Free Intensive Nov 21–22Free 2-day hands-on Vibecoding Intensive on Nov 21–22 at 16:00 MSK (Sat–Sun) · Live Cohort 3 starts Nov 23Join Free →Cohort 3 (Nov 23)
ReleaseAPI / Product · ActionableElevenLabs Blog→ original

Shoppers Want to Talk to an AI Agent, but Only 34% Pick It for a Billing Dispute

ElevenLabs let shoppers talk to a live voice agent for a store and asked what they would choose next. Nearly half preferred to speak and listen to the answer, but only 34% picked the agent for a wrong-charge dispute, and 56% for a late delivery. That is why the voice agent should be launched on order and delivery status, while money questions stay with people. After a single agent failure, 61% want a human for a billing problem and 24% for a delivery problem. Without a visible path to a human, 60% of shoppers abandon the agent altogether. In a standard survey without a live trial, print beat voice by more than 4 to 1.

AI-processed from ElevenLabs Blog; edited by Hamidun News
Shoppers Want to Talk to an AI Agent, but Only 34% Pick It for a Billing Dispute
Source: ElevenLabs Blog. Collage: Hamidun News.
Listen to article
Article Fact Sheet & Key Figures
ParameterDetails
Primary sourceElevenLabs Blog ↗
Editor-in-Chief
Key metric #1Let the shopper speak and listen while showing products on screen. For comparisons, duplicate the answer in the chat.
Key metric #2Run a pilot with users of voice assistants and loyalty-program members, then widen the audience.
Key metric #3Start with order and delivery status, then add actions. Take first the requests where the wait for a human is currently longest.

Imagine you have launched an AI agent in your online store. Shoppers are willing to ask it about a late delivery. But what if their card was charged the wrong amount? ElevenLabs, a developer of voice agents, ran three studies of shoppers and found that after a live trial, almost half choose to communicate by voice. But for a dispute over a wrong charge, only 34% of respondents picked the agent. For a late delivery, 56% picked it. This shows where the line of trust in an agent currently sits.

Here is the context. Walmart, Amazon and other retailers have already launched agents that answer questions, compare products and carry shoppers through to payment. According to Bain, which ElevenLabs cites, shoppers trust an agent on a store's own website roughly three times more than an agent from a third-party company. Building an agent is becoming easier, but deciding how it should talk and when it should call a human is harder.

A caveat: ElevenLabs sells a voice agent platform, so its conclusions benefit the company and should be verified with your own audience. But the methodology is interesting. Participants first completed a real task with a live voice agent and only then answered questions. Many comparisons were built on the same people: fixing a wrong charge once by voice and once by text. The survey ran over the summer in the United Kingdom, the United States, Canada, Australia, Europe, South Africa and Latin America. The authors consider only results that repeated across several studies to be reliable.

Voice wins, but not everywhere

In the first study, shoppers used an agent to look for a product, asked a clarifying question and put the item in their cart. Each participant completed the task in three ways: speaking and listening to the answer, speaking and reading the answer on screen, and typing and reading. Almost half chose "speak and listen", and about 30% chose "speak and read". Across six tasks, from finding a gift to comparing two products, voice led in all of them. Fewer than one in ten rejected the AI assistant entirely. About four in five people who wanted to use an assistant wanted to talk to it rather than type, in every category: electronics, groceries and large purchases.

The one exception was comparing two products. Voice preferences split roughly evenly among the three options. The practical takeaway: answer quick questions by voice, but when a person is weighing options, also show text or images. Nobody tested a format where the agent speaks and simultaneously writes the same content in a chat, but slightly more than half would choose exactly that. The device made no difference: about 70% chose voice for both phones and computers.

There is also an experience effect. In an earlier standard survey, almost no one bought by voice, and among those who chose an assistant, typing beat voice more than 4 to 1. After a live trial, the picture flipped. If you are deciding whether to build voice, a questionnaire may push you toward text, but a working prototype will tell you more. Users of Alexa and Siri were more willing to want a shopping assistant, to let it place orders and to prefer listening to the answer. For users of text tools such as ChatGPT, no such link was found.

When the shopper calls a human

The more serious the problem, the less often people chose the agent. For a late delivery, 56% chose it; for a return, 49%; for an order the person did not place, 43%; and for a wrong charge, only 34%. Participants were then told the agent had failed once: it could not find their data and asked them to repeat themselves. The share of people demanding a human rose from 24% for a delivery problem to 61% for a billing problem.

Having a live employee available also matters. If waiting for a human takes two to ten minutes, 35% chose the agent for a billing dispute. If a human is available immediately, only 20% did. The voice itself was not the problem: after a conversation with the agent, 52% demanded a human, and after a text exchange, 54% did. How natural or robotic the voice sounded made no difference to patience.

Shoppers wanted proof. When both the voice and text solutions came with a written record, 72% chose voice. If a quick two-minute voice fix came with no record, but a five-minute text fix did, 58% still chose text. When the text option was stretched to fifteen minutes, the share fell to 49%. And if there is no visible way to reach a human, 60% will abandon the agent altogether, and nearly half of them will look for a phone number.

What developers and product managers should do

  • Let the shopper speak and listen while showing products on screen. For comparisons, duplicate the answer in the chat.
  • Run a pilot with users of voice assistants and loyalty-program members, then widen the audience.
  • Start with order and delivery status, then add actions. Take first the requests where the wait for a human is currently longest.
  • Keep wrong charges, other people's orders and account security with people, and design the handoff to a human around the first failure.
  • From day one, make the path to a human visible and send a transcript or summary after every resolution.

Limitation of the research: participants stated their choices after a trial, rather than actually buying. This reveals what people want, but not conversion. Trust in an agent is tested at the moment of error, so the handoff to a human deserves as much careful design as the conversation itself.

Story

Key Questions Answered

What are the key details on Voice wins, but not everywhere?

In the first study, shoppers used an agent to look for a product, asked a clarifying question and put the item in their cart. Each participant completed the task in three ways: speaking and listening to the answer, speaking and reading the answer on screen, and typing and reading. Almost half chose "speak and listen", and about 30% chose "speak and read". Across six tasks, from finding a gift to comparing two products, voice led in all of them. Fewer than one in ten rejected the AI assistant entirely.

When the shopper calls a human?

The more serious the problem, the less often people chose the agent. For a late delivery, 56% chose it; for a return, 49%; for an order the person did not place, 43%; and for a wrong charge, only 34%. Participants were then told the agent had failed once: it could not find their data and asked them to repeat themselves. The share of people demanding a human rose from 24% for a delivery problem to 61% for a billing problem. Having a live employee available also matters.

What developers and product managers should do?

Let the shopper speak and listen while showing products on screen. For comparisons, duplicate the answer in the chat. Run a pilot with users of voice assistants and loyalty-program members, then widen the audience. Start with order and delivery status, then add actions. Take first the requests where the wait for a human is currently longest. Keep wrong charges, other people's orders and account security with people, and design the handoff to a human around the first failure.

ZK
Zhemal Khamidun — CPO at Alpina Digital, product manager of AlpinaGPT. Building AI products since 2019. Hamidun News is a pet project that grew into a production aggregator with 6,000+ readers.
✦ Private AI Builders Community · Vibe Club

Ship AI Faster with Founders, Engineers & Weekly Architecture Reviews

Inside Vibe Club (hamidun.com/club) — weekly live masterminds with Zhemal Khamidun, Tech Radar, full LMS access, and 370+ production Claude Code skills.

Join Vibe Club →Explore Hamidun Academy
What do you think?
Loading comments…