ElevenLabs Voice AI Handles 90% of Rohlik's Calls and Lifts Lead Qualification 3.3x
Customers change dates and budgets mid-search, and a filter form forces them to start over. ElevenLabs showed a voice agent that searches a live database and completes the order in a single conversation. Maia, the agent at online supermarket Rohlik, handles 90% of hotline calls, and at the Immobiliare.it portal, lead qualification (sorting out inquiries from real buyers) grew 3.3x. The companies publish these figures themselves, so verify them in your own pilot.
AI-processed from ElevenLabs Blog; edited by Hamidun News
You're looking for a hotel: first two nights by the canals, then three, and then it turns out you already have another trip on those dates. In a filter form, any such change means "start over," and according to ElevenLabs, people abandon about 70% of online carts, payments and bookings midway. The company proposes replacing the form with a voice AI agent that searches for options, answers questions and places the order in a single conversation.
In a recent webinar, it presented two main examples. Maia, the agent at European online supermarket Rohlik, handles 90% of incoming hotline calls. And at Italian real-estate portal Immobiliare.
it, lead qualification grew 3.3x: that is, 3.3 times as many inquiries can be identified as genuinely interested buyers worth passing to sales managers.
For a product manager or developer, this is a reason to look at where in the product customers struggle with forms and queues.
What the demo and the customer cases showed
The demo is simple. On a live booking site, the agent finds a hotel in Amsterdam for two people: up to 400 euros a night, near the canals and a museum. Midway through the conversation, the customer extends the trip from two nights to three. The agent searches a live hotel database, and the site page updates during the conversation itself. It notices an overlap with a trip to Istanbul the customer has already bought and asks before confirming. It then passes an anniversary request on to the hotel, cancels an extra booking at the customer's request and briefly switches to Dutch to recap the result.
The customer cases carry more weight:
- Rohlik (3 million customers in five countries). Maia closes 90% of hotline calls and 90% of inquiries across phone, web, app and WhatsApp in six languages. Issues are resolved more than twice as fast, and the company expects to cut operating costs by 50% without losing CSAT (customer satisfaction scores). Customers can build and pay for an order by voice.
- Klarna uses the agent as the first line of phone support for 35 million customers in the US. Some typical requests are resolved up to 10 times faster.
- Immobiliare.it: users were 74% more willing to leave a phone number. The company started with one agent for listings; it is now a network of agents for sellers, outbound calls, support and search.
- SevenRooms: more than 400 thousand calls across 600+ restaurants, a fifth of them after closing time. Restaurants got up to 25% more bookings, comparing the first three months of 2026 with the same period a year earlier.
- KPN, together with ElevenLabs, raised the accuracy of Dutch speech verification from about 50% to 90%. Deliveroo reached more than 80% of couriers stuck at registration.
Why voice instead of chat
According to ElevenLabs data from hundreds of production agents that have both voice and chat, voice is 29% more successful after two minutes of conversation. In a survey of 98 people, 85% preferred voice for rebooking a flight after a cancellation, citing speed of resolution and empathy. A 16% reduction in response latency came with a 26% increase in successful conversations. The takeaway for developers is practical: an agent that is too slow loses, even if the model is very smart.
The background matters too: according to the Marchex Institute, in the US one call in seven goes unanswered or drops, and 80% of people who reach voicemail hang up without leaving a message.
Build tips
Stripped of the marketing, the recommendations are fairly sound engineering:
- Start with one clearly defined scenario and get it into production quickly, then add returns or order lookup.
- Spell out the role, tone, prohibitions (for example, do not invent data) and what to do when a tool fails or a booking does not go through in the system prompt.
- Connect your CRM and booking system, but expose each tool and document only at the stage of the conversation where it is needed.
- Move authentication and booking confirmation into separate sub-agents and tool calls so that behavior stays predictable.
- Choose the model and voice deliberately and test them on the same scenarios: a powerful model may be too slow, and a fast one may be less stable.
- Plan the handoff to a human: user ID, authorization status, a short summary of the call and the reason for the handoff.
- Check data requirements in advance: EU data storage and a zero-retention mode are available, and outbound calls in the Netherlands require explicit consent.
Caveats
This is vendor material, and all customer results were published without independent verification. The SevenRooms figure is "up to 25%," and a survey of 98 people is too small to build a strategy on. The claim that teams building an agent on their own capture only 10–20% of the platform's potential looks like an argument for paid services from ElevenLabs and its partners. The cost per minute of conversation and the price of integration are not given, so you will have to work out the economics in your own pilot. Budget days, not hours, for tuning, and watch what callers ask that the agent cannot answer. In the demo, for example, it was parking.
- Nov 18, 2026AWS releases SMART: spin up a WhatsApp agent on AgentCore with one command
- Nov 18, 2026Anthropic: Claude with 413 note files earned 1 badge in Pokémon, but with a handful of files it earned 4
- Nov 18, 2026Configuring Qwen Code: QWEN.md, Subagents, and a Sandbox Cut Wasted Tokens
- Nov 18, 2026Grok Bot accessed email and GitHub 31 times in 5 days without being asked, then denied it
Key Questions Answered
What the demo and the customer cases showed?
The demo is simple. On a live booking site, the agent finds a hotel in Amsterdam for two people: up to 400 euros a night, near the canals and a museum. Midway through the conversation, the customer extends the trip from two nights to three. The agent searches a live hotel database, and the site page updates during the conversation itself. It notices an overlap with a trip to Istanbul the customer has already bought and asks before confirming.
Why voice instead of chat?
According to ElevenLabs data from hundreds of production agents that have both voice and chat, voice is 29% more successful after two minutes of conversation. In a survey of 98 people, 85% preferred voice for rebooking a flight after a cancellation, citing speed of resolution and empathy. A 16% reduction in response latency came with a 26% increase in successful conversations. The takeaway for developers is practical: an agent that is too slow loses, even if the model is very smart.
What are the key details on Build tips?
Stripped of the marketing, the recommendations are fairly sound engineering: Start with one clearly defined scenario and get it into production quickly, then add returns or order lookup. Spell out the role, tone, prohibitions (for example, do not invent data) and what to do when a tool fails or a booking does not go through in the system prompt. Connect your CRM and booking system, but expose each tool and document only at the stage of the conversation where it is needed.
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.
The AI world, distilled — once a week
Seven stories that actually mattered, hand-picked. No noise, no reposts, no press releases.
Done! Check your inbox for a confirmation.