KPN launches Mobiel AI: voice assistant for small businesses boosts accuracy from 50% to 90%

For a plumber, hairdresser, or phone repair technician, this is the business itself. A missed call means a lost order. But answering every call means stopping work. According to KPN, 53% of entrepreneurs acknowledge that being reachable by phone hurts employee focus, while 69% identify customer satisfaction as one of their top priorities. Mobiel AI is designed to address precisely this trade-off.

What customers get

KPN, the Netherlands' largest operator, built Mobiel AI together with ElevenLabs. It is a paid add-on for KPN EEN MKB business plans and works directly inside the operator's existing mobile app. Customers do not need to change their number or plan.

The package includes three capabilities. First, voicemail transcription: users can read a missed call within seconds and decide what is urgent. Second, a summary and full transcript of every call, with follow-up actions captured automatically. Third, a voice menu that routes callers to the right colleague or department, plus a status setting that determines when the entrepreneur is available.

Put simply, it is a secretary that never gets tired and never forgets anything, but lives in the operator's app rather than in a separate service.

Why Dutch became the challenge

KPN compared several voice AI providers and chose ElevenLabs for its speech quality and conversational AI. For Dutch, that is no small consideration: the language is rich in compound words and regional accents. A general-purpose model struggles with this material.

The key figure in the case study concerns verification of short, high-stakes phrases: after collecting additional Dutch data and fine-tuning the transcription model for short, high-confidence sequences, accuracy rose from roughly 50% to 90%. One important detail is that the improvement benefited every Dutch-language deployment on the ElevenLabs platform, not just KPN. For developers, this is a useful example: a general speech-recognition model for smaller languages almost always needs targeted refinement for short phrases such as names, numbers, and confirmations, where errors are costly.

Data stays in the country

The second argument is data sovereignty. Mobiel AI runs on KPN's network and in its own data centers. Voice data remains in the Netherlands: ElevenLabs processes speech in real time and returns the result to the operator. According to the companies, customer data is not used to train models and is not shared with third parties.

The teams assembled the first working voicemail prototype during a design sprint that brought KPN developers and small-business owners into the same room. This format is worth adopting: the people for whom the product is being built see the prototype from day one.

What to do with this

The limitations should be kept in mind. This is a customer case study from the vendor itself, so the figures have not been independently verified. The method used to calculate the 50% and 90% accuracy figures is not disclosed: it is unclear which dataset or metric was used. Pricing, volumes, and retention metrics are also absent.

Practical takeaways for teams building voice products:

• Embed AI where customers already have trust and an established channel: the operator's app instead of a new service.
• For smaller languages, plan to fine-tune on short, critical phrases rather than focusing only on long-form speech.
• If you work with European customers, local voice processing becomes a selling point, not merely a compliance requirement.
• Start with a narrow pain point: voicemail and call summaries are simpler than a full-fledged voice agent and deliver a faster return on investment.

KPN's story shows that the value of voice AI for small businesses often lies in unglamorous features: reading what was missed, remembering to call back, and avoiding distractions from spam. These are exactly the kinds of features that sell without lengthy explanations.