A voice agent is a real-time system, not a chatbot with a microphone. Mistral and monday.com build it two ways

In a text chat, an agent that takes four seconds to answer feels slow. On a phone call, four seconds of silence feels like a dropped line, and the caller starts talking again just as the agent begins to answer. Now both are speaking, the agent has to decide whether it was interrupted, and the transcription it is working from is already out of date.

That is why voice agents are a different engineering problem from text agents, and two launches in two days show the two ways vendors are solving it.

Two architectures for the same constraint

Mistral’s Voxtral 3, released October 20, is a set of components. It includes transcription models (Voxtral Small Transcribe 3 for accuracy, Mini Transcribe 3 for efficiency, both covering more than 50 languages with speaker diarization), a streaming model, Voxtral Mini Transcribe Realtime 3, with end-of-turn detection and latency configurable from 80 milliseconds to 2.4 seconds, and a text-to-speech model with control over emotion, pacing and emphasis. Mistral’s documentation lists the realtime model in public preview.

On top sits a voice agent SDK built on LiveKit that runs speech in, reasoning, tool calls and speech out as one streaming pipeline, over SIP telephony or WebRTC. The agent stops speaking when the caller cuts in, and end-of-turn prediction combines acoustic and semantic cues to reduce false cut-offs. Each stage is a separate model, so the language model, the voice or the language can be swapped without rewriting the agent. Pricing is published per minute: $0.008 for realtime transcription, $0.04 for the full voice agent API with Mistral Small 4, $0.08 with a larger model. Mistral offers an EU-hosted API plus U.S. and global endpoints, and self-deployment through its sales team.

monday.com’s Harmony, launched October 19, is a product. It handles sales calls from prospecting to close for mid-market and enterprise customers in the U.S., U.K., Australia and Singapore. Instead of one large model, it runs small proprietary models, each for one task (qualifying a lead, capturing data, handling an objection), coordinated by a real-time predictive engine, with telephony built on Twilio. monday.com reports about 200 milliseconds from audio in to audio out, warm transfers to human reps with a briefing, and write-back to major CRMs. The voice technology came from its acquisition of OneAI.

Vendor latency figures on one scale

0 ms1 second

1. 80 msLowest configurable delay, Voxtral Realtime 3 transcription
2. About 200 msHarmony, audio in to audio out
3. Under 500 msVoxtral realtime transcription
4. Under 1 sVoxtral SDK, full agent turn

What the latency budget forces

Tools have to be fast, or the agent has to talk while it waits. A CRM lookup that takes two seconds is fine for a text agent and fatal on a call. Voice agents need tool calls designed for the budget: precomputed context, cached account data, or a spoken holding phrase while a slower system responds.

Small specialized models make sense here. Harmony’s design is the same argument as decision models on the agent’s hot path, pushed to real time: if most steps of a conversation are bounded decisions, a frontier model on every step is too slow before it is too expensive.

Modular versus packaged is a build or buy choice. Mistral’s components let a company keep control of each stage, choose where audio is processed and replace parts as better ones appear, at the price of assembling and operating the pipeline. Harmony gives a sales team a working system with CRM integration, at the price of running on monday.com’s infrastructure and its choices. Neither is better in general; the deciding questions are who must own the audio and who will operate the system at 2 a.m.

The controls a call needs that a chat does not

Consent and disclosure. In the U.S., the FCC ruled in 2024 that AI-generated voices count as artificial voices under the Telephone Consumer Protection Act, so outbound AI calls need the prior consent that rule requires. Disclosure, recording consent and opt-out rules vary by state and country.

Audio as data. A recording is personal data, and becomes biometric data once voiceprints are derived from it. Decide retention before go-live; Mistral says audio retention can be controlled, which is the question to ask every vendor.

Handoff as a designed path. Bright Car Finance, the U.K. lender monday.com cites, says the agent follows up with every approved applicant right away, qualifies about half of its conversations and transfers 80% of those to advisers live, and reports a 9% increase in deals closed in less than a year. The figures are the customer’s, not audited, and they describe an agent whose main job is to reach a person faster, not to replace one. That is a sound place to start: measure the handoff, not only the automation rate.