TL;DR
- First-attempt pickup in one insurance operation sits at 14%, rising across a multi-attempt sequence
- Cold-call answer rates run around 28%, and mobile numbers answer at 8.1% against 12.7% for direct dials
- 68% of consumers say messaging has replaced some of their calling, and 18 to 24s prefer text 49% to 19%
- Consumers 55 and over still prefer calls at 35%, so the channel splits by age rather than dying
- The FCC ruled in February 2024 that AI-generated voices are artificial under TCPA, requiring prior express consent
- India's TRAI Second Amendment in 2025 mandates DND scrubbing and dedicated 140-series numbers for commercial calls
An insurance team told us their first-attempt pickup rate is 14%, and that only 55 to 65% of their calls get answered at all across the full sequence.
That is the number the voice AI category is quietly built on. Not that calling is obsolete, but that most calls die unanswered, and the answer everyone is selling is a faster way to place more of them.
Voice AI for sales is real and it works. It works in a narrower band than the category implies, it is more regulated than most vendors mention, and it fails outright in at least one vertical where the pitch sounds most compelling.
What every voice AI vendor is actually selling
The top-ranking players in this category sell infrastructure for building a voice agent. Low latency, natural turn-taking, realistic speech, a builder to define the flow.
That is genuinely hard engineering and they do it well. It is also not the same product as a system that knows who it is calling.
A voice agent built on those tools starts every call cold. It knows the number, the script and whatever variables got passed in. It does not know that this person abandoned an application at the document step on Tuesday, replied to a WhatsApp message on Wednesday, and has already been called twice this week.
The call quality is excellent. The conversation is still a stranger's.
The pickup problem nobody solves with better voices
Answer rates set the ceiling on everything else, and they are low.
Cold-call answer rates sit around 28%. Mobile numbers answer at 8.1% against 12.7% for direct dial. Local-presence dialling lifts connection rates by roughly 20 to 30%, which is real and also the kind of gain that regulation is steadily closing off.
Against that, the economics of doing it with people are unforgiving. A US contact-centre agent costs $28 to $42 an hour, an in-house rep $30,000 to $60,000 a year, and offshore alternatives around $14 an hour.
That arithmetic is the honest case for voice AI. Not that it converts better than a good human, but that a 14% first-attempt pickup rate makes human dialling an expensive way to reach voicemail.
Voice is splitting by age, not disappearing
The generational data is the most useful thing in this category for planning purposes.
68% of consumers say messaging has replaced some of their calling. Among 18 to 24s, 49% prefer text against 19% for calls. Among those 55 and over, 35% still prefer calls (YouGov, February 2026, 2,442 respondents).
That is not a channel dying. It is a channel segmenting. A revenue team that goes voice-first for a young D2C audience is fighting preference, and one that goes messaging-only for a 55-plus insurance book is doing the same thing in the other direction.
The decision is per-segment, and it is a decision most tooling does not let you make, because it sells you a voice agent or a chat agent rather than a journey that picks.
The regulation nobody mentions in the demo
This section is short and it should change how you evaluate vendors.
In the United States, the FCC ruled in February 2024 that AI-generated voices count as artificial under the TCPA. That means prior express consent is required, penalties run $500 to $1,500 per violation, and since April 2025 consent can be revoked in any reasonable manner.
In India, TRAI's Second Amendment in 2025 mandates DND and preference-registry scrubbing and dedicated 140-series numbering for commercial calls, with movement toward mandatory AI self-identification.
The practical consequence is that a voice deployment needs consent capture, registry scrubbing, opt-out handling and disclosure built in from the start. Any vendor whose demo skips this is selling you a compliance problem with a pleasant voice attached.
It also settles a question the category likes to leave open. Sounding indistinguishable from a human is not a feature to aim for. In both markets the direction of regulation is toward disclosure.
Where voice actually fails
The most useful thing in our own call corpus is a counter-example, and it deserves more prominence than a vendor would normally give it.
A hospital group reported that a previous vendor's voice AI pilot reduced their conversions. Their read was that people dealing with a health concern want a human, and the automated voice made that harder rather than easier. The same operation also reported roughly 50% of calls going missed or unreachable.
We are including that because it is true and because the alternative is pretending voice is universally welcome. It is not. In healthcare, and in any high-anxiety or high-consideration purchase, an automated voice can cost you the conversion it was deployed to win.
The lesson is not that voice AI fails. It is that channel choice is a decision the journey should make based on segment and context, rather than a product category you buy once.
Where voice belongs in a revenue journey
Four places where it earns its keep, all of them defined by stage rather than by channel preference.
First response, where speed matters more than polish and the alternative is a lead sitting untouched for hours.
Qualification, where a short structured conversation establishes fit before a human is spent on it.
Re-engagement, where someone stalled mid-journey and messaging has already failed to move them.
Renewal and collection, where the conversation is expected, the customer relationship already exists, and consent is uncontroversial.
What makes these work is not the voice. It is that the agent arrives knowing what happened before the call and writes back what happened during it. Zigment's voice agents run on existing numbers and dialers, handle consent capture and voicemail, transfer to a human with the transcript and score attached, and write the outcome back to the CRM so the next step in the journey is informed by the call rather than blind to it.
That is the difference between a voice agent and voice as one stage of an orchestrated journey. We sit on top of HubSpot and Salesforce precisely so the call is not a separate system with its own memory.
How to evaluate this honestly
Three questions worth asking any voice AI vendor.
What does the agent know before it dials?
If the answer is a row from a spreadsheet, you are buying a dialler with better speech.
Where does the call end up?
A transcript in a separate dashboard is a record. A scored outcome written into the customer timeline is a system.
What happens when voice is the wrong channel?
If the product has no answer, you will find out from your conversion rate.
The teams getting value from voice AI for sales are not the ones placing the most calls. They are the ones who know which conversations deserve a call at all.
If you want to work out where voice belongs in your own funnel and where it would cost you, talk to our team.