Virtual Agent
A virtual agent is an automated system, usually a chatbot or voice bot, that handles customer-service or internal-help-desk requests on behalf of an organisation: answering questions, checking an order or account, resetting a password, or collecting information before a person takes over. Unlike a personal assistant that serves one user across their life, a virtual agent represents a company, speaks within that company's policies, and is judged by business measures such as resolution rate and customer satisfaction rather than by how impressive it sounds.
The direct ancestor is interactive voice response. IVR systems spread through telephone call centres from the 1970s and 1980s and relied on touch-tone input, the familiar "press 1 for billing" menu, later joined by speech recognition that accepted spoken keywords. They automated routing and simple look-ups, but rigid menu trees frustrated callers who had a problem that did not match any branch. Web chat widgets from the early 2000s added scripted bots with keyword matching. Virtual agents emerged as these were combined with natural-language understanding that maps free text to intents, a knowledge base of answers, and integrations with order, billing and ticketing systems.
Escalation design determines whether a virtual agent helps or harms. A well-built agent recognises when it should stop: after repeated failed turns, when the customer explicitly requests a person, when sentiment turns angry, or when the topic carries legal or safety weight. The handoff should pass the conversation history, the verified identity and what was already tried, so the customer does not start over. The failure that customers describe most often is the dead end, a bot that loops through "I didn't understand that" with no route to a human. Some jurisdictions have begun addressing this; California's 2019 bot-disclosure law, for instance, requires bots used to influence sales or votes to disclose that they are bots, and the EU AI Act sets transparency duties for systems that interact with people.
Language models raised the ceiling and the risk together. Retrieval-based designs that ground answers in a company's help articles let virtual agents answer far more questions than hand-written intent lists, and can phrase replies naturally. They can also state policies that do not exist. In Moffatt v. Air Canada (2024), a British Columbia tribunal held the airline responsible for incorrect bereavement-fare information its website chatbot had given a customer, rejecting the argument that the chatbot was a separate entity. The practical lesson is that what a company's agent says is generally treated as the company speaking.
Metrics can mislead. Containment or deflection rate, the share of conversations that never reach a human, is easy to measure and easy to game: a bot that makes escalation hard scores well while customers give up. Better measures track whether the issue was actually resolved, whether the customer contacted again within days, satisfaction after bot-only and after handoff conversations, and time to resolution. A common misconception is that the goal is to remove people; the more durable arrangement lets the agent absorb repetitive, well-defined requests so that humans have time for cases involving judgement, empathy or exceptions.
Scope and honesty are design choices. Good virtual agents state what they can do, disclose that they are automated, and decline requests outside their remit rather than improvising. Neighbouring entries describe conversational AI as the underlying technique and AI assistants as the personal equivalent; the virtual agent is distinguished by representing an organisation and by the accountability that implies.
Read virtual agent as a company's front line made of software: its value lies in resolving routine requests quickly and in knowing, reliably and early, when to hand a customer to a person.