White Label AI Voice Agents: What the Research Says

What customers say about AI on service calls

Gartner surveyed 5,728 customers in December 2023 and reported in July 2024 that 64% would prefer companies did not use AI in customer service, and 53% would consider switching to a competitor if they found out a company planned to. The biggest worry was harder access to a live person, followed by job losses and wrong answers. Gartner's own comment was that many customers fear generative AI will become another obstacle between them and an agent ¹. We read this through a trade publication's copy, because Gartner's page blocked our access.

SurveyMonkey's December 2025 study of 2,017 US adults, weighted to Census data from a non-probability online panel with a stated margin of error of 2.5 points, found that 79% strongly prefer a human over an AI agent for customer service and 8% prefer AI. Those who preferred AI cited availability (41%), speed (37%) and accuracy (30%). Those who preferred a human said it understands their needs better (61%), explains more thoroughly (53%) and is less frustrating (52%). 89% said companies should always offer the option to speak to a person, and 54% were confident they could tell when they were talking to an AI chatbot ². SurveyMonkey sells survey software and the sample is an online panel, not a random sample of the population.

Preference is not the same as behavior. Neither survey measured what people do when a call is actually answered by AI and solves their problem, which is the more relevant question for a receptionist or support line.

What experiments show about AI on the phone

The cleanest evidence we found is a randomized field experiment published in Marketing Science in 2019. More than 6,200 customers were randomly assigned to receive highly structured outbound sales calls from chatbots or human workers. Undisclosed chatbots were as effective as proficient workers and four times more effective than inexperienced workers at producing purchases. But disclosing the chatbot's identity before the conversation reduced purchase rates by more than 79.7%, and shortened calls. The authors found that people who knew they were talking to a machine were curt and perceived it as less knowledgeable and less empathetic, and that a late disclosure and prior AI experience reduced the penalty ³.

A second study, indexed in the journal Decision Support Systems, randomly assigned customers who had abandoned opening a securities account to either an AI powered outbound call or a human call. As summarized, AI raised informational support but lowered emotional support, and the empathy shortfall outweighed the information gain so the net effect on account opening intention was negative. We could read only a search summary of this study ⁴.

Both studies are about outbound sales or sales recovery calls, a narrow and structured task. The first used 2019 technology, so its numbers say little about the current generation of voice models. There is also a more recent working paper on AI voice agents in sales (Habel, Zawadzki and Freise) that is reported to find the disclosure penalty has shrunk. We could not open it and do not rely on it. We found no randomized study of AI answering inbound calls for a small business.

The timing problem: why pauses matter

Human conversation runs on fast turn taking. Stivers and colleagues studied response timing in a worldwide sample of 10 languages, from traditional indigenous communities to major world languages, and published in the Proceedings of the National Academy of Sciences on 30 June 2009. All ten languages avoided overlapping speech and minimized silence between turns, and the average gap differed between languages by no more than about 250 milliseconds from the cross-language mean ⁵.

That is the research behind claims that a voice agent must respond within a fraction of a second. The abstract we read does not state a specific millisecond target. Frequently quoted thresholds such as 800 milliseconds for a natural feeling response come from voice AI vendor blogs and engineering rules of thumb, not from this paper, and we did not find a peer reviewed study that sets them. Treat them as design heuristics.

The rules for AI voices on calls

On 8 February 2024 the US Federal Communications Commission released a declaratory ruling (FCC 24-17, adopted 2 February 2024) confirming that the Telephone Consumer Protection Act's restrictions on an artificial or prerecorded voice cover current AI technologies that resemble human voices or generate call content using a prerecorded voice. Callers must therefore obtain the prior express consent of the called party before making a call that uses an AI generated voice, absent an emergency purpose or exemption. If the call is advertising or telemarketing, the Commission's rules require prior express written consent ⁶.

The ruling also says the TCPA does not allow any carve out for technologies that purport to provide the equivalent of a live agent. Artificial or prerecorded voice messages must state clearly at the start the identity of the business responsible for the call, and a telemarketing or advertising call must offer opt-out methods ⁶. The Chairwoman's statement noted that 26 State Attorneys General wrote in support, and a footnote in the ruling cites a 2023 Ninth Circuit decision that described an artificial voice as including a sound resembling a human voice originated by artificial intelligence ⁶.

The ruling is about calls a business initiates. Our reading, not legal advice, is that it does not directly address a customer calling in to a business, which is the receptionist case. Rules also differ by state and change, and the ruling may have been amended or challenged since 2024, so check with a lawyer for your own use.

The labor context

The US Bureau of Labor Statistics lists the May 2025 median pay for customer service representatives at $44,770 a year, or $21.53 an hour. There were 2,666,000 jobs in 2025 and the bureau projects employment to fall 5%, or about 141,800 jobs, between 2025 and 2035, with about 289,500 openings a year all from replacement needs. It attributes the lower demand to tasks continuing to be automated, naming self service systems, social media and mobile applications, especially in retail trade. The page does not name AI ⁷.

How big is the market? Forecasts disagree

Market forecasts for conversational AI are predictions, not measurements, and they vary with how the market is defined. MarketsandMarkets forecast the conversational AI market would grow from $13.2 billion in 2024 to $49.9 billion by 2030, a 24.9% compound annual growth rate. Grand View Research forecast $41.39 billion in 2030 at a 23.7% rate ^8,9. Neither is a forecast for voice agents alone. We read only search summaries of these forecasts, and both publishers sell the reports.

What AI changes: the evidence

The strongest independent evidence is about text, not voice. A National Bureau of Economic Research study of 5,179 customer support agents found that agents using a generative AI assistant resolved 14% more issues per hour on average, and 34% more if they were novice or low skilled, with minimal effect for experienced workers. The assistant supported human agents ¹⁰.

We found no independent study of how often an AI voice agent resolves an inbound call without a person, how often callers hang up on it, or what it does to bookings compared with voicemail. Vendor figures for those outcomes exist, but none we found published its method.

What the evidence does not show

• The consumer surveys measure stated preference, not behavior on a real call ^1,2. The SurveyMonkey sample is a non-probability online panel, and we read Gartner through a trade publication's copy ^1,2.
• The only randomized evidence we could read, Luo et al., is about outbound sales calls with 2019 technology and a narrow script. Its disclosure penalty may not apply to inbound service calls or to current voice models ³.
• The Decision Support Systems study and the market forecasts were read through search summaries only ^4,8,9.
• A recent working paper on AI voice agents and disclosure in sales was not opened and is not used.
• The turn taking paper is a 2009 study of human conversation. It does not set a latency target for software, and the millisecond thresholds quoted by vendors are not in it ⁵.
• The FCC ruling covers consent for calls a business initiates. We offer a reading of what it means for inbound calls, not legal advice, and its status may have changed since 2024 ⁶.
• We found no independent study of AI voice agents answering inbound small business calls, and no independent measure of their containment, bookings or caller hang up rates.

Applying this as an agency

The evidence suggests four design choices before putting an AI voice agent on a client's line. First, start with inbound calls, such as after hours and overflow, where the alternative is voicemail and the consent rules in the FCC ruling are about outbound calls ⁶. Second, offer a path to a person on every call, since 89% of respondents say companies should always provide one and loss of access to a human is customers' top worry ^1,2. Third, decide how the agent identifies itself rather than leaving it to chance: the only randomized evidence shows identity disclosure can change behavior on sales calls, though it comes from outbound sales in 2019 ³. Fourth, test response speed on real calls, since conversation depends on short pauses and the thresholds are rules of thumb ⁵.

On Convocore an agency can run voice agents for each client in a separate sub-account on the agency's own domain and branding, with transfer rules, transcripts and call summaries. For outbound campaigns, take advice on consent and disclosure rules for each country and state first. We found no independent measure of how often an AI voice agent resolves a call without a person, so measure that on a pilot before promising it.

Method and limits

• Every number on this page comes from a source we opened and read on 9 October 2026, except where a source note below says we could only read a search summary.
• Sources are labelled primary (read at the original publisher), commissioned (paid for by a party with a commercial interest in the result), secondary (a report of the original) or forecast.
• Where sources disagreed or a figure could not be traced to an original, we left it out. Sample sizes and dates are given wherever the source gave them.
• Convocore sells white label AI voice agents and has a commercial interest in this market. We have tried to report the evidence as it is, including where it is unfavorable to our product. If you find an error, write to support@convocore.ai and we will correct it.

Sources

1. Gartner Survey Finds 64% of Customers Would Prefer That Companies Didn't Use AI For Customer ServiceSecondary report
   Gartner, as reproduced by Insurance Canada. Survey December 2023; press release 9 July 2024.
   We read the trade publication's copy of Gartner's release because Gartner's own page returned an error. 5,728 customers. Gartner's survey was not specific to phone calls.
2. Customer service statistics: humans vs AI (2025 CX study)Primary source
   SurveyMonkey. Fieldwork 10 to 11 December 2025.
   2,017 US adults aged 18 and over from a non-probability online panel, weighted to Census Bureau data, margin of error 2.5 points. SurveyMonkey sells survey software. The page also repeats figures from other studies, which we did not use.
3. Frontiers: Machines vs. Humans: The Impact of Artificial Intelligence Chatbot Disclosure on Customer Purchases (Luo, Tong, Fang, Qu)Primary source
   Marketing Science, vol. 38, no. 6. Published online 20 September 2019.
   We read the abstract and introduction of the full article. Randomized field experiment, more than 6,200 customers, highly structured outbound sales calls. 2019 technology.
4. Emotion vs. information: Understanding the effect of AI-powered call systems on potential customer decision from a field experimentSecondary report
   Decision Support Systems. 2025.
   ScienceDirect blocked our access, so we read only a search summary of the abstract. We could not confirm sample size or effect sizes.
5. Universals and cultural variation in turn-taking in conversation (Stivers, Enfield, Brown and others)Primary source
   Proceedings of the National Academy of Sciences, 106(26), 10587 to 10592. 30 June 2009.
   PNAS blocked our access, so we read the full abstract on a university research record (Illinois Experts). We did not read the body of the paper.
6. Declaratory Ruling, Implications of Artificial Intelligence Technologies on Protecting Consumers from Unwanted Robocalls and Robotexts, CG Docket No. 23-362, FCC 24-17Primary source
   US Federal Communications Commission. Adopted 2 February 2024, released 8 February 2024.
   We read the ruling and the Chairwoman's statement. It does not discuss inbound calls. We did not check later amendments or court challenges. This is not legal advice.
7. Customer Service Representatives, Occupational Outlook HandbookPrimary source
   US Bureau of Labor Statistics. May 2025 pay data, 2025 to 2035 projections; read 9 October 2026.
   The handbook is updated yearly. It cites automation of tasks and self service systems as reasons for lower demand and does not mention AI by name.
8. Conversational AI Market worth $49.9 billion by 2030, growing at a CAGR of 24.9%Forecast
   MarketsandMarkets, via GlobeNewswire. 6 May 2024.
   A forecast from a firm that sells the report. The page blocked our access, so we read a search summary. Scope is conversational AI, not voice agents alone.
9. Conversational AI Market to Hit $41.39 Billion By 2030Forecast
   Speech Technology magazine, reporting Grand View Research. 2025.
   A forecast reported by a trade magazine. We read a search summary only. Search summaries of other outlets gave 23.6% for the growth rate, so the rate varies by report.
10. Generative AI at Work (Brynjolfsson, Li, Raymond), NBER Working Paper 31161Primary source
    National Bureau of Economic Research. April 2023, revised November 2023.
    We read the paper's abstract page. 5,179 customer support agents; a working paper rather than a final journal article, about text based support, not voice.

Questions

Related