AI Voice Agent vs IVR: Why “Press 1” Menus Are Falling Behind

A caller spots a double charge on last month’s bill and rings the helpline. The menu says: “For billing, press 1. For a new connection, press 2. To hear these options again, press 9.” Nothing fits, so they press 0 repeatedly, or hang up.

This article is for teams deciding whether to replace an IVR (interactive voice response, the “press 1” phone menu) with an AI voice agent.

What Changes When Callers Can Just Say the Problem

• Menu IVR: digits move the caller through a fixed prompt tree.
• Speech-enabled IVR: the caller speaks, and keywords or a few predefined intents drive a fixed flow.
• Conversational IVR (often sold as AI-powered IVR): natural-language understanding picks the intent, often then routing into predefined flows.
• AI voice agent: holds a multi-turn conversation, calls backend systems, completes tasks and hands over with context.

The first three suit a handful of well-understood requests, but designers must anticipate every path, so a caller with two issues, or one who describes the problem in their own words, often falls outside the flow. An AI voice agent works from the conversation, so it can resolve requests no branch was built for. Some products sold as “conversational IVR” only replace the first menu, so ask vendors which one you are buying.

How Fast an AI Voice Agent Has to Answer

Most production voice agents are a chain. The Moshi paper describes voice activity detection, speech recognition, a text dialogue model and text-to-speech, and each stage adds delay: speech-to-text (STT) waits for the caller to finish, the language model (LLM) takes time to reply and may call tools, text-to-speech (TTS) produces audio, and the network adds time on every hop.

Human turn-taking is a reference point, not a target

In a ten-language study, Stivers and colleagues found that the most common gap between turns was 0 to 200 ms in every language, with an overall mode of 0 ms and a mean of +208 ms. That describes people taking turns with each other. It is not a latency requirement for AI voice agents, and it is not end-to-end response time, which includes every stage above.

The Moshi authors say cascaded pipelines cause “a latency of several seconds between interactions”, while their full-duplex research model reports 160 ms theoretical and 200 ms in practice. That is a research result, not a phone-line guarantee: PSTN (phone network) or SIP (internet telephony) transport and codecs add delay. How fast is fast enough depends on the interaction, the architecture and the caller’s experience.

Commercial models also change quickly: OpenAI made its Realtime API generally available with gpt-realtime on 28 August 2025, and its deprecations page now lists gpt-realtime for shutdown on 20 January 2027, with gpt-realtime-2.1 as the replacement. Check current model lists before you commit to one.

One Caller, Two Systems: Menu IVR vs AI Voice Agent

Take the double-charge caller through both systems (an illustration, not a measured result). On a menu IVR, they guess “billing”, press 1, and meet a second menu where nothing says “I was charged twice”, so they press 0 and wait for a human. With an AI voice agent, they say it in their own words, and the agent asks for what is missing, checks the account, then answers or hands over with context.

Multilingual AI voice agents for code-mixed callers

A “press 1 for Hindi, 2 for English” menu assumes one language per call, but many Indian callers switch mid-sentence: “mera refund abhi tak process nahi hua, can you check the status?” Keyword menus handle that poorly, as do models trained on clean, single-language audio.

Sarvam’s models show why the version matters. Each word error rate (WER) below is vendor-reported, tied to one model and benchmark, and not comparable across rows:

Model	Status in Sarvam’s docs	Reported result	Benchmark
Saarika v2.5	Being deprecated	18.32% WER across 11 languages	VISTAAR
Saaras v3	Previous model, still available	About 19% WER	IndicVoices
Saaras v4	Latest, default, recommended	Lowest WER reported, 16.03%, in the keyword-prompting (L5) setting with a domain keyword list supplied	IndicContextEval

Statuses: Sarvam’s model list and Saarika page; figures: Saaras v3 and v4 launch posts. Test recognition on your own recorded calls before going live. Zoice agents speak ten Indian languages, including Hindi, Tamil and Telugu, plus English (India), with native Hinglish (Hindi mixed with English) support and mid-conversation language switching; see the language support page.

How to Replace the Menu Without Replacing Your Telephony

Replacing the IVR is a menu project, not a telephony project. The carrier, the numbers customers dial, the SIP trunks and keypad handling can all stay (see diagram).

1. The caller dials your existing business number (DID).
2. The call reaches the AI voice agent over your carrier’s SIP trunk.
3. One pipeline detects speech, transcribes, reasons and speaks.
4. The agent calls backend systems such as CRM or billing.
5. It resolves the request or transfers to a human queue.
6. Keypad (DTMF) input stays available.

Keep the keypad for PINs and OTPs

DTMF tones (the tones a keypad press sends) are carried as signalling events under RFC 4733, which gives “a gateway passing DTMF tones to an IVR” as an example of digits that must arrive reliably. Low-rate voice codecs cannot be guaranteed to reproduce the tones accurately enough for automatic recognition, so events are usually the more dependable path. In Dialogflow CX, with DTMF enabled, keypresses become normal text input.

Check signalling and media encryption separately

SIP, the protocol that sets up calls, can run over UDP, TCP or TLS. TLS protects the SIP signalling, the call-setup messages. The audio travels separately as RTP, and SRTP is what protects that media stream (RFC 3711; Twilio also treats them as separate protections). So a TLS trunk does not by itself guarantee encrypted audio: verify SRTP negotiation and configuration independently, because the right setup depends on your telephony provider and SIP architecture. On Zoice, SRTP media encryption is available when the trunk uses TLS.

Migrate one request at a time

Replace the top-level menu with an open prompt that routes into your existing queues, then automate the highest-volume requests one at a time. According to Twilio’s documentation, ConversationRelay converts caller speech into text for your application, which generates the reply using AI; Twilio then speaks that reply to the caller. Set handover rules for callers who ask for a person, repeat a failed request or are clearly distressed, and for requests outside policy.

On Zoice, you can bring your own carrier (BYOC): register SIP trunks over UDP, TCP or TLS, choose a DTMF mode per trunk, and bind each DID (phone number) to an AI voice agent. transfer_agent passes a call to another AI specialist and transfer_number to a human phone number, with an optional announcement; see the SIP / BYOC page. Because call audio passes through speech and language models, where it is processed is a compliance question; our data residency post covers it.

Measure Outcomes, Not Calls Handled

A count of calls the AI answered says little, since badly handled calls still count. Measure IVR automation by outcomes against your current baseline, and read containment alongside abandonment, since callers who give up can look like resolved calls.

KPI	What it measures
Call containment rate	Calls resolved without a human
First call resolution (FCR)	Issues resolved first time
Average handling time (AHT)	Average time per interaction
Transfer rate	Calls passed to human agents
Cost per resolved call	Operating cost divided by resolved calls
Abandonment rate	Callers who hang up before resolution
Task completion rate	Intended tasks completed
Customer satisfaction (CSAT)	Satisfaction after the call

A hypothetical example, with made-up numbers: if your IVR resolves 40 of every 100 calls without a human and a pilot resolves 55, containment is up 15 points; also check transfer rate, abandonment and CSAT, and compare cost per resolved call.

A Published Deployment: Kaushgar Setu’s Helpline

According to Zoice’s customer stories, Kaushgar Setu’s helpline runs on a Zoice voice bot that identifies callers from their number, reads out their wallet balance and training status, and files SLA-tracked support tickets without a human picking up. This post does not quote results for this deployment.

How to Decide: Keep the Menu or Add an Agent

A menu IVR can still be the right tool. Use these checks:

• Keep the menu when callers have a few well-understood reasons for calling.
• Add an agent when call reasons vary, callers raise several issues or mix languages, or you need transcripts to learn why people ring.
• Either way, keep keypad entry for PINs and OTPs, and replace only the top-level menu first.

Frequently Asked Questions

Is an AI voice agent cheaper than an IVR?

Per call, an IVR costs less, but compare cost per resolved call: an AI voice agent costs more per minute and can finish requests an IVR can only route. We could not verify the widely repeated analyst forecasts of contact-centre savings from a primary source, so model your own. On Zoice, each conversation shows its cost in ₹ with a per-minute breakdown; see the pricing page.

Can an AI voice agent take a PIN or OTP on the keypad?

It should, and for secrets the keypad is often better, since a spoken OTP ends up in the transcript. On Zoice, recording redaction can also strip Aadhaar, PAN and OTP patterns before transcripts reach storage.

Is conversational IVR the same as an AI voice agent?

Not always. Test any product with three questions: can a caller raise two issues in one sentence and get both handled, can the system act through your back end or only route, and what happens when the caller goes off-script? A product that only routes is a better front door, not an agent.

Can an AI voice agent understand callers on a bad phone line?

It depends on the recognition model, so test it on recordings from your real callers, since phone audio is narrowband and often noisy. Zoice adds caller-side noise suppression.

Bring a month of IVR logs and your top call reasons, and we will help you pick the first request to automate. Talk to our team.