ElevenLabs 研究：购物者试用零售 AI 智能体后更偏好语音交互

Shopping has become a conversation. Leading retailers like Walmart and Amazon have rolled out AI agents that answer questions, compare products, and help shoppers get to checkout, with other major brands in the category following suit. ElevenLabs agents are already live with hundreds of retail and e-commerce brands globally.

For retailers, the case for building one is getting stronger: Bain finds shoppers trust a retailer's on-site agent about three times more than a third-party one.

That trust is an opening, not a guarantee, and building the agent is becoming the easy part. The harder questions live in the user experience: should shoppers speak to an agent or type in a chat window? Which moments call for voice, and which for text and visuals? And at what point do shoppers want a person instead?

We tested these questions with shoppers across three studies, including two where they were given an opportunity to interact with an ElevenLabs voice agent before responding. Most research asks people what they think they'd prefer. We looked at what they chose after actually trying it, and the two answers often differed.

What we did 

Over the summer, we studied how shoppers think about AI, and how they behave once they use it, across the UK, US, Canada, Australia, Europe, South Africa, and Latin America. Most completed a real task with a live voice agent first, then answered our questions afterward.

Wherever we could, we had the same shoppers try things two ways, like fixing a wrong charge once by speaking and once by typing, so many comparisons here are one person's experiences side by side. We also looked for what held up across studies. The findings we trust most appeared in more than one, with different shoppers and different questions.

How do shoppers want to talk to an agent? 

In our first study, shoppers used a live ElevenLabs voice agent to find a product, ask a follow-up question, and add it to the cart. 

Each shopper did the task three ways: 

• speaking and hearing the reply
• speaking and reading the reply on screen
• typing and reading the reply

Afterward, we asked which they’d choose going forward.

Nearly half chose speaking and hearing. About 30% chose speaking and reading.

Asked about six specific tasks, from finding a gift to comparing two products, shoppers put speaking and hearing on top in all six. Support ranged from more than half for gift ideas to about a third for comparing products. Fewer than one in ten ruled out a shopping assistant altogether. 

About half would use one on sites they already shop, and around four in ten would use one for some kinds of shopping. In every category we asked about, from appliances to groceries to big one-off purchases, roughly four in five shoppers who wanted an assistant wanted to talk to it, not type. These are choices shoppers made after trying an agent, not purchases, so they show what people want, not what converts.

Comparing two products was the one task where voice didn't clearly win. Shoppers split almost evenly across all three options, and almost evenly again on whether they'd rather read or hear an explanation of the difference. Voice is the clear favorite for quick questions, but when shoppers are weighing options, offer text or visuals too.

Experience changes the answer 

How shoppers want to engage with an AI shopping assistant depends on whether they’ve tried one. In our earlier survey, almost no one had ever used a voice assistant to shop. We described six shopping tasks, like finding a gift or reordering a household item, and asked whether they’d type to an AI assistant, speak to it, or do the task themselves. Among those who chose an assistant, typing beat speaking by more than four to one.

In the two studies where shoppers spoke with a live agent first, speaking led. In our shopping study, it was the top choice on every task. In our support study, when shoppers chose an agent over calling a person or sorting out an order problem themselves, most chose to speak to it. 

Habit points the same way. Shoppers who already talk to voice assistants like Alexa or Siri were more likely to want a shopping assistant, to let it place an order, and to prefer hearing its answer. Typed AI tools like ChatGPT showed no such link. So if you're deciding whether to offer voice, a survey may steer you toward typing. Putting a working agent in front of customers will tell you more.

Two more findings are worth noting. When we described an agent that answers out loud and shows the same answer in the chat, an option no one had tried, just over half picked it over anything they'd used. And device made no difference: about seven in ten chose a speaking option whether they imagined shopping on a phone or a computer, and roughly two in three gave the same answer both times.

When do shoppers want a person instead?

When something goes wrong, shoppers become much less willing to deal with an agent. 

We tested this with a second group who spoke with a live agent first, then told us how they’d handle problems with an order.

The more serious the problem, the less they wanted an agent. We asked about four scenarios: a late delivery, a return, an order they never placed, and a wrong charge. 56% chose an agent for the late delivery, 49% for the return, 43% for the order they never placed, and just 34% for the wrong charge.

Then we told them the agent had failed once, unable to pull up their details and asking them to try again, and asked whether they’d give it another go or ask for a person. The share asking for a person rose along the same ladder, from 24% for the late delivery to 61% for the wrong charge.

Having a person on hand changed the choice too. We asked how they’d resolve a wrong charge, once when reaching a person meant waiting two to ten minutes, and once when a person was available immediately. With a wait, 35% chose the agent. With a person available right away, only 20% did.

Voice itself wasn’t the problem. 52% asked for a person after speaking to the agent, and 54% after typing, and how natural or robotic shoppers found the voice didn’t predict their patience on any of the four problems.

What shoppers did want was proof. When a spoken and a typed resolution both came with a written record, 72% chose to speak, so it’s the record they care about, not the typing. When we took the record away from the faster spoken option, offering a two-minute fix with no record or a five-minute typed one with proof, 58% still chose to type. Stretched to fifteen minutes, that fell to 49%.

Finally, we asked what they’d do if there were no visible way to reach a person. 60% said they’d give up on the agent entirely, even if they never planned to use that option. Nearly half of those would go looking for a phone number instead.

What this means for your agent 

Across the studies, shoppers made a strong case for a voice agent. They picked speaking over typing on every task we tested. For a late delivery, most chose the agent over waiting for a person. Serious problems, like a wrong charge, are where shoppers still want a person. That's the limit today. Agents move past it by earning trust on the easier problems first.

• Let shoppers speak to the agent and hear the reply, with products on screen. That setup was the most chosen in our shopping study. For comparisons, have the agent say the answer and show it in the chat at the same time.
• Pilot with the shoppers most likely to use voice. Start with voice-assistant users and loyalty members who engage with recommendations, then expand.
• Launch on order status and delivery questions, then add actions. Check how long customers wait for a person today and start with the contact types where the wait is longest.
• Keep wrong charges, orders the shopper didn't place, and account security with a person. More than half asked for a person after one failed attempt, so build the handoff for the first failure.
• Build a visible route to a person into every deployment from the start. Don't wait for a customer or account to ask for it.
• Send a transcript or summary after every resolution. When the spoken fix came without a record, most shoppers picked a slower typed fix that had one.

Walmart, Amazon and a growing list of retailers now have a shopping agent, and shoppers say they trust a retailer's own agent more than a third-party one. Our research shows where that trust gets tested. Quick questions are easy. What the agent does when something goes wrong decides whether shoppers hand it more, so design the handoff to a person as carefully as the conversation that comes before it.

We think the future of retail is conversational: every shopper getting the service of your best store associate, on the web, on the phone or in store, in their own language, with a person one step away when it matters. To see how it would work in your store, talk to our team.