Glossary of voice AI and automated customer service
This glossary defines the terms used around voice agents and conversational agents in customer service. Each entry gives a self-contained definition, then explains what the idea changes in practice for a business.
How to use this glossary
Entries are grouped by theme rather than alphabetically: first what an agent is and its variants, then the technical building blocks of the conversation, the quality of answers, telephony, connections to business tools, and finally data.
Each entry opens with a self-contained definition, usable as it stands in a requirements document, then explains what the idea changes for customer service and how LeetCall approaches it.
The key ideas for choosing a voice agent
Three families of ideas are enough to compare two voice agents seriously. The first concerns the conversation itself: STT transcribes speech, VAD governs turn-taking and enables barge-in, TTS provides the voice. An agent that talks over the caller or leaves awkward silences fails on these blocks, however good its answers are.
The second family concerns accuracy. A language model left to itself can hallucinate; grounding, or RAG, forces it to answer from authorised sources, and a supervision layer such as Guardian Agents checks the reply before it is spoken. The question to ask is always the same: what does the agent do when it does not know?
The third family concerns connections. Without APIs or webhooks into the CRM, the ERP or the booking system, an agent can only recite general information; with them, it acts. That leaves the human hand-off, which must be planned, triggered on clear rules and passed on with its context.
Definitions
AI voice agent
An AI voice agent is software that answers a company's phone calls, understands in natural language what the caller says and handles the request. It combines speech recognition, a language model and speech synthesis, and connects to business tools to look up information or record an action, such as a booking or a ticket.
In customer service, an AI voice agent takes inbound calls at any hour. At LeetCall, it is built for a single business, from its procedures and its vocabulary, and hands over to a person whatever falls outside its scope.
Callbot (or voicebot)
A callbot, also called a voicebot, is a conversational robot that talks over the phone channel. The two words describe the same kind of tool: callbot stresses the call, voicebot stresses the voice. A callbot built on a language model understands a freely worded request, with no touch-tone menu and no imposed phrasing.
The word is still useful to separate voice from written channels. LeetCall prefers to say voice agent, because the tool does more than speak: it consults the company's data and carries out actions in its software.
Chatbot
A chatbot is a conversational robot that exchanges in writing, on a website, in an application or in a messaging app. It reads a message, infers the intent and replies with text. Older chatbots followed a tree of pre-programmed answers; those built on a language model understand freely worded requests.
On the web, a chatbot handles written customer requests: opening hours, order tracking, appointment booking. At LeetCall, the web agent shares the same sources and the same rules as the voice agent, so the answer is identical by phone and in writing.
Conversational agent
A conversational agent is software able to hold a continuing exchange with a person, by voice or in writing, while keeping the context of the conversation in memory. Unlike an answering machine or a form, it asks questions, rephrases, takes previous answers into account and can trigger an action.
The term covers both voice agents and chat agents. What makes a conversational agent useful in customer service is its ability to act in the company's tools rather than return a generic help text.
Conversational AI
Conversational AI refers to the set of technologies that let a machine hold a dialogue in natural language: speech recognition, language understanding, answer generation, speech synthesis and management of the conversation flow. It is the field to which voice agents, callbots and chatbots belong.
Speaking of conversational AI rather than of a robot is a reminder that several building blocks work together and that each one can fail. At LeetCall, a supervision layer sits on top of them to check the reply before it reaches the customer.
LLM (large language model)
An LLM, or large language model, is a model trained on large amounts of text that predicts and generates sequences of words. In a customer-service agent, it interprets the caller's request and phrases an answer in natural language, working from the material it is given at that moment.
An LLM does not know a company's data: it has to be supplied during the conversation. At LeetCall the model writes the answer, but the content comes from authorised sources and goes through a check before it is spoken.
STT (speech-to-text)
STT, or speech-to-text, is automatic speech recognition: turning an audio signal into text. In telephony, STT runs as a continuous stream, meaning it transcribes while the person is speaking instead of waiting for the sentence to end, which gives the agent time to prepare its reply.
The quality of STT conditions everything else: one misheard word leads to an answer wide of the mark. LeetCall covers this in a dedicated article, https://leetcall.ai/blog/stt-temps-reel-agents-vocaux, and the agent asks the caller to repeat rather than guess when a transcript is doubtful.
TTS (text-to-speech)
TTS, or text-to-speech, is speech synthesis: turning text into voice. It is the block that gives a voice agent its voice. The quality of a TTS shows in the naturalness of its pace, the accuracy of its intonation and its ability to pronounce proper nouns and trade-specific terms correctly.
A clear voice avoids repetition and misunderstanding. At LeetCall, the pronunciation of product names, street names or dishes is worked on with the business, because those are precisely the words the caller is waiting for.
VAD (voice activity detection)
VAD, or voice activity detection, is the technique that separates speech from silence and background noise in an audio stream. It tells the agent when the caller starts speaking and when they have finished. Without VAD, an agent talks over the caller or leaves awkward silences in the middle of the conversation.
VAD sets the rhythm of the exchange: it is what makes natural turn-taking and barge-in possible. LeetCall devotes a blog article to it, at https://leetcall.ai/blog/vad-detection-activite-vocale.
Barge-in
Barge-in is the ability of the caller to interrupt the voice agent while it is speaking. The agent detects the incoming speech, stops talking and starts listening again, as a person would. Without barge-in, the caller has to wait for the end of every sentence, which gives the conversation an artificial and heavy rhythm.
Barge-in is one of the clearest signs that a voice agent is pleasant to use. It rests on VAD and on careful management of turn-taking, which LeetCall tests on trial conversations before going live.
Turn-taking
Turn-taking is the rule that decides who speaks and who listens at each moment of a conversation. In a voice agent, it means deciding when the caller has finished a sentence, when the agent should reply, and how to react to hesitation, to a silence or to the caller picking up speech again.
Badly tuned turn-taking produces either interruptions or silences. It is as much a matter of tuning as of technology, which LeetCall approaches through conversational craftsmanship: https://leetcall.ai/blog/savoir-faire-conversationnel.
Latency
Latency is the delay between the end of the caller's sentence and the start of the agent's reply. It adds up the time to transcribe, the time to look up information, the time to generate the answer and the time to synthesise speech. Latency that runs too long makes it feel as if the agent did not understand.
Latency is reduced by running steps in parallel rather than one after another, and by preparing the answer while the caller is still speaking. It is one of the reasons streaming STT matters.
Hallucination
A hallucination is an answer produced by a language model that sounds plausible but rests on no real source. In customer service it takes the form of an invented opening time, an imaginary return condition or an availability announced in error. It is the main risk of an agent left without supervision.
Two things limit hallucinations: grounding the answer in the company's authorised data, and checking it before it is spoken. That is the role of Guardian Agents at LeetCall, described at https://leetcall.ai/blog/guardian-agents-hallucinations.
Guardian Agents
Guardian Agents are LeetCall's supervision layer: specialised agents that check a customer-service agent's reply before it is spoken or sent. They verify that it rests on an authorised source, that it stays within the agreed scope and that it is consistent with the company's own procedure.
When a reply does not pass that check, it is reworded or the conversation is handed to a person, rather than risking false information. Supervision is therefore as much a safety mechanism as a quality tool.
Grounding (RAG)
Grounding, implemented through the technique known as RAG (retrieval-augmented generation), means having a language model answer from documents and data retrieved at the moment of the question, rather than from its training memory alone. The model receives the relevant material, then writes its answer from it.
This is what lets an agent state a shop's real opening time or an order's real status. At LeetCall the sources are the ones the company has authorised: catalogue, procedures, customer records and the responses of connected tools.
Human hand-off
A human hand-off is the transfer of a conversation from the agent to a member of the team. It is triggered when the request falls outside the agent's scope, when the answer cannot be verified, when the subject is sensitive, or simply when the caller asks to speak to a person.
A successful hand-off carries over the context already gathered, so the customer does not have to repeat everything. At LeetCall, the hand-off goes to the client company's own team, and its trigger rules are set during framing.
IVR (interactive voice response)
An IVR, or interactive voice response system, called SVI in French, is the system that greets a call with a recorded message and offers a touch-tone menu. It routes the call to a department or a queue, but does not handle the request itself and does not understand natural language.
An IVR is still useful for routing, but it puts the burden on the caller to guess the right option. An AI voice agent takes the request as it is worded and follows it through to the action, handing over when needed.
Switchboard
A switchboard is the arrangement, human or technical, that receives the calls arriving on a company's number and distributes them to the right people. Staffed by a team, it offers full understanding of requests, but remains limited to opening hours and to whoever happens to be available at that moment.
An AI voice agent does not replace the switchboard: it absorbs repetitive calls and those that arrive when nobody can pick up, then passes to the team the exchanges that call for human judgement.
SIP telephony and SIP trunk
SIP (Session Initiation Protocol) is the protocol used to set up telephone calls over an IP network. A SIP trunk is the link that connects a company's phone system to its operator, and over which calls are carried. It is the route by which a voice agent receives the calls it is meant to handle.
In practice, connecting a voice agent does not mean changing number: the call is routed to it by the existing telephony. This connection is one of the steps of going live on a project.
Webhook
A webhook is a mechanism by which one piece of software automatically tells another that an event has just happened, by sending a message to an agreed address. Unlike a classic API call, where one system queries the other, here the sender pushes the information as soon as it exists.
In a customer-service agent, a webhook is used for instance to signal that a booking has been recorded or that a ticket has changed state. LeetCall uses them to link the agent to tools that do not expose a complete programming interface.
API
An API, or application programming interface, is the set of entry points a piece of software exposes so that other software can read or write its data in a controlled way. It is through APIs that a company's tools talk to each other, without a person having to copy information from one screen to another.
A useful voice agent reads and writes in the company's tools through their APIs: finding an order, placing a slot, updating a record. This is how LeetCall connects to software already in place, with no migration and no new tool to adopt.
CRM
A CRM (customer relationship management system) is the software where a company keeps the history of its relationship with each customer: contact details, exchanges, quotes, orders and open requests. Salesforce, HubSpot, Pipedrive and Zoho CRM are common examples. It is the reference for what the company knows about a contact.
Connected to the CRM, an agent recognises the caller, finds their file and records the exchange in the right place. The information stays in the tool the teams already consult, instead of living in a separate log.
ERP
An ERP (enterprise resource planning system) is the software that centralises the running of a business: stock, purchasing, production, invoicing and deliveries. Odoo is a widespread example. It is often where the exact answer sits for a question about a lead time, an availability or an invoice.
A customer-service agent connected to the ERP gives the real status of an order rather than a general answer. Access is opened to the necessary scope: the agent reads what it must read and writes only what has been planned.
Omnichannel
Omnichannel describes an organisation of customer service in which the different channels, phone, website, WhatsApp, messaging and email, share the same information and the same rules. A customer can start a request on one channel and continue it on another without having to explain everything again from the start.
The classic pitfall is a web agent saying the opposite of what the phone answers. LeetCall builds a company's voice and web agents on the same sources and the same procedures, so the answer stays consistent.
WhatsApp Business
WhatsApp Business is the professional version of the WhatsApp messaging service, which lets a company exchange messages with its customers from a number identified as its own. Its programming interface makes it possible to link the channel to other software, and therefore to have a customer-service agent reply there.
Many customers would rather write than call, particularly for a confirmation or an attachment. A LeetCall agent answers the requests customers leave on that channel, under the same rules as on the phone.
Call transcription
A call transcription is the text produced from what was said during a phone conversation. It is used to re-read an exchange, to understand why a given answer was provided, to spot the requests that come back most often and to improve the agent's answers over time.
The transcription can be recorded in the company's tools, next to the customer record or the ticket. What is kept, and for how long, is decided by the company, in line with its duty to inform the people concerned.
GDPR applied to voice AI
GDPR applied to voice AI refers to the obligations that fall on a company when an automated agent processes personal data over the phone: informing people, collecting only what is necessary, defining a retention period, securing the data and allowing the rights of access, rectification and erasure to be exercised.
In practice this means announcing that the other party is an automated assistant, limiting what is recorded and hosting data in a controlled setting. At LeetCall, data is encrypted in transit and at rest and hosted in the European Union.
Inbound call
An inbound call is a call placed by a customer to a company's number. It is the starting point of most customer-service requests: a question about a product, an order to track, a booking to make or a complaint to raise. The volume of inbound calls varies sharply with the hour and the season.
Inbound calls often arrive at the worst moment: during service, during a meeting, or outside opening hours. That is precisely the scope of a LeetCall voice agent, which picks up at any hour and hands over whatever needs a person.
Frequently asked questions
What is the difference between a callbot and an AI voice agent?
Both words describe a robot that holds a conversation over the phone. Callbot is the older term and also covers systems built on rigid scripts, while AI voice agent stresses natural language understanding and the ability to act in the company's tools.
What is the difference between a chatbot and an AI voice agent?
A chatbot exchanges in writing, on a site or in a messaging app; an AI voice agent exchanges by voice, over the phone. Voice brings its own constraints: real-time transcription, turn-taking management and the possibility for the caller to interrupt.
What is an AI hallucination in customer service?
It is an answer that sounds credible but rests on no real source, such as an invented opening time or return condition. The remedy is to ground the answer in the company's authorised data and to check it before it is spoken.
What is grounding, or RAG?
Grounding means giving the language model the documents and data relevant to the question so that it answers from them. RAG describes that retrieval followed by generation. It is what makes it possible to answer on a real catalogue or a real order status.
Why do VAD and barge-in matter for a voice agent?
VAD tells the agent when the caller is speaking and when they have finished, which avoids interruptions and silences. Barge-in lets the caller cut in mid-sentence. Together they give the exchange a rhythm close to a human conversation.
Is an AI voice agent compatible with the GDPR?
Yes, provided people are informed, collected data is limited, a retention period is set and processing is secured. At LeetCall, data is encrypted in transit and at rest and hosted in the European Union, and the client company stays in control of what is recorded.
Last updated :
The voice AI that revolutionizes your phone calls. Intelligent automation, available 24/7.