IVR vs AI voice agent: why customers press zero

Everyone has pressed zero. The menu did not have the option, the option led to another menu, or the recording asked for an account number the caller did not have. The IVR was built to route calls cheaply, and it does, at the cost of the caller's patience. A voice agent is built to resolve calls: the caller says what they want, the agent understands, acts and confirms. This article compares the two honestly, including what an IVR still does well, and lays out how to migrate without a big-bang cutover.

Side by side

Why callers abandon IVRs

Menus are written from the company's org chart, not the caller's problem. The option the caller needs is usually the third one in the second menu, and the caller has already guessed wrong once. Authentication is demanded before the caller knows whether the system can help at all. And there is no way to say "I just need to move my appointment to Thursday". Every one of those is a design consequence of routing rather than resolving.

What an IVR still does well

• Compliance announcements that must be heard verbatim
• Routing by known account state where no conversation is needed
• Very high-volume, very simple prompts ("press 1 to confirm") on outbound
• Cost control for callers who genuinely need a queue, not an answer

A good migration keeps those. The voice agent takes the resolvable intents; the IVR, or a minimal version of it, remains for announcements and queue routing.

The migration path

• Measure the IVR: which options are chosen, where callers press zero, which intents dominate transfers
• Pick the top intents: usually booking, status, hours, reschedule; three or four intents cover most volume
• Build the agent behind the IVR: the first option routes to the agent; callers who choose it get resolution
• Move the greeting: once the agent handles the top intents well, it answers first and offers the old menu only as a fallback
• Retire menus as intents are absorbed; keep compliance prompts

Cost, honestly

Per minute, a voice agent costs more than an IVR: roughly $0.05–0.15 versus a cent or two. Per *resolved call*, it costs less, because the IVR minute is followed by a queue minute and an agent minute, or by an abandoned call. Model both from your own volumes; see How much does an AI voice agent cost per minute?.

What changes for the contact centre

Agents receive transfers with transcripts and a summary instead of cold calls; routine volume falls and the remaining calls are the complex ones; after-hours volume is handled rather than voicemailed. Staff who used to answer "what are your hours" thirty times a day are freed for the calls that need them, and they own the agent's phrase lists and pronunciations. Adoption depends on involving them from the first week, as in the hospital network rollout.

A worked example

A service company's IVR had seven options and a 38% zero-press rate. Analysis of a month of calls showed three intents made up most of the volume: booking a visit, checking a visit time, and rescheduling. The agent took those three behind option one, resolved most of them without a person, and then became the greeting with the IVR as fallback. Transfers arrived with transcripts, after-hours bookings appeared in the morning queue, and the remaining menu was two compliance prompts. The company kept the IVR vendor for those prompts and the queue.

Team and timeline

Six to ten weeks with a conversational AI engineer, a telephony engineer to integrate behind the existing IVR, an integration engineer for the system of record and a delivery lead. The IVR analysis in week one decides the intents; the first release lives behind option one; the greeting moves when the dashboard justifies it.

Before you start: a checklist

• IVR option and zero-press reports for the last three months
• Recordings of transferred calls for intent analysis
• Which prompts are compliance-mandated and must stay
• Access to the telephony platform for routing changes
• The system the agent will act in, with API or staging access
• Contact-centre staff to review the first week's calls

Reading your IVR reports before you decide

Most contact-centre platforms report option selections, zero-presses and transfers. Read three months of them alongside a sample of transferred-call recordings. You are looking for three things: the intents that dominate, the menu paths where callers give up, and the calls that reach a person only to ask something the system could have answered. That analysis, which takes a week, is the scope document for the agent and the evidence for the business case; without it, migrations are opinions.

Cost of the status quo

Abandoned calls are not free: they become repeat calls, emails and complaints, and for booking-led businesses they are lost revenue. Put a number on it from your reports: abandonment rate times call volume times the value of a resolved call. That figure, next to the per-resolved-call cost of a voice agent from the cost breakdown, is the business case. Most companies find the IVR was never cheap; it was just invisible.

What to do with the recordings

Transferred-call recordings are the raw material for the agent's intents, phrasing and evaluation set. Cluster them by request, note the words callers actually use rather than the words the menu used, and mark which ones a person resolved with a lookup the agent could do. That set, transcribed and labelled, becomes the golden conversations the agent is tested against before it takes a single live call, and it is why the first week of a migration is listening rather than building.

Glossary

• Zero-press rate: share of callers who abandon the menu for a human
• Intent: what the caller wants, stated in their own words
• Containment: the IVR-era term for calls that never reach a person; replaced by resolution
• Warm transfer: hand-off with transcript and summary
• Compliance prompt: an announcement that must be played verbatim
• Fallback menu: the minimal IVR kept behind the agent

Mistakes we see

Migrations fail when they replace the IVR wholesale on day one, when compliance prompts are lost, when the agent is given every intent instead of the top three, and when contact-centre staff first hear about it at launch. Keep the IVR for what it does well, move intents one at a time, and put staff on the recordings in week one.

Questions clients ask

• Will callers be confused by the change? Less than by the menu; saying what you want is the natural behaviour the IVR suppressed.
• Can we keep our queue and skills routing? Yes; transfers go into the same queues with the same skills, plus context.
• What about callers who want to press buttons? Keypad input is still accepted for confirmations and the fallback menu remains.
• Does it work with our contact-centre platform? Over SIP or through the provider, without replacing the platform.
• How do we know it is working? Per-intent resolution, transfer reasons and the zero-press rate of what remains of the menu.

What good looks like after 90 days

At ninety days the agent answers first, handles the top intents without a person, transfers arrive with context, after-hours volume is handled, and the remaining menu is compliance prompts and queue routing. The zero-press rate becomes a historical number.

Related reading

What is an AI voice agent?, AI voice agent for appointment booking, the voice agent service and pricing pages. For the customer-experience research behind menu abandonment, Nielsen Norman Group publishes the usability studies most teams cite.

The IVR asks callers to learn your menu. The voice agent asks them what they need. Migrate one intent at a time and keep only the prompts that must be heard.

For decision-makers, the test is simple: call your own IVR with the three most common requests and count the presses. Then imagine saying the request once. That gap is what the migration closes.