Conversational IVR is a phone menu you talk to

The short version

• Conversational IVR replaces the numbered menu with speech: the caller states the request in their own words and the system acts on it.
• Understanding the sentence is the easy half. Finishing the task needs identity resolution and write access to the system holding the record.
• An AI IVR that cannot complete a task is an expensive way to collect a routing decision.
• Silence is the failure callers notice first, so response latency is a design constraint rather than a tuning detail.
• Containment rate, the share of calls that end without a human, decides the business case, and it can only be measured after launch.

What is conversational IVR?

It is an interactive voice response system that takes spoken language as its input instead of keypresses. The caller phones in, hears a short open prompt, says 'I need to move my delivery to Thursday', and the system acts on that sentence. There is no tree of numbered options to walk down. Vendors also sell the same thing as AI IVR, artificial intelligence IVR or voice AI, and the labels do not mean anything different.

Four things happen in sequence on every turn, fast enough that the caller should not hear the joins. Telephony delivers the audio. Speech recognition turns it into text. A model decides what the caller wants and which tool to call. Speech synthesis says something back. The loop repeats, with barge-in so the caller can interrupt instead of waiting out a prompt.

The step that decides whether you have built an IVR or a demo is the third one. A system that understands 'I need to move my delivery' and then says it will put the caller through to someone has parsed a sentence and achieved nothing. A system that finds the order, reads the available slots, moves the delivery and confirms the new date has replaced the call. Both sound identical in a thirty second video.

How is conversational IVR different from the speech IVR we already had?

The old systems understood a list. Conversational IVR understands a sentence. Directed dialogue IVR, the 'say or press one' generation, matched speech against a fixed grammar written per prompt, which is why callers learned to say the magic words slowly and in a flat voice. Anything outside the grammar fell through to a repeat of the prompt and then a transfer.

What changes with a model is paraphrase and repair. The caller can give two intents in one breath, correct themselves mid-sentence, or describe the problem without naming the product. The system can ask one targeted clarifying question rather than restarting the prompt. That is the whole user-visible difference, and it is a large one.

What also changes is the failure mode, in a direction that needs design work. A grammar could only mishear within its own list, so the worst case was a wrong branch. A model can ground a request in the wrong record and then act on it confidently. The mitigation is unglamorous: confirm specifics back to the caller before anything irreversible, and keep the set of actions the system can take deliberately small.

How the three generations of IVR differ

What does an AI IVR system need before it can finish a call?

Access, mostly. The language model is now the cheapest and least risky part of the build. What decides whether a call ends resolved is whether the system can identify the caller, read the record, write the change back, and show afterwards that it did.

In practice that means four pieces of plumbing. Identity resolution, verified to whatever standard the task actually requires rather than whatever is convenient on a phone call. A read path into the system of record. A write path that is idempotent, because callers repeat themselves and lines drop mid-transaction. And a defined escalation, including what the human receives when the call arrives.

This is why conversational IVR projects are integration projects wearing a voice interface. If nobody can name the API that completes the task, the model cannot complete it either, and the honest version of the system is a router that does a nicer job of asking why you called. That is still worth something. It is worth much less than the brochure implies.

What do published surveys actually say about callers and AI on the phone?

Less than the pages quoting them suggest. The IVR frustration statistics that circulate come mostly from vendor-commissioned consumer surveys, and the method is usually in the fine print rather than the headline. One of the most quoted is eGain's own survey, fielded on SurveyMonkey in February and March 2021 with 500 US respondents who had phoned a business or government agency from a smartphone in the previous year. Small sample, vendor sponsor, self-reported recall, and now several years old. We have not reproduced its findings and would not build a business case on them.

The more recent work is also vendor-run, and more usefully framed. Five9's 2025 report surveyed 1,006 consumers across the US, UK and Canada through Zogby Analytics, and its own headline is about balancing AI with human connection rather than replacing it. Its 2026 follow-up, fielded with Hanover Research across 3,000 consumers and 600 customer experience decision-makers in the US, UK and Germany, puts AI adoption in customer experience at 92% while stating that consumer trust still depends on human support.

Read those two together and what survives is a design requirement, not a market statistic. Adoption is widespread enough that callers are no longer surprised to meet a machine. What they keep asking to have protected is the route to a person. So build the transfer path first, and treat any single percentage you see quoted about caller frustration as someone else's survey until you have read its method.

When is conversational IVR not worth building?

When call volume is low, when nearly every caller wants the same one thing, or when nothing behind the phone has an API. Those three cover most of the projects that should not start.

If almost all calls are a single intent, a one-line menu or a published direct number beats any model on both speed and cost. If volume is low, note that the expensive part does not shrink with it: you still need an evaluation set, monitoring, and someone to read the calls that went wrong, and that cost is largely fixed. If the task behind the call is a human reading three screens and retyping, fix that first or accept that you are buying routing.

There are also calls that should reach a person quickly by design. Anything where identity verification has to be strong, anything regulated enough that a wrong answer is a liability, and anyone distressed or in an emergency. Say so in the opening prompt rather than making them fight the system to find out. And if the real complaint about your current IVR is that the menu is badly organised, rewriting the menu is a week of work, not a quarter.

What does conversational IVR cost to run?

Per minute, plus a build, plus an evaluation habit you keep paying for. The per-minute side is telephony, speech recognition, model tokens, speech synthesis, and whatever the platform charges per session on top. It is metered, so cost scales with call length. A system that takes longer to fail than a menu did can cost more and deliver less.

List prices are published and they move. Twilio's conversational AI pricing page, for instance, notes that pricing is current as of October 2026, which is a fair indication of how much any per-minute figure you read is a snapshot. Price the current rates yourself, against your own call mix and average handle time, before anyone builds a model in a spreadsheet.

The comparison most buyers want is against the cost of an agent handling the same call. If you want a defensible benchmark for that, the US Bureau of Labor Statistics publishes employer-reported wage data covering roughly 2.7 million customer service representatives, which is a sounder denominator than a vendor saving claim. The arithmetic that matters is still marginal: cost per contained call, including the calls the system held for a while and then transferred anyway.

What to measure in the first month

• •Containment rate broken down by intent, not as one blended number.
• •Transfer latency, from the caller asking for a person to a person speaking.
• •Mis-grounding: how often the system acted on the wrong record or the wrong order.
• •Repeat contacts within two days, which is where a false containment shows up.
• •The exact points in the call where callers go silent or hang up.
• •Cost per contained call, with the cost of transferred calls included.