Voice AI agents in business: where do they stand in 2026?

A voice AI agent picks up the phone, understands the request and acts on it: it books the appointment, updates the file, sends the confirmation. In 2026, natural conversation is a solved problem for the best models. Execution is what still gets measured. On the phone, academic evaluations count 31 to 51% of tasks completed, against 85% in writing for the same agent. And eight failures out of ten come from the agent’s behaviour, the rest from the voice layer. In Djtal’s reading, the voice is ready. What makes the difference is the agent behind it.

What is a voice AI agent?

A voice AI agent is a software agent connected to the company’s systems (calendar, CRM, client files) with a voice interface on top. It differs from the interactive voice response system with its fixed menu (“press 1”) and from the smart answering machine, which takes a message without acting. The 2026 generation runs in full duplex: the model listens and speaks at the same time, waits when the caller searches for words and accepts being interrupted.

What do the 2026 evaluations measure?

Four numbers locate the state of the art, sources at the bottom of the page.

What is measured	2026 result
Delay before the first sound	0.8 to 1.3 seconds for the best models; around 200 ms between two humans
Turn-taking (listening, accepting interruptions)	91 to 98% for the leaders: close to solved
Tasks completed over the phone	31 to 51% in clean conditions, against 85% in writing for the same agent; 26 to 38% with noise or accents
Dictated names, emails and reference numbers	14 to 41% exact success; a spell-and-confirm protocol wins back 14 to 31 points, at a cost of 21 to 28 seconds per call

One usage figure sits on top of these technical measures: 87% of customers demand access to a human when a company uses generative AI for customer service (Gartner, February-March 2026 survey of 3,566 customers).

Why does the agent matter more than the voice?

The most useful research result of the year fits in one line. 79 to 90% of a voice agent’s failures come from its behaviour as an agent, the rest from the voice layer. Recent voice stacks have settled their side of the problem. Mistral, for instance, released Voxtral 3 on 20 October, a complete voice stack that can be hosted in Europe or on the client’s premises. What decides the outcome of a call is now what the agent can do: remember previous conversations, reach the calendar and the client file, stop at the right moment, hand over. Agentic AI already does that work in writing inside companies. That same work is what now gets wired to the phone.

What can a Swiss company expect from it?

The most solid case is the calls a company loses today: in the evening, at lunchtime, during peaks. Every unanswered call is often a client dialling a competitor’s number. A voice agent that takes those calls, recognises the client and handles the simple requests delivers a measurable service. It hands over to a human as soon as the conversation calls for it, a design requirement the Gartner figure above imposes from day one.

On languages, French, standard German and English are covered to varying degrees depending on the model. Swiss German remains out of reach of commercial offerings (around 25% of words mistranscribed in the best research work): a test on your own calls comes before any commitment. On the legal side, the caller must know they are talking to an AI. Any recording requires the participants’ consent (art. 179ter of the Swiss Criminal Code).

Our own first test calls, in October, returned an immediate verdict: an agent that cuts people off makes them hang up.

Where to start?

1. An agent that already does the work in writing. The voice sits on top of an existing agent, with its memory, its accesses and its stopping rules. An agent that fails in writing will fail on the phone.
2. Real calls, on a small scale. A few dozen test calls are enough to expose the flaws that demos hide: interruptions, proper names, changes of mind mid-sentence.
3. Agent metrics. Counting completed tasks, wrong actions and recoveries after an error says more than the transcription rate.
4. A human exit at every step. Visible, fast, with no justification asked of the caller.

Frequently asked questions

Does a voice AI agent replace the switchboard? It first takes the calls the company loses (off-hours, peaks) and transfers to a human as soon as the conversation calls for it. Replacing a human reception that does its job well is a poor starting point.

Which languages does a voice agent cover in Switzerland? French, standard German and English depending on the model. Swiss German calls for a test on your own calls before any commitment.

How long does a voice agent take to answer? 0.8 to 1.3 seconds for the best models, where a human answers about five times faster. Callers perceive that slight delay, which remains compatible with a flowing conversation.

Sources

• Artificial Analysis, speech-to-speech comparison, retrieved 21 October 2026.
• τ-Voice, a benchmark for voice agents on real tasks (arXiv, March 2026).
• τ-Elicitation, voice collection of names, emails and identifiers (arXiv, September 2026).
• Full-Duplex-Bench v3, turn-taking evaluation (arXiv, April 2026).
• Gartner, press release of 4 August 2026: February-March 2026 survey, 3,566 customers.
• Stivers et al., Universals and cultural variation in turn-taking in conversation (PNAS, 2009): ~200 ms between two human turns.
• Mistral, Voxtral 3, 20 October 2026.
• Whisper fine-tuned on Swiss German, preprint (arXiv, June 2026).