Conversational AI — AI Glossary

Reference details

Related terms

• Voice Agent
• AI Customer Support
• Multi-Turn Conversation

How it works

1. The transcript is the state, and because endpoints are stateless the caller resends it — so the conversation lives in the request.
2. Intent and task state may be tracked explicitly alongside the transcript, which is what allows a system to know where it is in a process rather than inferring it each turn.
3. Knowledge enters through retrieval or tools, because the model's weights cannot hold current or account-specific facts.
4. Actions are taken through tools, which is the boundary where a conversation stops being advisory and starts changing things.
5. History must be managed as it grows — truncated or summarised — and that decision competes directly with retrieval for the context budget.

Example: Why turn count is the wrong success metric

Two assistants handle the same request. One resolves it in three turns; the other takes eight, asking clarifying questions and confirming before acting. Measured on turns to resolution the first wins; measured on whether the user got the right outcome without a correction afterwards, the second may well win. Optimising the visible metric produces assistants that act fast on under-specified requests, which is how a conversational product becomes efficient and unhelpful at once.

Commonly confused with

Used for both the interface and the system, which obscures where failures live. The conversational surface is turn-taking; the system behind it is retrieval, state tracking and tool use, and most failures are in those rather than in dialogue. It is also not the same as a chatbot in the older sense: rule-based and retrieval-based systems answered from a fixed repertoire and failed visibly on anything outside it, where a model-backed system answers anything and fails invisibly, which is a different operational problem.

When to use it

Reach for it when:

• The task genuinely needs multi-turn clarification rather than a single request.
• State must persist across turns, requiring explicit tracking beyond the transcript.
• Explaining why most conversational failures are retrieval or state failures.

Reach for something else when:

• A form or a single call would serve the user better and more predictably.
• Turn count is the optimisation target, which rewards acting on under-specified requests.
• No history management exists, where long conversations crowd out retrieval.

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Conversational AI — AI Glossary.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path