ElevenLabs Shows How Businesses Deploy Voice Agents: KPN Goes From 50% to 90% Accuracy

A vendor's corporate blog is a genre that should be read with a marketing discount. But ElevenLabs' latest roundup of case studies contains several details useful to anyone planning to add voice to a product. For context: the company says enterprise customers account for more than 55% of its revenue, and employees at almost 70% of Fortune 500 companies use the platform. Now, what exactly are customers building?

Five Scenarios, Five Different Roles for Voice

Yahoo added a voice mode to Yahoo Scout, its question-answering service. It is built on ElevenAgents, so the team did not have to write voice infrastructure itself. Scout exited beta after more than 3 billion answers, but that figure applies to the whole service, not to voice. The contribution of the voice mode cannot be separated out from the text.

Westgate Resorts built a voice agent, Sarah, for routine front-desk questions. The agent's knowledge was drawn from 8,000 real call transcripts so that it answers what guests actually ask about. It was launched at a 2,100-room hotel in Orlando, and two more resorts are now being added. The text includes no metrics such as the share of inquiries resolved.

Twilio embedded ElevenLabs voices into Conversation Relay, its tool for voice AI conversations. For developers on Twilio, this means speech synthesis is plugged in within a familiar platform. Doppel uses the API for vishing simulations: security teams call employees with a realistic AI voice so they hear in advance what an attack sounds like. Pocket, a wearable AI voice recorder, has transcribed more than 2.5 million hours of audio with ElevenLabs Speech-to-Text.

The KPN Case: Where Models Actually Break

The most substantive example is the Dutch operator KPN. For its Mobiel AI product (voicemail transcription, call summaries, smart routing), what matters is not the general "readability" of a transcript. What matters are the small details: a postal code and a date of birth. Dutch, with its long compound words and strong regional accents, is hard for models, and an error in a single digit of the postal code breaks the entire identity check.

ElevenLabs engineers, together with KPN, fine-tuned a model on Dutch data. Identity verification accuracy rose from roughly 50% to 90%. The company adds that all Dutch deployments on the platform benefited from the improvement. It is an honest lesson: average recognition accuracy hides failures on critical entities. If your agent needs to record an address, a contract number or a date, those are exactly what you need to measure.

What to Do With This

First: if voice is not your core competence, a ready-made agent platform saves months. Yahoo says outright that it could not have shipped the product that quickly on its own. The price is vendor dependence, and the article says nothing about cost or latency.

Second: data matters more than magic. Westgate used its own calls, KPN used its customers' language. Set up labeling of real dialogues before choosing a model: you will need them for both training and testing.

Third: measure accuracy on key fields, not on the whole text. A test on postal codes, names and dates will show what an overall metric masks.

The final caveat concerns the material itself. This is a vendor showcase: there is no data on project timelines, costs or the share of inquiries the agent closes without a human. The "50% to 90%" figure comes with no measurement methodology or sample size. Treat it as an order-of-magnitude guide, not a guarantee.

If you are building a voice agent for customer support, take from the article not the logos but the scheme: a ready-made platform, your own transcripts and a separate measurement on critical data. Most of the success stories described rest on these three steps.