Voice AI in 2026: complete guide for Dutch companies

Voice AI is the breakthrough of 2025-2026

Call NS customer service at half past nine in the evening and for a few years now you've been getting a voice that tries to be friendly. Until you ask whether you can catch your train tomorrow with a bike instead of a dog. Then you get transferred to a human. That kind of moment is the real test case for voice AI: not the demo where the voice sounds natural, but the production where the caller asks something outside the script. In 2025-2026 that test case is, for the first time, workable for a specific set of use cases. The underlying technologies are now mature. For SMB organisations that opens up reception outside office hours, intake at high volumes, first-line customer service, appointment confirmations, and in some sectors qualifying customer needs for sales.

This article covers the tech, tools, concrete use cases, telephony integration, costs, AI Act implications, and the situations where you're better off not using voice AI.

What is voice AI, really?

Voice AI is software that holds a spoken conversation. Under the hood, voice agents consist of three layers that either work separately (classic pipeline) or as one integrated model (next-gen voice models like OpenAI Realtime or Gemini Live).

1. ASR (automatic speech recognition). Speech to text. Tools: OpenAI Whisper, AssemblyAI, Deepgram. For clear Dutch at phone quality, the error rate in 2025-2026 sits below 5%. Whisper-Large-v3 does a solid job on Dutch dictation. NVIDIA's Parakeet is faster in benchmarks, but with substantially more errors on Dutch. Test on your own audio before you pick.

2. LLM (large language model). Understand the text input, plan a response and phrase it. Tools: GPT-4o, Claude, Gemini. This is the brain of the agent. It understands the question, optionally consults the knowledge base or an API, and decides what action is needed.

3. TTS (text to speech). Text into natural-sounding speech. Tools: ElevenLabs, OpenAI Voice, PlayHT, Cartesia. Dutch voices had their breakthrough in 2025-2026. With the top tools, the difference from a human voice is only audible if you really listen for it.

In the classic pipeline those three steps follow one another: user speaks, ASR, LLM, TTS, response. Latency: 1.5 to 3 seconds, which is sometimes noticeable in a phone call.

In next-gen integrated voice models (OpenAI Realtime, Gemini Live, some Anthropic experiments), everything sits in one model: speech in, speech out. Latency drops to 300-700 ms, comparable to a natural human conversation. For most production use cases, a hybrid approach (classic for robustness, integrated where conversational flow demands speed) is the most practical.

How does it work in production?

A production voice agent has six parts. First, telephony integration: the agent has to hang off a number. For the Netherlands that runs via SIP trunking (Twilio, Telnyx, Vonage), via cloud PBX integration (RingCentral, Aircall, 3CX), or via WebRTC for browser calls. Then a voice platform or orchestrator that glues telephony, ASR, LLM and TTS together and manages the conversational flow: Vapi, Retell AI, Bland.ai or ElevenLabs Conversational AI. For those who want to go deeper technically: build your own on LiveKit Agents or Pipecat (open source).

Underneath sits the knowledge base or RAG: a searchable vector database (Pinecone, Weaviate, Postgres+pgvector) filled with your SOPs, FAQs, product info or website content. The agent reads from here. Alongside sit the tool integrations: API calls the agent makes during a conversation, like booking an appointment in the calendar, pulling customer records from the CRM, opening a ticket, looking up an invoice.

Two things you forget in every pilot and desperately need in production. One: human escalation. When the agent hesitates or falls outside its scope, it needs to hand off smoothly to a human, with context transfer included. Not a luxury, a must. Two: monitoring and transcripts. Every call gets recorded, transcribed, scored, and flagged automatically for human review on issues. Without this, you can't guard quality and you can't provide an audit trail.

For the broader agent context see /ai-agents.

Human escalation and monitoring: you forget them in every pilot, and you desperately need them in production.

Use cases that work in production

Concrete examples running in production at Dutch clients in 2025-2026.

Reception outside office hours delivers the clearest win. An AI voice that picks up the phone after 5pm, answers first-line questions (opening hours, prices, availability), books appointments, and transfers to an emergency line for urgent matters. For practices (GP, dentist, physio), hotels and service organisations, typically 30 to 60% volume reduction during the day.

Intake conversations come second. An agent that runs a structured intake with new leads or patients (name, complaint, urgency, preferred doctor), puts the answers into the system and probes further on multi-part questions. For clinics, legal advice practices and sectors with lots of intake calls, this works well.

First-line customer service handles the first two minutes of the call: identification, complaint categorisation, simple questions straight from the FAQ, harder ones routed to a human with full context. Saves 40 to 60% of the volume for human agents. Appointment confirmations and reminders can take over 80% of the calling work assistants currently do for healthcare practices. Qualifying customer needs for sales works selectively: for simple B2B products an agent can put qualification questions to inbound leads and book them into the right account manager. Not for consultative sales. And reservations and bookings for restaurants, hairdressers, car rental or kayak rental: high volumes with simple variation are ideal for this technology.

For customer service specifically see /ai-klantenservice.

Which tools do you use?

The Dutch 2026 landscape breaks down into roughly four tiers.

Tier 1: voice platforms (orchestrators). Vapi.ai is the most used by technical teams, programmable via API, supports any ASR-LLM-TTS combination. Strong for custom builds, but has fewer no-code options. Retell AI is similar, a bit more plug-and-play for enterprise, with good call recording and analytics built in. Bland.ai targets outbound (sales and service calls), with lower latency and stronger for large outbound campaigns. ElevenLabs Conversational AI is their own orchestrator with their voices built in, strong on voice quality, less flexible on tooling.

Tier 2: TTS engines. ElevenLabs is the dominant player for Dutch voices. Voice cloning, multispeaker, emotions, regional accents. OpenAI Voice via the Realtime API gives low latency and decent Dutch quality. Cartesia / Sonic is the newcomer, low latency, suited to real-time. Google and Azure TTS are enterprise-grade with EU hosting, only slightly less natural than ElevenLabs in Dutch.

Tier 3: ASR engines. Deepgram is fast and accurate for Dutch, with EU deployment available. AssemblyAI is comparable, strong on real-time. OpenAI Whisper delivers good quality at higher latency than Deepgram or AssemblyAI, fine for non-realtime transcripts. For Dutch dialects (West Flemish, broad Limburg, Twents), WER climbs to 10-15% pretty quickly, whereas with standard Dutch you keep it under 5%. Always test with the people who actually call you.

Tier 4: open source / build your own. LiveKit Agents (WebRTC + agent framework, popular for those who want full control), Pipecat (Python framework from Daily.co, strong for multi-modal agents), or Whisper plus LLM of choice plus open source TTS for anyone who wants to run everything on their own infrastructure (healthcare, defence, financial).

In practice: for 80% of SMB use cases, Vapi or Retell + ElevenLabs + Deepgram + GPT-4o or Claude is the combination that works. For strict GDPR or compliance requirements you move to EU-hosted or open source on your own stack.

Evaluation criteria: how do you know a voice agent works well?

Eight metrics that matter in production.

1. WER (word error rate). What percentage of spoken words does ASR interpret incorrectly? Good: <5% for clear phone quality, Dutch language.

2. End-to-end latency. How long between the user stopping and the agent starting to answer? Good: <1 sec for integrated, <2 sec for classic pipeline.

3. Task success rate. What percentage of calls reach the conversational goal (appointment booked, question answered, complaint routed correctly)? Good: >85% for scoped use cases.

4. Escalation rate. What percentage moves to a human? No fixed benchmark; for first-line, 20-40% on average, with smooth transfer that's no issue.

5. Sentiment score per call. Did the caller experience the conversation as positive, neutral or negative? Steer on trends, not on individual calls.

6. CSAT (customer satisfaction score). Short survey after the call (3 questions via SMS). For production quality, at least 4.0 out of 5.0.

7. Hallucination rate. What percentage of calls contains a factually wrong answer? Target: <2%. Requires tight RAG and monitoring.

8. AHT (average handle time). How much time does an average call take? Compare with human handle time and you get the real ROI.

AI Act and voice AI

For voice AI, at least three AI Act provisions apply to your organisation.

Art. 4 (AI literacy) requires everyone who works with the voice agent or reviews its output to know how it works and what can go wrong. Document training and policy. See /ai-training.

Art. 50 (transparency) is the most important one in production. AI that interacts with humans has to make clear it's an AI. For voice that means: "Hello, you're speaking with our AI assistant" as an opener, or a comparable indication within ten seconds. Not negotiable. Concealing it can trigger fines.

High-risk classification applies to voice AI in HR (phone pre-screening of candidates), in healthcare (medical triage or diagnostic questions) and in critical infrastructure. In those cases your application likely falls into the 'high risk' category, which makes risk management, dataset control, technical documentation and human escalation mandatory. For the broader AI Act context see /ai-act; for a self-scan see /ai-act-checker.

On top of that, GDPR still applies. Voice recordings and transcripts are personal data. Requirements like informed consent, a retention policy, and EU hosting are standard.

Integration with your telephony

For most Dutch organisations, voice AI sits alongside, not instead of, your existing telephony. Three routes.

A dedicated number for the agent is the simplest setup. A new 088 or geographic number that routes to the AI agent 24/7 or outside office hours. Works with Twilio, Telnyx or Vonage SIP trunks.

Routing within your PBX is cleaner if you already run a cloud PBX. Aircall, RingCentral, 3CX and most vendors now support an AI reception integration. Inbound calls go to the agent first, which routes on to humans based on the conversation. Works with your existing numbers and routing rules.

Hybrid with human-first-then-transfer works for sectors where the first voice has to be a person (GP practices, lawyers). During office hours the employee picks up and presses a key to transfer to the agent for specific tasks (for example "book this appointment"). For introducing voice AI in conservative organisations, this is often the most accepted route.

Voice choice and brand identity

An underrated detail: which voice do you give your voice agent? Three routes.

Pre-set voices from Vapi, Retell or ElevenLabs are quickly available and have dozens of Dutch options. No extra work, but the same voice other companies use. Fine for reception. For brand-distinctive applications it feels generic.

Cloning a voice actor's voice is the logical pick for brands with an audio identity (radio spots, podcast intro). With ElevenLabs Professional Voice Cloning you record a voice actor and use that voice as your agent. Cost: 250 to 1,000 euros for an actor who records 30 minutes, plus an ElevenLabs Pro subscription. Result: a unique voice only your company uses.

Cloning an employee's voice is technically the same, but uses a colleague as the voice source. Good for internal agents (HR bot, IT helpdesk bot) where familiarity of the voice gives comfort. Don't forget explicit consent and an agreement on what happens if the employee leaves.

In all three routes, a short audio-transparency line is mandatory ("you're speaking with our AI assistant") under the AI Act. A cloned voice on top of a live employee without that indication is misleading per art. 50.

An honest word about quality issues

Voice agents consistently cause problems on four fronts in 2025-2026. Interruptions are the biggest one. If the caller talks over the middle of your sentence, the agent has to stop, listen, and re-plan. The good orchestrators (Vapi, Retell) have this out of the box. In a self-build, you guaranteed forget it in pilot 1 and your system is annoying in production.

Background noise comes second. Cars, renovation, kids in the background: ASR trips up and the agent responds oddly. Invest in a good VAD (voice activity detection) and a silence-aware fallback. Code-switching comes third. Callers who flip between Dutch and English or between standard Dutch and regional dialect drive up error rates. Pick an ASR that's explicitly multilingual and test with the people who actually call you. For a call centre with lots of Limburg or Brabant on the line, this is no small detail.

Regional data comes fourth and is the most underrated. Postcodes, Dutch street names with diacritics, numbers in spoken form ("twenty-seven" versus "27"): the agent has to normalise those before pushing them into an API or CRM. Skip this and you're back with a human correcting by hand what the agent got wrong. That's the kind of rework that erodes the business case.

What does voice AI cost?

Three cost components.

Per-minute costs during calls: ASR (typically $0.005-0.01/minute), LLM tokens (per call $0.05-0.20), TTS (typically $0.05-0.15/minute with ElevenLabs), telephony (Twilio inbound about $0.01/minute). Total: €0.20-0.50 per call minute. For 10,000 minutes/month you're looking at €2,000-5,000.

Build costs are a scoped project; ask for a fixed price after a conversation. For maintenance you typically pay a fixed monthly fee, because a voice agent needs ongoing care: prompt tuning and knowledge-base updates stay necessary.

ROI math: an agent that handles 5,000 minutes of routine volume per month quickly replaces more than one customer-service FTE in salary plus overhead. Break-even within three months for SMB organisations with enough call volume. Too little volume and you're mostly paying a fixed monthly fee for an agent sitting idle.

When NOT voice AI?

Three scenarios where I advise clients against voice AI.

Low volumes: building a voice agent for 100 calls a month is overengineering. A good human receptionist or an async channel (chatbot, mail form) is cheaper and more effective.

Highly emotional content: for grief care, heavy legal matters, mental-health helplines and dismissal conversations, you don't use an AI voice as first contact. The social and ethical damage doesn't outweigh the efficiency gain. Everyone knows how it feels to get an IVR at a moment when you need a human.

Complex consultative conversation: strategic sales, complex technical helpdesk where deep probing is required, bespoke legal advice. A voice agent can qualify and route, but not replace. Don't try. The NS example at the start of this piece is exactly this: the moment the question falls outside the script, you lose the conversation.

Voice AI vs chatbot vs receptionist: an honest comparison

For many SMB organisations the real question is which channel to use. Three alternatives and when they win.

A classic human receptionist scores better on quality for low volumes and complex conversations. Up to 50 calls per day, or for calls that average longer than 5 minutes, a good human receptionist is cost-comparable and qualitatively stronger. Above that tipping point, the human becomes the bottleneck.

A text chatbot or WhatsApp bot wins on async channel traffic and international reach. Customers who send a message don't expect an instant reply. Customers who call do. For B2B organisations, a chatbot is often enough. For B2C service, voice is usually essential, because the audience prefers to call.

The voice AI agent wins on high volumes with routine conversations and on 24/7 availability without a night shift. ROI is highest exactly where human reception is expensive and doesn't scale.

In practice: combine. Voice AI for the first 60-90 seconds, human for escalation, chatbot for async questions. A good architecture has no one-size-fits-all channel.

How do you start?

Three steps for SMB organisations considering voice AI.

Inventory your use cases. Not "we want voice AI", but "our reception gets 200 calls a day, 60% of which go to voicemail outside office hours". Per use case: volume, call duration, percentage of routine questions, escalation patterns. A free Quickscan via /ai-scan maps this in an hour.

Start scoped. First use case in production within six weeks with limited scope (only appointment booking plus opening hours plus transfer, for example). Measure all eight criteria (WER, latency, task success, escalation, sentiment, CSAT, hallucination, AHT) two weeks after go-live. Iterate.

Scale only what works. If the first use case shows stable numbers, expand to the next (intake, reminders, confirmations). If the first one doesn't work, you've invested six weeks instead of six months, and you know where the limit sits. Stopping a project to learn and choose a different use case is effective steering.

Conclusion

Voice AI is production-ready in 2026 for a specific set of use cases at Dutch SMB organisations: reception outside office hours, intake at high volumes, first-line customer service, appointment confirmations, simple bookings. For those use cases the tech stack is mature, the compliance path is clear, and ROI is achievable within three to six months.

For other use cases (highly emotional, complex advice, low volumes), voice AI in 2026 is not the right tool. A good agency says so too. An AI strategy requires pilots to be practical; otherwise, it's just a report. Start small, measure results, and choose channels based on caller needs.

Want to know if voice AI fits your situation? Book a free call. For the full agent approach and the tools I work with, see /ai-agents. For customer-service-specific voice implementations see /ai-klantenservice. For the AI Act context see /ai-act.