Last updated: 9 October 2026

MIHOS carries the swiss made software + AI label. AI is the core technology of the product, not an added feature: without it MIHOS does not function. The application is designed, developed and operated in Switzerland by MuuM Sàrl, in Nyon. The AI models it runs on are third party and they are not Swiss made. What is Swiss made is everything built around them: the capture application, the consent layer, the scoring engine, the playbook logic and the CRM write-back.

Hosting

MIHOS is a SaaS platform. Recordings, transcripts and account data are stored in Switzerland. Speech to text and speaker separation are performed by Gladia, a French company that processes recordings in France; its sister company Gladia Inc. (United States) is listed among its sub-processors, and we are obtaining a written commitment that our customers' data is processed only in the European Union. Language model inference runs in France, on infrastructure operated by OVHcloud. Practice sessions with an AI client use ElevenLabs, a United States provider, for live speech recognition and for the synthetic voice.

Level

We build on foundation models consumed as a service and wrap them in our own layer: orchestration, the playbook context, the scoring engine, the coaching logic and the CRM write-back. The models are a component of the application, not a system we operate ourselves.

AI technologies used

Automatic speech recognition, speaker diarisation, large language models for analysis, scoring and coaching text, and speech synthesis driven by a conversational agent for practice sessions. Speaker recognition by voice signature exists as an option, off by default, offered only on the conditions set out on our Voice & biometric data page.

Model origin and type

Established third party models, used as delivered through secured APIs: open weight language models served by OVHcloud, and commercial speech recognition, speaker separation and speech synthesis models from Gladia, ElevenLabs and, where speaker separation or recognition by pyannoteAI is enabled, pyannoteAI. We operate no proprietary foundation model and we run no model on our own hardware.

Customisation and fine tuning

We do not train and we do not fine tune any model. Behaviour is steered through prompts, retrieval and your own playbook context. We do not use meeting content to train or improve a model, and at each provider that offers it, we switch off the use of our customers' content to improve its models. MIHOS does get better with use, but at the application layer: the product accumulates your playbook, your real objections and the outcome of your deals, and feeds them back as context. Your data does not change the models. We may change a model or its version; a change of provider is announced under the procedure that applies to sub-processors.

What the AI is not allowed to do

No emotion inference. The sentiment and emotion detection features our providers offer are not enabled in our pipeline. No voiceprint without explicit consent: optional speaker recognition, off by default, applies only to a rep who consents, and never to the people the rep meets. No inference of special category data from a voice or a transcript. Scores and recommendations are indicative and AI output can be wrong: every decision about a person involves human judgement, every write into a customer system is logged and reversible, and every score is traceable back to the timestamped passage that produced it, as long as the transcript is kept.