Conversational IVR lets a caller say why they are calling in their own words instead of pressing a number, and routes the call from what they said. It replaces the menu tree, not the queue. The caller still ends up wherever the old option sent them, which is usually a wait for an agent.
That distinction decides whether buying one is worth it. If every option on your menu ends in a transfer, a conversational front end gets callers to the right line faster and the line is just as long. A narrower question is more useful: which of your menu options could be finished on the call, with no transfer at all? The worksheet below answers that for your own menu, one option at a time.
What is conversational IVR?
Conversational IVR is an interactive voice response system that takes natural speech as its input. Instead of "press 1 for billing," the caller hears an open question, says something like "I want to check what I owe," and the system matches that to an intent and sends the call where that intent is supposed to go. The recognition is new. The destinations are the ones you already configured.
Interactive voice response units, also called voice response units, have let callers self-serve since long before any of this. Gans, Koole and Mandelbaum described them in their 2003 review of call center research as specialized computers that let customers "communicate their needs and to self-serve" using the keypad or their voice, reporting balances and departure times and triggering simple transactions. In many banking call centers at that time, they wrote, roughly four in five calls were fully self-served by an IVR (Telephone Call Centers: Tutorial, Review, and Research Prospects, Manufacturing and Service Operations Management 5, 2003).
The self-service idea is decades old, in other words, and even the speech part is not new. The same 2003 review notes in passing that the then current generation of speech recognition already let IVRs interpret complex commands. What has changed recently is how well the recognition works on an unscripted sentence, and how little scripting it takes to get there.
Conversational IVR vs a touch-tone IVR
The difference is the input. A touch-tone menu asks the caller to translate their problem into one of your categories and press the matching digit. A conversational menu asks them to describe the problem and does the translating for them. Both then hand the call to the destination attached to that category.
You can see how tightly coupled the two are in the platform documentation. Amazon's own walkthrough for putting a bot in front of a call flow describes the result as a "Press or Say" interaction: the bot "prompts callers to press or say a number that matches the menu option for the task to complete" (Add an Amazon Lex bot, Amazon Connect Administrator Guide). The canonical build is a menu where you may speak the digit rather than press it. The example ships two intents, one that looks up an account balance and one that asks for an agent, which is a fair summary of what most of these front ends do all day.
Speaking instead of pressing does help real callers. It is faster than listening through eight options, it handles people who do not know which category they belong in, and it does not fall apart on a menu six levels deep. None of that reduces the number of calls that reach your agents.
Conversational IVR vs an AI agent that finishes the call
Three different things get sold for the front of a phone line, and they differ by what the system is allowed to do once it understands the caller.
| Touch-tone IVR | Conversational IVR | AI agent | |
|---|---|---|---|
| What the caller does | Presses digits through a menu | Says what they want | Has a conversation |
| What the system does with it | Matches the digit to a destination | Matches the sentence to an intent, then to a destination | Answers the question or takes the action |
| Where the call goes next | The queue for that option | The queue for that intent | Ends on the call, or transfers if a person is needed |
| Effect on your queue | None | Calls arrive better sorted | Fewer calls arrive at all |
| What you have to supply | A menu tree and recordings | Intents, sample phrasings and a fallback path | Knowledge, rules, and access to the systems that hold the answer |
The last row is why the third column costs more to set up. Routing needs to know which bucket a sentence belongs in. Finishing a call needs the actual answer, which means reading from your knowledge base and writing to your scheduling or customer records. For a longer look at the categories and which calls an agent can finish on its own, see what AI agents handle in a call center.
Vendors also sell the third column as an intelligent virtual agent, which makes IVR vs IVA the same comparison under different initials. Treat the label as marketing and read the capability row instead: does it hand the call off, or does it resolve it?
Why faster routing may not shorten the wait
Routing improvements move time around inside the call. They do not change how many agent minutes you have or how many callers want them.
There is a measurement in the 2003 review that has aged unnervingly well. Gans, Koole and Mandelbaum observed that the process by which a caller who wants a rep identifies themselves through an IVR "can average 30 seconds, even though subsequent queueing delays often reach no more than a few seconds." The front end was costing the caller more time than the queue it was sitting in front of.
Anyone who has called a bank recently knows that shape. If your own numbers look like that, the fix is not a better menu. It is either fewer steps before the queue or fewer calls that need the queue at all. Only the second one shows up in the metrics your team runs on, because the first just relabels where the seconds were spent.
Which menu options should be resolved instead of routed?
Three questions settle it for any option on your menu.
- Can it be finished with one lookup or one action against a system you already run? Order status, balance due, next available appointment, cancelling something. If yes, it is a candidate to resolve.
- Does getting it right require weighing something that is not written down? A payment plan for a customer in trouble, whether to make an exception, how upset someone is. If yes, route it.
- If the answer is wrong, what does it cost? A wrong appointment time costs a phone call. A wrong coverage answer or a wrong promise about a refund costs more than that, so it belongs with a person or needs a person's approval before it goes out.
Option one and option three pull against each other on purpose. Plenty of requests are one lookup away from being answered and still too expensive to get wrong. Those are the ones worth arguing about in a room, and the worksheet exists to get them into that room.
An IVR menu audit worksheet
Pull up your current menu and fill in one row per option. Monthly volume comes from your own reporting; everything else is a judgment your team makes together.
| Option as the caller hears it | What they actually want | Monthly calls | One lookup or action? | Needs judgment? | Cost of being wrong | Verdict |
|---|---|---|---|---|---|---|
Verdicts are resolve, split or route. Split is the one most audits end on, and it is the finding worth having.
Here is the worksheet filled in for an example seven option menu. The volumes are illustrative, not anyone's real numbers, and are there to show how the arithmetic changes the ranking.
| Option as the caller hears it | What they actually want | Monthly calls | One lookup or action? | Needs judgment? | Cost of being wrong | Verdict |
|---|---|---|---|---|---|---|
| 1. Check an order or job status | Where is it and when | 4,100 | Yes | No | Low | Resolve |
| 2. Billing and payments | Mixed: what do I owe, and I am disputing this | 3,600 | Half | Half | Low to high | Split |
| 3. Book, change or cancel an appointment | A slot, or a different slot | 2,800 | Yes | No | Low | Resolve |
| 4. Technical help | Mixed: a known fix, or something new | 2,200 | Half | Half | Medium | Split |
| 5. Make a complaint | To be heard, and a remedy | 700 | No | Yes | High | Route |
| 6. Sales and new accounts | A price and a commitment | 900 | No | Yes | High | Route |
| 7. Everything else | Something your menu never anticipated | 1,100 | No | Yes | Varies | Route |
Two options come out clean, three stay with people, and the two in the middle are the real result. Options two and four each carry a few thousand calls a month of which roughly half are a single lookup, and today both halves sit in the same queue. Splitting them into "what do I owe" and "I have a problem with a charge" sends the lookups somewhere that can answer them immediately and leaves the hard half with the people who should have it.
That is also the answer to a question that sounds reasonable and is not: should the AI take billing calls? The menu option is the wrong unit. The caller's intent is the unit, and one option usually carries several.
What a conversational front end needs from you
Whichever layer you buy, speech recognition is probabilistic and the design work is in deciding what happens when it is unsure.
Amazon Lex processes a caller's audio in two steps, speech recognition to produce a transcript and then natural language understanding to work out the intent, and when it cannot settle on one result it returns several candidates with confidence scores from 0.0 to 1.0 (Using confidence scores, Amazon Lex V2 Developer Guide). Somebody has to pick the threshold at which the system acts on its best guess versus asks again, and that choice is an operational decision, not a technical default.
Google's Dialogflow CX makes the same point structurally. Every flow is created with no-match and no-input event handlers already attached, and they cannot be deleted (State handlers, Dialogflow CX documentation). The platform will not let you ship a flow with no answer to "the caller said something I did not recognize" or "the caller said nothing."
The build list for a conversational layer, then: the intents you want recognized, enough real phrasings of each to recognize them, a confidence threshold, a path for the no-match and no-input cases, and a way to a person that works on the first try. Write those last two before you write the greeting, because they are the parts callers remember.
If you are evaluating an agent rather than a router, the pre-launch work is heavier and more specific. Testing conversational AI before customers hear it has the test set we use for that.
How to add an AI option to the menu you already have
The deployment most contact centers start with is one new line in the existing menu: "press 1 to talk to our AI assistant now." Callers who take it are answered immediately. Callers who would rather wait for a person keep their place in line, so nobody is forced through the AI to reach your team. You keep the menu, and you learn what the agent can carry from the calls it actually takes.
With Telvana, what the agent does on those calls is answer from the knowledge base you give it and take the actions you allow: looking up a customer, checking availability, booking, rescheduling and cancelling. When a call needs a person it transfers to the number you choose, under rules you set. Every call is recorded, transcribed, summarized and tagged by outcome, which is how you find out which intents it is finishing and which ones keep ending in a transfer. Calls reach it from any phone system that can forward or transfer to an outside number, so adding the option is a change to your menu rather than a migration.
Our team builds the agent from your scripts, FAQs and transfer rules, tests it on your own call types, and you approve it before it takes a live call. Pricing starts at $399 a month based on call volume, plus a one-time setup fee that covers building, testing and launching it. Hold queue relief describes the press-1 deployment in more detail, and pricing has the rest.
Frequently asked questions
What is the difference between IVR and IVA?
An IVR routes: it works out what the caller wants and sends the call to the right queue. An IVA, or intelligent virtual agent, is meant to handle the request itself. The labels get used loosely by vendors, so ask what happens after the system understands the caller rather than which acronym is on the page.
Is conversational IVR the same as conversational AI?
Not quite. Conversational IVR is one application of conversational AI: understanding a spoken request well enough to route it. Conversational AI more broadly covers systems that hold the whole conversation across phone, text and chat and complete the request. Routing is the narrow case.
Does conversational IVR reduce call volume?
Not by itself. It changes how calls are sorted before they reach your agents, which can cut misroutes and transfers between teams. Volume arriving at the queue only drops when something resolves the caller's request instead of passing it along.
Do we have to replace our IVR to use AI?
No. The common path is to add one option to the menu you already have and point it at an AI agent, then expand once you can see which call types it finishes. Replacing the menu wholesale is a bigger project with a slower payback and no extra benefit at the start.
Can a caller still reach a person?
Yes, and it should take one request. With the press-1 deployment, choosing a person keeps the caller in the normal queue with their place intact. When a caller who started with the agent needs a person, the call transfers to the number you nominated, and the person who picks up takes the call from there.
What does it cost to add an AI option to our phone menu?
Telvana starts at $399 a month based on call volume, plus a one-time setup fee that covers building, testing and launching the agent. For a conversational IVR bought as part of a contact center platform, ask for licensing and usage together rather than the headline rate, and ask what an interaction counts as.
Start with one option on your menu
Bring your current phone menu to a demo and we will go through it the way the worksheet does, option by option, and name the ones an AI agent could finish. You will hear it take one of your own call types on the call.