AI voice agent glossary: 100 terms explained simply

An AI voice agent combines several technologies: speech to text (STT) to hear, a large language model (LLM) to understand and decide, text to speech (TTS) to speak, and integrations to act in your CRM, ERP or calendar. This glossary explains 100 terms you meet when you evaluate voice AI, AI call automation or an AI call center, in plain English, grouped by topic. Where it helps, we add how the term applies in practice.

Which artificial intelligence terms matter for voice agents?

• Artificial intelligence (AI). Software that performs tasks that normally need human intelligence, such as understanding language, recognising speech or making simple decisions.
• Generative AI. AI that produces new content, such as text, speech or images, instead of only classifying existing content. Voice agents use it to compose answers.
• Conversational AI. AI designed to hold a dialogue with people, by text or voice, keeping track of what was said earlier.
• AI automation. Using AI to complete a task end to end, for example a confirmation call that ends with an updated order in your ERP.
• Large language model (LLM). A model trained on large amounts of text that understands and produces language. It is the “brain” deciding what the agent says next.
• Natural language processing (NLP). The broad field of making computers work with human language, spoken or written.
• Natural language understanding (NLU). The part of NLP that extracts meaning: what the person wants and which details they gave.
• Intent. What the caller wants to achieve, such as “book an appointment” or “cancel my order”.
• Entity (slot). A specific detail inside a sentence, such as a date, an address, an order number or a name.
• Prompt. The text instruction given to a language model to guide its answer.
• Instructions (system prompt). The standing rules that define the agent’s role, tone, goal and limits for every call.
• Hallucination. When an AI states something false with confidence. Good agents reduce this by answering only from your information and admitting when they do not know.
• Guardrails. Rules and checks that keep the agent within limits: topics it must avoid, promises it must never make, cases it must hand over.
• Token. A small unit of text (a word or part of one) that language models read and produce.
• Context window. How much text a model can take into account at once, such as the conversation so far plus relevant documents.
• AI agent. An AI system that not only talks but takes actions, such as looking up an order or booking a slot.
• Agentic AI. AI that plans and carries out multi-step tasks with some autonomy, within the limits you set.
• Function calling (tool use). The mechanism that lets the agent call an external function during a conversation, for example “check free calendar slots”.

What do knowledge base terms mean?

• Knowledge base. The documents, texts and website pages the agent uses to answer questions. Its quality sets the quality of the answers.
• Retrieval-augmented generation (RAG). Looking up the relevant passages in the knowledge base before answering, so answers are grounded in your content.
• Embedding. A numeric representation of the meaning of a text, used to find passages similar to the question.
• Vector database. A database that stores embeddings and quickly finds the passages closest in meaning to a question.

Which voice technology terms should you know?

• Voice AI. Any AI that works with spoken language: recognising it, understanding it or producing it.
• AI voice agent. Software that holds a phone conversation instead of a person, understands free speech, answers and completes a task. Also called an AI phone agent.
• Voicebot. A general term for a bot that talks by voice. Also written voice bot or AI voice bot. Older voicebots follow fixed phrases; modern ones are AI voice agents.
• Chatbot. A bot that talks in writing, on a website or in a messaging app.
• Speech to text (STT). Converting spoken words into text. Errors here (a wrong digit in an address) spread through the whole call.
• Automatic speech recognition (ASR). Another name for the technology behind STT.
• Text to speech (TTS). Converting text into a spoken voice. It decides whether the agent sounds natural or robotic.
• Word error rate (WER). The share of words a speech recognition system gets wrong. Lower is better; real phone audio is harder than studio audio.
• Latency. The pause between the moment the caller stops speaking and the moment the agent starts answering. Ours answers almost instantly, under one second in conversation.
• Barge-in. The caller’s ability to interrupt the agent. A good agent stops, listens and adapts.
• Turn-taking. How the conversation passes between speaker and agent without long silences or talking over each other.
• Endpointing. Detecting when the caller has finished speaking, so the agent neither interrupts nor waits too long.
• Voice activity detection (VAD). Detecting whether someone is speaking at all, as opposed to silence or background noise.
• Custom vocabulary. A list of names, products and terms the agent should recognise and pronounce correctly.
• Multilingual agent. An agent that can work in several languages. Ours speaks Romanian natively and 83 more languages.
• Sentiment analysis. Estimating whether the caller sounds satisfied, neutral or upset, for example to prioritise a callback.

What do telephony terms mean?

• Inbound call. A call received by your company. The agent answers it.
• Outbound call (initiated call). A call your company starts. The agent calls the person.
• IVR (interactive voice response). A phone menu with recorded prompts and key presses, such as “press 1 for sales”.
• DTMF. The tones generated by pressing phone keys, used by IVR menus.
• Auto attendant. A simple automated switchboard that greets callers and routes them to an extension.
• VoIP. Voice over IP: phone calls carried over the internet instead of classic phone lines.
• SIP. The protocol used to set up and end VoIP calls.
• SIP trunk. A connection that links a phone system to the public phone network over the internet.
• Virtual PBX. A company phone system that runs in the cloud instead of in a box in the office.
• Phone number (DID). A direct number that reaches a specific line or agent. With AgentVocal AI a number is allocated automatically during setup.
• Caller ID. The number shown to the person called. A recognisable number helps people answer.
• Concurrent calls. How many calls run at the same time. AI agents can handle several in parallel.
• Call transfer. Passing a live call to another number or person.
• Warm transfer. A transfer where the person taking over receives the context first, so the caller does not repeat everything.
• Voicemail. An automated mailbox that records a message when nobody answers. If voicemail picks up, the call is charged as a normal minute.
• Answering machine detection (AMD). Technology that tries to tell whether a person or a voicemail answered.
• Unanswered call. A call nobody picks up. Only calls nobody answers are not charged.
• Call recording. The audio of the call, kept only if you switch it on, and with the person informed at the start.
• Transcript. The written text of the call, line by line. Faster to read than listening to the audio.
• Call summary. A short written overview of the call, generated after it ends.
• Call outcome (disposition). The structured result of a call, such as “confirmed”, “rescheduled” or “asked for a colleague”.

Which call center and operations terms should you know?

• Call center. A team that makes or answers calls at volume.
• Contact center. A call center that also handles other channels, such as email, chat and messaging.
• AI call center. A call center where AI voice agents take part of the calls and people take the rest, with clear handover.
• Omnichannel. Continuing the same conversation across phone, SMS, WhatsApp and email, without losing context.
• Campaign. A set of initiated calls with one purpose, one list and one set of rules.
• Contact list. The people a campaign calls, with the context each call needs.
• Predictive dialer. A system that dials many numbers in advance and connects answered calls to free human agents.
• Call cadence. The retry rule: how many attempts, at what times and over how many days, counted from the first call.
• Calling hours. The time slots in which the agent is allowed to call.
• Retry. A new attempt after an unanswered call, according to the cadence.
• Deduplication. Making sure the same case is not called twice by mistake.
• Do-not-call list (restriction list). Numbers that must not be contacted again. Ours is checked before every call, SMS or WhatsApp message sent automatically.
• Human handover (escalation). Passing a case from the AI agent to a person, by transfer or with a note for a callback.
• Human in the loop. A setup where people review, approve or take over the cases that need judgement.
• Script. The outline of the call: greeting, purpose, questions, outcomes, closing.
• Follow-up message. The SMS, WhatsApp message or email sent after the call with what was agreed.
• First call resolution (FCR). The share of requests solved in the first call, without callbacks.
• Average handle time (AHT). The average duration of a call, including the work after it.
• Containment rate. The share of calls the AI completes without handing over to a person.
• Quality assurance (QA). Listening to and scoring a sample of calls to improve the agent’s instructions.
• AI receptionist. An AI voice agent that answers a company’s phone, answers questions, books and takes messages.
• Virtual receptionist. A receptionist who answers your calls remotely. It can be a human service or an AI receptionist.
• AI answering service. An AI that answers calls you would otherwise miss, often after hours.
• Appointment booking. Creating an appointment in a calendar during the call, usually by offering two or three free time slots.

What do integration terms mean?

• API. An interface through which two systems exchange data automatically. A CRM or ERP needs one to connect.
• Webhook. A message one system sends automatically to another when something happens, such as “new order created”.
• CRM. Customer relationship management software, where contacts, cases and notes live.
• ERP. Enterprise resource planning software, where orders, stock and invoices live.
• SMTP. The protocol for sending email. After-call emails can go out from your own mail server.
• SMS. A short text message, useful for reminders when other channels are not available.
• WhatsApp follow-up. A WhatsApp message after the call, often used to send a list of documents or a confirmation.

Which compliance and business terms matter?

• GDPR. The EU’s General Data Protection Regulation, which governs how personal data, including call data, is processed.
• Controller. The organisation that decides why and how personal data is processed. For calls made on your behalf, that is your company.
• Processor. The organisation that processes data on the controller’s instructions. For AgentVocal AI, that is CyberSkill S.R.L.
• DPA (data processing agreement). The contract between controller and processor that sets out how data is handled.
• Special category data. Sensitive data, such as health data, that needs extra protection and should not be collected by the agent unless specifically agreed.
• Lawful basis. The legal reason for processing data, such as a contract, legitimate interest or consent.
• EU AI Act. The EU regulation on artificial intelligence, which includes transparency duties: people should know when they talk to an AI.
• Per-minute billing. Paying for the time the agent talks. With us, minutes are rounded up on every call.
• White-label. A product offered under one brand while built on technology components that are not named.
• Onboarding. The setup process. Ours is a 9-step wizard; in our test, the technical setup of a complete agent took 4 minutes and 38 seconds.

Where to go from here

Terms are easier to understand when you hear them in a real call. Listen to the demo calls, see the step-by-step process on the how it works page, or read short answers in the FAQ. For how these pieces behave in Romanian, see the guide on the Romanian voicebot.