable. The failures are harder to see. The need for deliberate conversation design is greater than it has ever been.
Conversation design at enterprise scale is not one skill. It is a cluster of interconnected disciplines applied across the full lifecycle of a system.
Conversations have phases, even when they feel fluid. Opening, alignment, working, closing. AI systems must be designed to manage these transitions explicitly — they do not acquire this from training data. Turn design determines whether interactions feel like conversations or like forms pretending to be conversations.
Misunderstanding is normal in conversation. It must be designed for, not avoided. Good conversation design anticipates where the system's confidence will be low, builds in the right signals to the customer, and creates recovery paths that don't make customers feel like the system failed them.
Intents are not descriptions of what customers say. They are commitments about how the system behaves. Each intent definition is a risk decision. In generative AI systems, this work extends into prompt architecture and the constraints placed on model output.
The moment AI passes a customer to a human is the moment trust is either reinforced or broken. Escalation is not a fallback. It is a designed mechanism for preserving accountability when automation should not proceed. Context must travel with the customer. The agent must inherit understanding, not just a transcript.
Generative AI amplifies design decisions. A well-bounded system becomes dramatically more helpful. A poorly bounded one becomes dramatically more dangerous. Conversation design in GenAI environments includes prompt architecture, behavioural constraints, confidence calibration, and the governance structures that stop systems drifting after launch.
Building a conversational system is finite work. Stewarding it across years, as the organisation changes, the model updates, new policies arrive, and customer expectations evolve, is ongoing. The best conversation designers spend most of their careers in stewardship, not initial build.
Generative AI makes conversations sound better. It doesn't make them work better. The organisations currently skipping conversation design in favour of faster GenAI deployment are making the same mistake, in a more expensive context.
Classical AI failed obviously. A bot that didn't understand said "I'm sorry, I don't understand." Frustrating, but honest. Modern generative AI produces fluent, well-structured answers that sound authoritative — and may be completely wrong. The customer takes the answer at face value. By the time the error surfaces, it's in churn data, complaint queues, or a regulator's inbox.
The structure that used to live in explicit flows and intent rules now lives in prompts, system instructions, and orchestration logic. It is harder to inspect, easier to lose, and more consequential when it drifts. That is not an argument against generative AI. It is an argument for applying the same rigour to it that good conversation designers have always brought to the systems they build.
Read the full essay on conversation design and GenAIAdvisory, review, and design services for organisations building or improving conversational AI systems. Offered selectively and in a personal capacity.
A structured review of an existing conversational AI system — virtual agent, chatbot, voice AI, or agent assist. Covers dialogue structure, uncertainty handling, escalation logic, GenAI bounding, and production governance. Output: a prioritised set of improvements with clear rationale.
Ongoing vendor-neutral advisory for teams building or deploying conversational AI. Decision support on dialogue architecture, NLU design, GenAI prompt strategy, escalation design, and governance. Project-based or retainer. For organisations who want a senior practitioner, not a methodology.
Workshops on conversation design principles for product, technology, and CX teams deploying conversational AI. Covers the fundamentals, GenAI-specific considerations, and the governance practices that keep systems honest over time. Based on the content of Designing AI Conversations at Scale.
The book on conversation design
Covers conversation design fundamentals, NLU architecture, voice AI, generative AI risks, agentic systems, Agent Assist, governance, analytics, and the ethics of AI-powered customer conversations. Built from twenty years of enterprise practitioner experience.
Get the bookConversation design is the discipline of deciding how AI systems behave in language — not just what they say, but how they structure dialogue, handle misunderstanding, manage uncertainty, and determine when to escalate to a human. It draws on linguistics, UX, systems thinking, and risk management. Writing is the medium. Behavioural judgement is the work.
An AI conversation designer decides how a system behaves when understanding is partial, confidence is low, and consequences matter. In practice, this means designing dialogue structure, defining uncertainty handling, setting escalation logic, writing and governing prompts in generative AI systems, and stewarding systems in production over time — not just at launch.
No. Generative AI has made it easier to build systems that sound natural. It has not made it easier to build systems that behave correctly under uncertainty. Because generative AI sounds confident even when wrong, fluent failure is harder to detect and more damaging when it surfaces. Conversation design is the discipline that prevents it — through deliberate structure, bounding, and governance.
Conversational AI refers to AI systems that engage in natural language dialogue with humans — virtual agents, chatbots, voice assistants, and agent assist tools. Modern conversational AI is powered by large language models that generate language dynamically rather than selecting from predefined responses. What distinguishes high-performing conversational AI from low-performing conversational AI is rarely the model. It is the conversation design surrounding it.
Enterprise product, technology, and CX teams deploying or reviewing conversational AI in customer-facing or internal operations environments. Particularly useful before major investment decisions — or after a system launch that has underperformed the business case. Engagements are offered selectively, in a personal capacity, and are vendor-neutral.
Get in touch
Whether you're reviewing an existing system, designing a new one, or trying to understand what's gone wrong — that's what the advisory practice is for.
Book a discussion