Voice is where AI customer service gets personal. A slow reply, a missed word or a confident wrong answer lands on the caller straight away.
What callers say
80% of consumers say they are willing to use AI customer service (Five9 and Hanover Research, April 2026, 3,000 consumers in the US, UK and Germany).
87% say a route to a human is essential when a company uses generative AI for service (Gartner, February to March 2026, 3,566 customers).
71% say knowing it is an AI is very or extremely important (Five9).
83% at least sometimes have to repeat themselves after an AI-to-human transfer (Five9).
So callers will use it, but they want an honest label and an easy way out.
How fast is "fast"?
Vendors often quote replies of half a second or less. An independent test of five voice platforms on real phone calls (Openbenchmarks, 2026) found median replies of 1.3 to 1.7 seconds, measured from the caller's last word to the agent's first audio. The difference is mostly in what is measured: server-side figures leave out phone-line and end-of-speech delay.
Real callers are harder than demos
A 2026 academic benchmark of 12 voice agents found none scored above 0.5 on accuracy and experience at once, and that accents and noise cut scores by up to 0.314. The best causal evidence for gains comes from helping people: AI assistance raised issues resolved per hour by about 14%, and by about 34% for the least experienced agents (5,179 agents, a chat setting rather than voice).
Seven questions to ask any supplier
How do you define containment and resolution, and do hang-ups or caller-requested transfers count as contained?
What are the median and 95th-percentile response times on a real telephone call?
Can I test on 50 of my own recorded calls, including strong accents and poor lines?
What happens when the agent is unsure or the caller asks for a person, and does context transfer?
Which answers come from my approved content, and how do you stop invented answers?
What does one fully completed call cost, all in?
Where is call audio stored, for how long, and is it used to train models?
Starting small
Pick one narrow job: after-hours calls, bookings, order status or overflow when staff are busy. Tell callers it is an AI, keep a clear route to a person, test with your own callers' accents and phone lines, and judge success by calls resolved, not calls handled.
Read the full report
Our industry report Voice AI in the contact centre: what is real, what is hype has the charts, the myths and every source.
View the full report on our Resources page →
This article summarises published research as of 23 October 2026 and is for general information only. Gartner figures are taken from syndicated copies of its releases, and some data comes from companies that sell AI products.
