Introducing Eleven v4Meet Eleven v4, our most emotive model yet.
DiscoverSkip to contentLog inSign upContact salesLog inCustomers don’t buy in straight lines. They change dates and budgets, have questions halfway and expect the answer to come with the next step. Forms and filters make them start over. A voice AI assistant does not. It searches, answers, adapts and completes the transaction in a single conversation.
In our recent webinar, Pieter Peeters (GTM), Boning Gong (FDE) and Shana Inch (GTM) showed how an AI agent built on ElevenAgents handles real hotel booking conversations, what separates agents that perform from agents that disappoint and how teams get from pilot to production.
Key takeaways
- One conversation replaces forms and filters. Customers change dates, budgets and plans mid-request. The AI assistant searches live data, adapts, and completes the booking or order. At Rohlik, customers can already search products, build a shopping list and check out by voice.
- A good assistant guides, not only answers. It knows the customer and the catalogue. In our demo, the assistant caught a clash with an earlier booking that the customer had forgotten. At Immobiliare.it, a listing assistant lifted lead qualification 3.3x.
- Support follows the same pattern. Customers with a problem want it solved, not routed. Rohlik’s assistant, Maia, handles 90% of incoming hotline calls, and Klarna’s assistant is first-line phone support for all 35 million US customers.
- Voice earns its place as conversations get harder. Past two minutes, voice succeeds 29% more often than chat. In our study, 85% of 98 people preferred voice over chat to rebook a cancelled flight. Speed matters too: a 16% drop in latency came with a 26% rise in success.
- Configuration decides the outcome. Many AI agents that disappoint are set up wrong: the wrong model, the wrong voice or too little time spent on configuration. Technology is not the problem. Building with a forward deployed engineer or a trained partner gets far more out of the project.
Why customers want an assistant, not a form
Customers don't book or buy in straight lines. A request starts as two nights near the canals, becomes three, then runs into a trip already on the calendar. In a shop, a basket starts with a recipe, grows with the usual order and ends at checkout. Forms and filters are too static for that. Each change forces the customer to start over or leave. Each change forces the customer to start over or leave, which is part of why 70% of online carts, checkouts and bookings are abandoned.
Choosing is the hard part. Few customers know which filter to set. They know what they want: a quiet hotel near the museums under €400 a night, a recipe for four, a home close to the school. A good assistant asks the next question, narrows the options and recommends one. It also knows the customer.
Support follows the same pattern. A customer with a problem does not want a phone menu or a queue. They want it solved, in their own language, and handed to a person with the full context when the assistant cannot do it. Demand also ignores office hours. At SevenRooms, 1 in 5 of the calls handled by its Voice AI arrived after hours. Voicemail doesn't catch them: 80% of callers who reach it hang up without leaving a message. US data from the Marchex Institute shows that 1 in 7 phone calls fails, unanswered or abandoned.
The cost of unanswered calls
Stress pushes people toward voice. In an ElevenLabs study of 98 people, 85% preferred voice over chat to rebook a cancelled flight, and named speed of resolution and empathy among their top reasons. Chat works for short questions, but across hundreds of production agents running both channels, voice succeeds 29% more often once a conversation runs past two minutes. Many teams automate chat or email first and add voice last. In our experience, starting with voice lifts metrics right away.
Voice excels in complex customer conversations
What AI retail assistants deliver in production
AI assistants pay off first where customers need help choosing, ordering or fixing something: a shop with a long catalogue, a booking site with too many filters, a support queue that keeps growing. ElevenAgents is live across hospitality, travel, retail, commerce and marketplaces, and customers have published results in five areas.
Capturing demand outside opening hours. SevenRooms Voice AI, powered by ElevenAgents, has handled more than 400,000 calls across 600+ restaurants, 1 in 5 of them after hours. Restaurants captured up to 25% more reservations, based on SevenRooms data comparing venues' first three months on Voice AI in 2026 with the same period a year earlier.
Resolving support faster, in the customer's channel and language. Rohlik Group, one of Europe's largest online grocery platforms, serves 3 million customers across five countries. Its agent, Maia, handles 90% of incoming hotline calls and 90% of customer communications across phone, web, app and WhatsApp, in six languages. Resolution is over 2x faster and available 24/7, and Rohlik is on track to cut operational costs by 50% while maintaining CSAT. Customers can also search products, build a list and check out by voice. "We believe voice is the future of grocery," says Lukas Honke, Head of AI Engineering at Rohlik Group. Klarna runs its agent as first-line phone support for all 35 million US customers, with resolutions up to 10x faster for the queries it handles. eDreams ODIGEO reports double-digit improvements in resolution speed and in transfer rates across five languages, and LOT Polish Airlines is deploying ElevenAgents for inbound phone support, including booking changes and ticket inquiries.
Getting local languages right. KPN built Mobiel AI, an assistant inside its business mobile plans. Joint work took Dutch verification accuracy from roughly 50% to 90%, and that improvement now benefits every Dutch-language deployment on the platform.
Qualifying and converting leads. Immobiliare.it, Italy's leading property marketplace, saw lead qualification rise 3.3x. Users were 74% more likely to share their phone number, and about 80% rated the feature positively. What started as a single listing agent has grown into a network of agents covering sellers, outbound, support and search.
Running outbound at scale. Deliveroo's outbound agents reached more than 80% of riders who had stalled in onboarding, verified opening hours for 75% of restaurants and reached 86% of partner sites for rider tag activation.
Demo: AI concierge for hotel booking
Scenario: A travel concierge agent on a live booking site helps a customer find a hotel in Amsterdam for two people, under €400 a night, within walking distance of the canals and the museum. The customer changes the request mid-conversation.
What was shown:
- Natural, fast conversation that absorbs changes of plan, such as extending a stay from two nights to three
- Live search across a hotel database, with the website updating as the agent works
- Use of previous bookings: the agent caught an overlapping stay in Istanbul and asked before confirming
- Personalization: the agent asked about the customer's anniversary and passed a note to the hotel
- Follow-through: the booking was confirmed, and the conflicting one cancelled on request
- A switch to Dutch for a summary, then back to English
Why it matters: The customer finishes in one conversation what would otherwise take pages of filters and forms. The agent also catches the clash the customer had forgotten, and completes the task end to end.
Best practices for building a booking or commerce agent
- Start with one well-defined journey. Get a small use case into production quickly to confirm it works and is fast enough. Then extend the same platform to order lookups, refunds or shopping assistance.
- Write down persona, goals and guardrails. A strong system prompt covers who the agent is, its tone, what it must never do (such as invent information), and what to do when a tool fails or a reservation doesn't go through.
- Connect the right data, and scope it. Link knowledge and tools to the systems that hold the answers, such as your CRM and booking system. Make each tool and document visible only at the stage of the conversation where it's needed, so performance holds as the agent grows.
- Use workflows where the path must be fixed. Route steps like authentication or confirming a reservation through dedicated sub-agents and tool calls, so behavior is predictable.
- Choose the model and voice deliberately. A very capable model can make the agent slow enough to feel broken, while the fastest models can be less consistent. A voice that doesn't fit the brand can make a good agent sound poor. Test options against the same scenarios.
- Test before launch, measure after. Define success criteria up front, review every call against them, and track what callers ask that the agent can't answer. In the demo data, for example, callers kept asking about parking.
- Plan the handover to a human. Pass collected details, such as user ID and authentication status, plus a summary of the call and why the agent handed over, so the person picks up where the agent stopped.
- Check data and consent requirements early. EU data residency and a zero retention mode are available. For outbound calls triggered by website activity, rules such as explicit consent in the Netherlands apply.
- Give configuration time. Many teams test for a few hours and conclude the technology isn't ready. Results come from learning how to configure the agent, including model, voice and prompt.
Avoiding common AI agent implementation pitfalls
Three ways to get started
- Build it yourself. Buy a license on the website, or contact us for an enterprise license, then build with the documentation and the Architect, an AI builder that configures an agent from a description of what you want. A first working version can take about a day, testing included, which makes it the quickest way to validate a small use case.
- Build with ElevenLabs. Work with a forward deployed engineer who builds alongside your team and brings what they've learned across customer projects.
- Work with a systems integrator. Our trained partners in the Benelux run hands-on workshops on site. This suits smaller first use cases, or teams that need to justify a larger pilot to their board.
In our experience, teams that build entirely on their own capture only 10-20% of the platform's potential, so most projects, especially large scale ones, benefit from one of the last two options.
Watch the full session
Watch the full webinar here, including the audience Q&A on integrations, human handover and data residency. Ready to see this on your own bookings, orders or reservations? Contact our sales team for a guided demo tailored to your team.
Written by

GTM
View profilePieter Peeters is on the GTM team and focuses on establishing ElevenLabs' presence in voice AI and agents in Europe.
Similar articles

Webinar Recap: Build Safe AI Agents for an Enterprise Deployment
CategoryProductDateMay 9, 2026
3Shape expands global dental support with AI voice and chat agents from ElevenLabs
CategoryCustomer StoriesDateFeb 26, 2026
Everlywell expands access to health screenings with personalized, AI-powered outreach
CategoryCustomer StoriesDateFeb 3, 2026
Eldra delivers trusted phone support for elderly home care using ElevenAgents
CategoryCustomer StoriesDateFeb 10, 2026
Create with the highest quality AI Audio
Talk to salesSign up
AI 热点