Only 14% of customer service issues are fully resolved in self-service, according to a Gartner survey of 5,728 customers, and the touch-tone IVR is the phone version of that failure. Natural-language voice triage replaces the menu with one open question, works out what the caller wants from their own words, and either finishes the task on the call or transfers with context attached. It only feels natural if it answers inside the roughly one-second window humans use to take turns. In Metrigy's 2025-26 study of 656 companies, 37.6% already plan to fully replace their IVR with AI triage agents.
Key takeaways
Phone self-service is failing at scale: 14% of issues fully resolve in self-service, and even issues customers call very simple resolve only 36% of the time (Gartner).
Count the pre-queue leg as abandonment. On the federal 988 crisis line, 44% of calls were abandoned before they were even routed to a local center (GSA evaluation).
Latency budget: humans change speaking turns in about 200 ms, and ITU-T G.114 wants mouth-to-ear delay under 150 ms and never above 400 ms. Hold a voice agent to a sub-second response at p95, not on average.
Coexist if Genesys, Five9 or Amazon Connect already runs your queues and workforce tools. Replace if the IVR is a Twilio application your own engineers maintain.
Keep keypad input for digit strings, noisy environments and anyone who asks. Twilio and Amazon Connect both accept speech and DTMF in the same prompt.
Why do callers abandon touch-tone IVR menus?
Because the menu makes the caller do the company's routing work, and most people either give up or route around it. Gartner's data shows the shape of it: 73% of customers use self-service at some point, but only 14% fully resolve there, and only 36% of the issues customers describe as very simple. The phone menu is the least forgiving form of self-service, so that is where the numbers bite hardest.
Three mechanics drive the drop-off. The caller has to know your org chart: a failed payment could be billing, accounts or fraud, and the menu asks them to guess. Menus are linear and problems are not: a caller with a declined card who also needs to change an address gets one path. And zeroing out is a tax, not an escape: the caller waits through the tree, presses 0, and lands on an agent who starts cold.
The loss happens before a human is anywhere near the call. When the US General Services Administration evaluated the 988 Suicide and Crisis Lifeline, it found that about 44% of calls were abandoned before the call was routed to a local call center, and 11% of the remainder while waiting for a counselor. That is a crisis line, not a billing line, but split your own abandonment by leg and you will usually see the same pattern: the front door loses more callers than the queue.
Demand is moving the wrong way for a menu to cope. Zendesk's compilation of industry data reports that 57% of business leaders expect call volumes to increase by as much as one-fifth over the next one to two years (a figure Zendesk attributes to McKinsey). The US Bureau of Labor Statistics projects that employment of customer service representatives will decline 5 percent from 2025 to 2035. More calls and fewer people means the front door has to do more than sort.
How does natural-language voice triage route a call?
It replaces the menu with a single open question, then runs a four-stage loop on whatever the caller says.
1. Speech to intent
Audio is transcribed as the caller speaks, and a language model reads the transcript for intent and entities: what they want, plus the order number, card or date they mentioned along the way. "Someone used my card, the one ending 4-4-2-1, somewhere I have never been" becomes a fraud-report intent with the card already captured. The caller never had to find the word fraud in a menu.
2. One clarifying question, not another menu
If the request is ambiguous, the agent asks in plain language: "When you say the transfer is stuck, do you mean it has not arrived, or it was declined?" One follow-up replaces three menu layers, and the agent copes with a caller changing their mind mid-sentence.
3. Route or resolve
This fork defines the product you are buying. A routing-only system classifies the intent and transfers to the right human queue with a structured summary. A resolving system keeps going where the workflow allows it: authenticate the caller, look up the account, file the dispute, confirm the new address. Only the second model removes calls from the human queue. The distinction between deflection and resolution is the one to write into your vendor contract.
4. Act, confirm, log
When the agent acts, it calls the systems of record through scoped, least-privilege tools and reads back what it did. When it routes, it hands over the intent, the entities, the transcript and anything it already retrieved. Every step is logged, which is what makes the call reviewable later.
How fast does a voice agent need to respond?
Fast enough that the caller does not notice a gap, which means a sub-second response from the end of the caller's turn to the start of the agent's. The research on human conversation sets the bar. Levinson and Torreira's review in Frontiers in Psychology finds that the gaps between turns are short, of the order of 200 ms, while the latencies involved in language production are much longer, over 600 ms. People manage that by predicting when your turn will end.
Telephony engineering reached a similar number decades ago. ITU-T Recommendation G.114 states that if mouth-to-ear delays are kept below 150 ms most applications experience essentially transparent interactivity, and that one-way delay of 400 ms should not be exceeded for general network planning. That budget covers the network alone. A voice agent adds transcription, a model call, often a tool call into a backend, and speech synthesis on top.
The rules that follow: measure latency at the 95th percentile, because the slow 5% is where callers start talking over the agent; stream every stage rather than waiting for a full transcript; and have the agent acknowledge before a slow tool call instead of leaving dead air. If a vendor quotes an average, ask for the distribution (see our sub-second latency guide).
Should the AI coexist with Genesys or Amazon Connect, or replace the IVR outright?
Coexist when a contact center suite already owns your queues, recording and workforce management. Replace when the IVR is application code your own team maintains. The decision follows the stack, and it usually splits three ways.
A contact center suite such as Genesys or Five9. Keep it. The voice agent takes the first leg of the call and, when it needs a person, transfers into the existing queue with the context summary attached, over SIP or the platform's transfer API. Reporting, recording and the agent desktop do not change; the menu tree does.
Amazon Connect. Coexist inside the flow. Amazon's documentation describes the Get customer input block as accepting DTMF input and an Amazon Lex bot, so the flow already has a slot where a natural-language agent replaces the digit prompt and hands the call back for queueing or callbacks as before.
A Twilio-built IVR. This is usually the case for outright replacement. If the tree is a set of TwiML Gather verbs written by your engineers, every menu change is a deploy. A streaming voice agent takes the inbound call from the same number, and Twilio stays as the carrier and transfer path. Twilio's Gather documentation lists dtmf, speech and dtmf speech as the accepted input modes, which matters for the fallback question below.
Whichever path you take, insist on one workflow engine shared with chat and email; the integrations page lists the telephony platforms that support the coexist pattern.
What changes for healthcare call lines?
The contract does. Under HIPAA, a vendor that hears protected health information on a call is a business associate, and HHS guidance is explicit that a covered entity may only disclose PHI to a business associate if it obtains satisfactory assurances in the form of a business associate agreement. Ask for the BAA before the pilot, and confirm PII redaction applies to transcripts as well as recordings. The healthcare page covers the rest of the checklist.
When do you still need DTMF fallback?
Whenever the caller is reading out a digit string, whenever the audio is bad, and whenever the caller asks for it. Natural language should be the default, not the only door.
Digit strings. Account numbers, dates of birth and the last four of a card are faster and more reliable on the keypad, and should be read back before use either way.
Noisy environments. A parent calling from a car, a caller on speakerphone. Speech recognition degrades; keypads do not.
Accessibility and preference. Some callers cannot or would rather not speak their request aloud. Offering "or press 1" costs nothing.
Legacy steps you keep. If a downstream payment or authentication step still expects tones, the agent needs to pass DTMF through.
Twilio's Gather verb accepts both modes in one prompt and, in its own words, gives precedence to the first input it detects. The rule to write down: every prompt that collects a digit string offers both, and every digit string is confirmed before it is used.
What should you measure in the first 30 days?
Abandonment by leg, mis-routes, and whether the calls you chose to resolve stayed resolved. Blended containment hides all three.
Metric | How to compute it | Day-30 rule |
|---|---|---|
Abandonment before routing | Calls dropped before a queue or a resolution, divided by calls offered | Below the IVR baseline for the same call mix |
Mis-route rate | Calls transferred again after the first routing decision, divided by routed calls | Below the IVR baseline; investigate any intent above 10% |
Resolution rate on enabled call types | Calls closed without a human and no repeat contact in 7 days, divided by calls of that type | Report per call type, never blended with routed calls |
Repeat-contact rate | Callers who contact again about the same issue within 7 days | Flat or falling against the pre-launch baseline |
Human-request honor time | Seconds from a caller asking for a person to the transfer starting | One turn; any retry loop is a defect |
p95 response latency | End of caller speech to start of agent speech, 95th percentile | Under one second on live traffic |
Escalation quality | CSAT on calls the agent transferred, versus calls that reached a human directly | No gap; a gap means context is lost in handoff |
Set expectations against the market, not the demo. Gartner forecasts that agentic AI will autonomously resolve 80% of common customer service issues without human intervention by 2029. That is a direction over several years, not a day-30 target. In month one, falling pre-queue abandonment and a clean mis-route number prove the intent layer works; resolution rate is month two's story. Automated post-call QA on every call is what makes the table trustworthy rather than sampled.
What this looks like in practice: a worked example
Wonderschool runs a marketplace for childcare providers, and its inbound line had two problems at once: only around 10% of parent calls were being answered, and roughly half of inbound volume was scam noise that a triage team had to filter before reaching a real parent. That is the IVR failure mode in miniature.
The company deployed a Lorikeet voice agent as the first leg of every call. A parent says why they are calling; the agent pulls the provider and enrollment context, answers or completes the step where a workflow covers it, and routes the rest with the summary attached. The published result is that 100% of parent calls are now answered, up from around 10%, and the scam half of inbound is handled with zero triage time.
Lorikeet combines deterministic structured workflows for steps that must run the same way every time (identity checks, lookups, disclosures) with natural-language workflows for the open parts of the conversation, in one interaction. Keypad input is supported for legacy steps. It coexists with Genesys and Amazon Connect or replaces a Twilio IVR stack, and integrates with Talkdesk and Salesforce where those hold the queue. Voice responses are sub-second in the US, UK and Australia. Before launch the agent runs through pre-launch simulations and red-team scenarios; after launch Coach reviews 100% of calls post-hoc. Security posture (SOC 2, BAA available, PII redaction, data residency in the US, AU and UK) is on the trust page. Pricing is per resolution, about $1.20 per voice resolution, unresolved tickets cost nothing.
The honest limitation: this is not a same-day switch. Lorikeet expects to be wired into your telephony and systems of record, and the first weeks of a regulated deployment route more calls than they resolve while workflows are proven one at a time. A simple hours-and-directions line does not need that depth. If your hardest calls are the regulated ones, walk through a live call with one of your own call reasons before deciding.
What still needs a human
Anything where the caller asks for one, anything involving distress, and any decision a regulator would expect a person to own. Write the triggers down before launch.
An explicit request for a person. Honored on the first ask, with no "let me try to help first".
Distress and safety. Crisis language, a medical emergency, a frightened caller. A slow front door here costs more than a CSAT point.
Judgment calls. Whether a dispute is valid, whether a hardship arrangement is fair, what a symptom means. The agent gathers and organises; a person decides.
Low confidence after one clarifying question. Route rather than guess. A confident wrong route is worse than a short transfer.
Anything outside a configured workflow. Everything else is a warm transfer with the transcript attached.
Replacing the menu is a smaller decision than it looks, and a more measurable one. Start with the front door, split abandonment by leg, hold the agent to a sub-second response, and widen resolution scope only when the day-30 numbers say the intent layer has earned it.
