Voice AI glossary: the terms you need to know

Every buyer of voice AI meets a wall of jargon in the first week. This glossary translates it - organised by theme, defined from the decision-maker's side of the table. It links to deeper guides where we have them.

Core concepts

Voice AI. The umbrella term: technologies enabling a machine to listen, understand and speak in natural conversation. Everything below lives under it.

Voice agent (AI phone agent). A system that conducts phone calls end to end - answering, conversing, executing tasks, closing. Inbound (answers callers) or outbound (calls leads and customers).

Virtual assistant. The broader term for an AI interacting with people on a business's behalf; a voice agent is a virtual assistant specialised in the phone.

AI receptionist. The most common application: answering a business's calls like a front-desk professional - bookings, triage, FAQs. Full treatment in what is an AI receptionist.

Conversational AI. The capability of free-flowing conversation - understanding intent rather than keywords. What separates modern agents from menu systems.

IVR (Interactive Voice Response). The previous generation: press-1 menus that route by making the caller navigate. Compared properly in auto-attendant, IVR and AI receptionist.

Auto-attendant. The simplest tier: a recording that informs or redirects. Useful for stating opening hours; incapable of booking one appointment.

The technology inside

STT (speech-to-text). Converts the caller's voice to text in real time, through accents, noise and unfinished sentences.

LLM (large language model). The conversation's brain: interprets intent, decides the action, formulates the reply - steered by business-specific instructions.

TTS (text-to-speech). Turns the reply into natural spoken voice; the layer whose quality most shapes perceived naturalness. The full loop is walked through in how voice AI works.

Latency. The gap between the caller finishing and the agent starting to reply. High latency makes conversation feel artificial; good systems respond in a fraction of a second.

Barge-in. The caller can interrupt mid-sentence and be heard, as with a person. Without it, calls feel like listening to a recording.

Knowledge base. The business information the agent answers from: services, prices, policies, hours. Keeping it current keeps the agent competent.

Prompt (script). The instructions defining persona, scope and limits: how it introduces itself, what it may resolve, when it transfers. Writing these is a craft.

Hallucination. A language model producing plausible but false content. Serious deployments control it by restricting answers to the knowledge base and defining what must never be invented.

Guardrails. The configured safety limits: forbidden topics, action boundaries, escalation triggers - what keeps an agent from promising what the business cannot deliver.

Call operations

Triage. Identifying the reason for a call and directing it: to a person, a department, or resolution by the agent itself.

Cold transfer. Passing a call without context; the customer starts over. Standard in traditional switchboards.

Warm transfer. Passing with context: the agent briefs the colleague first. The customer-experience difference is enormous.

Escalation. The rules deciding when a human takes over: sensitive cases, out-of-scope requests, frustration signals. Good escalation is what makes AI trustworthy.

Overflow. The agent answers only when the team cannot - busy line or nobody free. Classic starting configuration, minimal risk.

After-hours answering. Coverage when the business is closed: evenings, lunchtimes, weekends, holidays - where the highest-value missed calls usually live.

Outbound. Proactive calls by the agent: confirming appointments, contacting ad leads, reactivating past customers. The inverse of inbound.

Speed-to-lead. Time from a lead's expression of interest to the first real conversation. Minutes instead of hours changes closing rates dramatically.

Lead qualification. The questions determining whether a contact has substance - need, timeline, budget - so sales talks only to leads worth their time.

Voicemail. Where unanswered calls go to die: most callers hang up without leaving a message. The silent adversary AI answering eliminates.

Metrics owners care about

Answer rate. Share of received calls actually answered - the most honest snapshot of your phone operation, and the first number automation moves.

Time to answer. Rings before someone (human or AI) picks up; every extra ring raises abandonment.

First-call resolution. Share of calls resolved without transfers or follow-ups - competence, measured.

Transcript. The word-for-word record of each call. With transcripts, management sees what customers ask and how they are served.

Cost per answered call. Total answering cost divided by calls answered - the basis of any serious comparison between staffing models, which we run in what AI phone answering costs.

Compliance

GDPR. The EU's data-protection regulation, covering data collected in calls. The practical version for voice assistants is in our GDPR guide.

Recording notice. Telling the caller, at the start, that the call is recorded and why - mandatory when recording, AI or not.

AI self-identification. The transparency obligation: callers must know they talk to a machine. Also plain good business - context in is it legal?

Data processor. The GDPR role of any provider handling personal data on your behalf - requiring a written agreement on storage, retention and access.

Data minimisation. Collecting only what the purpose needs. In a voice agent, a configuration decision made once.

A living glossary

This hub grows as new terms enter buyer conversations. If a quote or a meeting produced a term missing here, odds are it is dispensable jargon - but tell us, and we will translate it into plain language.