Callers now meet three kinds of automated phone systems: the classic IVR menu, the intelligent virtual agent (IVA) that understands natural speech, and the newer AI voice agent built on a large language model. The labels overlap in vendor marketing, so this guide explains what each one does, where each one fits, and why most enterprise teams end up combining them rather than picking one.
What is the difference between IVR, IVA and an AI voice agent?
An IVR follows a call flow that your team designs: callers press keys or say words from a defined list, and the system takes a fixed path. An IVA adds speech recognition and natural language understanding, so callers can say what they want in their own words and the system maps it to a known intent. An AI voice agent uses a large language model (LLM) to generate its replies during the call, so it can hold an open conversation, but its answers are less predictable.
The simplest way to remember it: an IVR follows a script, an IVA understands a request and then follows a script, and an AI voice agent writes much of its script as it goes.
IVR vs IVA vs AI voice agent: comparison table
| IVR | IVA | AI voice agent | |
|---|---|---|---|
| How callers interact | Keypad (DTMF) presses or short spoken words such as "billing" | Natural speech, such as "I want to change my payment date" | Open conversation, including follow-up questions and topic changes |
| What it understands | Only the keys and phrases defined for each prompt | Intents and details (dates, amounts, account numbers) it was trained to recognize | Almost any phrasing, plus context from earlier in the call |
| How it decides what to say | Fixed prompts and fixed paths | Fixed prompts chosen by the recognized intent | Generated responses, guided by instructions, data and guardrails |
| Typical use | Payments, balance checks, card activation, routing | Front-door routing, account servicing, scheduling | Open questions, FAQs, complex triage, after-hours help |
| Setup effort | Low to moderate; build and test the flow | Moderate; train and tune intents with real caller phrases | Moderate to high; write instructions, connect data, test for wrong answers |
| Main risk | Rigid menus frustrate callers | Misrecognized intents send callers down the wrong path | Confident but wrong answers, and less predictable behavior |
| When to choose | The task is structured and must run the same way every time | Callers have many reasons to call and menus are getting long | Questions are varied and a wrong answer is low risk or can be checked |
What is an IVR?
An interactive voice response (IVR) system is an automated phone application that plays prompts and acts on the caller's keypad presses or spoken responses. Every path is designed in advance, so the same input always produces the same result.
That predictability is the point. An IVR can look up an account, read a balance, take a card payment, or route a call based on the caller's number, all without an agent. Most IVRs today are built on standards such as W3C VoiceXML 2.0 or with visual call-flow builders that generate the same kind of logic. If you want to learn how to design one well, see our guide to IVR call flow design.
What are IVAs?
IVAs, or intelligent virtual agents, are automated agents that use automatic speech recognition (ASR) and natural language understanding (NLU) to work out what a caller wants from free-form speech. Instead of "press 1 for billing," an IVA asks "How can I help?" and recognizes the intent behind the answer.
IVAs come in a few common forms:
- Directed dialog: the system asks a narrow question ("Do you want to pay a bill or check a balance?") and listens for a small set of answers.
- Keyword spotting: the system picks out key words from a longer reply, such as "lost card" or "refund."
- Natural language: the system classifies an open answer into one of many trained intents and extracts details such as dates and amounts.
After it understands the request, a typical IVA hands off to a designed flow, so the actual transaction still runs the same way every time. Our explainer on conversational IVR covers how ASR and NLU work in more detail.
Is an IVA the same as a chatbot or RPA?
No. A chatbot usually means a text interface, although modern chatbots and IVAs often share the same NLU engine. Robotic process automation (RPA) is back-office software that automates repetitive tasks in other applications, such as pulling statements for an agent. An IVA talks to customers; RPA works behind the scenes and can support an IVA or an agent.
What is an AI voice agent?
An AI voice agent is a phone agent that uses a large language model to understand the caller and generate spoken replies in real time, usually with speech recognition on the way in and text-to-speech on the way out. It is given instructions, access to approved data, and sometimes tools it can call, such as an order lookup.
Strengths of AI voice agents
- Handles unexpected phrasing. Callers do not need to match a trained intent, and the agent can answer follow-up questions.
- Covers the long tail. Questions too varied to script, such as policy details, can be answered from approved knowledge sources.
- Faster to extend. Adding a new topic can mean adding content and instructions rather than building new menu branches.
Risks of AI voice agents
- Wrong answers. LLMs can produce confident but false statements. NIST's Generative AI Profile lists this as a core risk and calls it "confabulation."
- Less predictable behavior. The same question can get differently worded answers, which makes testing, auditing and regulated disclosures harder.
- Security and data exposure. A model connected to account data needs strict limits on what it can read, say and do.
- Outbound consent rules. In the United States, the FCC ruled in 2024 that AI-generated voices count as "artificial" voices under the TCPA, so outbound calls using them need the called party's prior express consent (FCC declaratory ruling).
Why keep a deterministic IVR backbone?
Keep the steps that must be exact, such as payments, identity checks and required disclosures, in designed call flows, even when an AI agent handles the conversation around them. A designed flow does the same thing every time, which is what auditors, payment networks and customers expect.
Three areas benefit most:
- Payments. Card data captured by keypad in a designed flow can bypass agents and recordings, which helps limit how much of your environment falls under PCI DSS. See IVR payments and PCI compliance.
- Authentication. Identity checks should follow fixed rules and fixed retry limits, not a model's judgment. See IVR authentication.
- Handoff. A designed transfer step can pass the caller's intent and verified details to an agent, so the caller does not repeat themselves.
How do you choose between IVR, IVA and an AI voice agent?
Choose by task, not by technology. Follow these steps:
- List your top call reasons and how many calls each one brings.
- Mark the structured tasks (payments, balances, activations, appointment changes). Build these as IVR flows.
- Look at how callers reach those tasks. If menus are long or misrouting is common, add an IVA front door that recognizes intent.
- Find the open questions that agents answer from knowledge, not systems. These are candidates for an AI voice agent, with guardrails and an easy path to a person.
- Decide the escalation rules for every path: when to transfer, what context to pass, and what to do after hours.
- Measure and tune. Track containment, transfers and repeat calls for each path. Our guide to containment rate and first call resolution shows how.
Can you combine IVR, IVA and AI voice agents?
Yes, and most mature contact centers do. A common pattern is an IVA or AI front door that understands the caller's request, designed IVR flows for transactions and verification, and a live agent for anything sensitive, emotional or unresolved. The AI handles language; the designed flows handle the steps where a mistake is costly.
How Plum Voice handles this
Plum Voice gives teams the designed-flow backbone and the speech AI to put in front of it. Teams build designed call flows visually in Plum Fuse, starting from prebuilt templates such as Payment Processing, Account Inquiry and Card Activation, or in code with VoiceXML on Plum DEV. Fuse includes a Dialogflow module that places a Google Dialogflow voice agent inside a call flow, and Plum AI Fusion sends each spoken request to multiple AI engines (Nuance, Google, Amazon and IBM) and combines the results. Plum can also build LLM-based AI voice agents and connect them to the same call flows, so open questions go to the agent while payment and authentication steps stay in designed flows, as on our payment processing page.