IVR in the age of AI: what the reports actually say

Press 1 for sales. Press 2 for support. Press 9 to hear this menu again.

The interactive voice response system, or IVR, was invented in the late 1980s and rolled out at scale through the 1990s. Three decades later, it still greets most business calls in the enterprise world.

The reports are unambiguous. Public industry surveys from 2024 and 2025 point to abandonment rates on multi-level IVRs that no one would defend if they were customer-acquisition costs. Satisfaction scores put IVR consistently among the lowest-rated customer touchpoints. And AI is finally mature enough to change the picture.

But the reports also warn against a lazy framing: AI does not simply replace the IVR. The job of the phone menu changes when the system can understand what a caller wants, in their own words, in real time. What used to be a routing exercise becomes an intent understanding exercise. That is a different problem, with different metrics and different value.

This article walks through what the public reports say, the four numbers that summarise the IVR problem, the concrete impact of switching to a conversational agent, the new KPIs to track, and a five-question decision framework you can apply this week.

What the public reports say (2024-2026)

Several publicly available reports converge on the same headline: touch-tone IVR is the customer touchpoint that ages fastest, and organisations that keep it in place at scale are absorbing costs they no longer see clearly.

Zendesk’s CX Trends report highlights year over year that voice remains a decisive channel for complex customer issues, and that friction on the voice touchpoint disproportionately damages loyalty. Salesforce’s State of Service report tracks the widening gap between customer expectations of instant, contextual service and the actual first-response performance of contact centres. McKinsey’s customer care research points to the shift from menu-driven to intent-driven routing as one of the highest ROI moves available to CX leaders. NICE’s Customer Experience report tracks the growing share of contact centres reporting active AI voice deployments. And the Microsoft Work Trend Index documents the acceleration of AI adoption inside enterprise operations, including customer-facing workflows.

The reports agree on two structural findings. First, callers are less patient with menu trees than they used to be, and the tolerance drops sharply after the second level of the menu. Second, self-service completion rates on legacy IVRs have stayed flat or declined for a decade, while expectations of self-service quality have not stopped rising.

Four indicators that summarise the IVR problem

You can spend a week reading contact centre reports. The signal reduces to four numbers most leaders can pull from their own logs in a morning.

Abandonment rate

On multi-level touch-tone IVRs, the public literature reports abandonment rates in the 30% to 50% range depending on industry and time of day. That share of callers hangs up before speaking to anyone. Every abandoned call is a lost lead, an escalation delayed, or a customer who calls back angrier.

Customer satisfaction score at the IVR touchpoint

Voice remains a preferred channel for complex or urgent issues, but the IVR experience consistently rates as one of the lowest-scoring parts of that channel. When surveyed on the specific IVR step, callers report frustration with menu length, misrouting and lack of a clear path to a human.

Average handle time

Handle time is inflated by pre-agent menu navigation and by the time an agent spends re-qualifying a poorly-routed call. Reports point to a meaningful share of agent handle time being spent recovering from IVR mismatches.

Cost per call

Cost per call is dominated by agent time, and misrouted or repeat calls compound that cost. Public research consistently identifies routing quality as one of the top three levers on cost per call, ahead of technology licensing.

Read your last 90 days of abandonment against menu depth. If abandonment steps up after the second level of the tree, you already have your answer.

Why AI doesn’t replace the IVR, it transforms its job

The classic IVR had one job: route the caller based on the button they pressed. It was a mechanical dispatcher, cheap to run at scale, brittle to design, and hostile to callers whose need didn’t fit the menu.

The AI voice agent has a different job: understand what the caller needs, in their own words, and route with context. What used to be a dispatch problem becomes an intent understanding problem, and everything else follows from that shift.

This is why the framing matters. If you set out to replace your IVR button-for-button with an AI equivalent, you inherit its limitations and you miss its transformation. If you set out to redesign the phone reception job around what the caller is actually trying to do, the IVR becomes obsolete as a side effect. Not the target.

The best public reports on the topic are careful to make this distinction. The report authors are not selling AI. They are describing what happens in production when the routing job shifts from menu to intent.

What AI lets you measure that the IVR never did

An AI voice agent generates signal at every step of the call that a touch-tone IVR could never produce. Some of that signal is transactional, some is behavioural, and both feed decisions you could not make before.

• Intent. What is the caller actually trying to do, in their words, with the confidence level of the classifier attached. This replaces the button press with a labelled intent that can be analysed, aggregated and improved.
• Sentiment. The emotional tone of the caller in the first few seconds, and how it evolves during the conversation. A caller who de-escalates during the call is a very different outcome from one who escalates.
• Menu deviation. The categories of requests that appear repeatedly but do not match any existing routing rule. These are the growth edges of your call taxonomy. The classic IVR could not surface them, because they never fit a button.
• Language and accent. The language the caller actually speaks, versus the languages your organisation currently supports. Multilingual deployments generate that gap analysis as a byproduct of running.
• Repeat context. Whether this caller has called before, and about what. This turns each call into a continuation, not a fresh start. The IVR treated every call as anonymous.

Each of these signals is directly actionable. Together, they turn the reception function from a black box into an observable system.

The concrete impact on call metrics

The public literature and Heedify’s own production observations point to a repeatable set of impacts when organisations switch a touch-tone IVR for a conversational agent. None of the numbers below are point predictions. They are observed ranges, and they depend heavily on call mix and starting baseline.

• Abandonment. Organisations moving from a multi-level touch-tone IVR to a conversational agent typically reduce their abandonment rate in a range of 20% to 40% according to public literature, sometimes more when the starting baseline was very poor.
• Time to first response. The conversational agent picks up the call and starts qualifying in under two seconds. There is no menu to hear, no button to hunt for. This alone changes the caller’s perception of the organisation.
• Multilingual coverage. A single agent handles 50-plus languages natively, at native quality on the top ten. Organisations that used to route international calls to offshore providers can consolidate reception in-house.
• 24/7 availability. The agent runs continuously without marginal cost. Missed calls become handled calls, and after-hours becomes a service window rather than a gap.
• Routing precision. Intent-based routing sends the caller to the right agent, queue or self-service path on the first attempt. Transfer rates drop, average handle time drops, first-call resolution improves.
• CRM enrichment. Every call arrives at the agent with its intent, sentiment, key entities and summary already logged in the CRM. Wrap-up time drops, and downstream analytics get a signal they never had.
• Fraud and priority detection. The agent surfaces patterns the IVR could not: unusual language mixing, atypical caller behaviour, high-urgency phrasing. These become inputs to fraud or priority queues.
• Agent cognitive load. Each call arrives pre-qualified. The agent starts with context, not with a customer repeating themselves. That is a quality of work improvement, not just a metric.

The impact is not evenly distributed. Organisations with deep menu trees and high international call volume see the largest shifts. Simple, single-language reception with a small number of routing paths sees more modest gains.

Reporting: the new KPIs to track once you move to AI voice

The KPIs you optimise on a classic IVR do not survive the transition. Some become irrelevant, some become misleading, and new ones become available.

KPIs that lose meaning

• Time in IVR. There is no menu to be lost in. This metric goes to near zero and stops being useful.
• Menu-level exit rates. The tree is gone. Replaced by intent categories.

KPIs that stay relevant

• Abandonment rate. Now a much cleaner metric because there is no menu tax to explain away.
• First-call resolution. Directly influenced by routing precision, easier to attribute.
• Average handle time. Cleaner input, because the agent starts with context.

KPIs that only exist now

• Intent detection accuracy. The share of calls where the agent classified the intent correctly, measured against human review.
• Self-service completion rate by intent. For which intents the AI handled the call end to end, without transfer.
• Sentiment shift during call. How the emotional tone of the caller evolves. A useful leading indicator for CX programmes.
• Menu deviation rate. The share of intents that do not map cleanly to your existing taxonomy. This is your growth edge.
• Language detected vs supported. How often callers speak in a language you had not planned for. Directly informs your language strategy.

A five-question decision framework

If you are trying to decide whether your IVR is costing your organisation more than it is worth, these five questions surface the answer in an afternoon of desk research.

• 1. What is your abandonment curve by menu depth? Pull the last 90 days of your IVR abandonment data. If the curve steps up sharply after the second level, your callers are telling you the answer.
• 2. How many languages do you support versus how many do your callers actually speak? The gap between the two is the size of the international business you are silently declining.
• 3. What does a misrouted call cost you? Add internal transfer time, agent time recovering context, and repeat call rate. Multiply by your misroute rate. That number is often a surprise.
• 4. What would an AI agent do differently on the top five call reasons you receive? For each of the five most frequent call reasons, describe the ideal handling. Compare with your current IVR routing. The delta is the value gap.
• 5. Do your current reports measure intent, or just clicks? If you cannot answer why callers are calling in structured, aggregable terms, you are managing your reception function blind.

Any three of these five answered honestly are usually enough for a business case.

What we observe at Heedify in production

The following are qualitative observations from Heedify AI Receptionist deployments in production. They are not benchmarks and they are not offered as universal truths. They are what we consistently see on the ground.

Callers do not adapt to the agent. The agent adapts to the caller. This changes the character of the reception function almost immediately. Testers who go in expecting to talk to a menu adapt in the first sentence.

Multilingual routing opens markets that were silently blocked. Organisations discover that a share of their inbound was going unanswered simply because it arrived in a language no one on shift spoke that day.

The recording and transcript archive becomes an asset. Once every call is transcribed, categorised by intent and summarised, the recordings become material for training, product feedback, and quality improvement, not just compliance.

The IVR mindset takes longer to unlearn than the IVR itself takes to switch off. The technology change happens in weeks. The reporting and process change to actually exploit the new signal takes a couple of quarters.

Frequently asked questions

Is the touch-tone IVR really dying?

Not everywhere at once, no. Simple, single-language reception with three or four routing paths still runs perfectly well on a classic IVR. What is dying is the multi-level menu tree, the mainstay of enterprise IVR for 30 years. It is being replaced by intent-driven routing wherever the call mix is complex or multilingual.

Can we keep a legacy IVR and add an AI layer on top?

You can, and some organisations do as a stepping stone. The value of AI as a surface layer over a legacy IVR is real but capped. The full impact comes when the routing job itself is redesigned around intent, not when a smart layer paraphrases the menu.

How long does it take to replace an IVR with an AI voice agent?

A pilot on a Teams Phone number can be live in weeks. A full production rollout depends on the CRM integrations, the knowledge base setup, and the internal change management. Public literature and our own deployments both point to under 90 days for a full rollout on a mid-sized reception function.

What is the main risk of switching from IVR to AI?

Under-preparing the change management. The technology transition is straightforward. The organisational transition is where projects stumble: rewriting call scripts, retraining agents on intent-based routing, updating reports, rebuilding SLAs against new KPIs. The projects that skip that work under-deliver, regardless of vendor.

Can you measure the ROI of an IVR replacement?

Yes, with the caveat that the ROI is spread across metrics. Direct saves come from reduced abandonment and misrouting. Indirect saves come from consolidated multilingual operations and after-hours coverage. Longer-term value comes from the new signal (intent, sentiment, menu deviation) feeding CX programmes. Build the business case with the three horizons in mind, not just the first.

What does AI change for sensitive calls (GDPR, health data, regulated industries)?

Sensitive calls raise the bar on data residency, transcription retention, and consent. Modern AI voice platforms address these with configurable data residency, granular retention policies, and explicit consent flows. Heedify is ISO 27001 certified and its AI voice deployments are designed to meet these requirements. The right question to ask a vendor is not “do you handle sensitive data?” but “where does the audio live, for how long, and who can access it?”

Ready to see how this applies to your reception function?

If you want a structured read on whether your current IVR is costing you more than it is serving, we can walk through the five-question framework against your data in a 30-minute session.

Book a demo → a working pilot in under 60 minutes.