The vocabulary of conversational AI

Software that converses in natural language and takes action (answering questions, qualifying a lead, booking a meeting) from a knowledge base you define. It differs from a chatbot because it does not follow a fixed script: it interprets the question and looks up the answer.

See alsoRAG (retrieval-augmented generation)Guardrail

An AI agent that converses over real-time audio, transcribing speech and replying with synthesised voice. The main technical requirement is latency: past roughly a second, the conversation stops sounding natural.

See alsoResponse latency

A technique where the agent first retrieves relevant passages from your knowledge base and only then generates an answer based on them. It is what keeps the answer anchored in your content rather than in the model’s generic memory.

See alsoKnowledge baseGuardrail

The set of sources the agent can consult: site, FAQ, product catalogue, documentation, spreadsheets. Answer quality is capped by the quality of this base, no agent invents what is not there.

See alsoRAG (retrieval-augmented generation)

Assigning a commercial-readiness score to each visitor based on what they asked, which pages they visited and what data they gave. It decides who goes to the sales team and who stays in nurture.

See alsoHand-off

A rule that makes the agent open the conversation rather than wait. It fires on behaviour (dwell time, exit intent, a stalled cart) and works to the extent that the message matches the behaviour that fired it.

See alsoExit intent

A signal that the visitor is about to close the tab, detected from cursor movement toward the browser chrome. As a trigger it works when the person already showed interest; on a fresh arrival it is just an interruption.

See alsoProactive trigger

Passing the conversation from the agent to a human, a callback or a ticket. A good hand-off carries the transcript and context along, so the person does not have to repeat what they already said.

See alsoLead scoring

The share of questions the agent resolves without reaching a human agent. It is the steadiest return metric for an agent, because cost per ticket is a number the operation already tracks.

A constraint applied to the agent to prevent answers that are out of scope, off-brand or ungrounded in your content. It covers faithfulness checks, topic blocking and masking of sensitive data.

See alsoPII maskingRAG (retrieval-augmented generation)

Removing or replacing personally identifiable information (tax IDs, emails, phone numbers) before the text reaches a model or a log. It is a practical privacy requirement for any agent that talks to customers.

See alsoGuardrail

The time between the visitor’s question and the start of the agent’s answer. Under a second the conversation flows; past three, abandonment climbs fast, because the wait becomes noticeable.

See alsoVoice agent