Barometer

Where are your meetings processed? The sovereignty barometer for AI meeting-notes tools

Fourteen AI transcription and meeting-notes tools, described from what their vendors publish on four points: hosting, the use of a third-party language model for inference, the on-premise offer and its price.

Methodology

This barometer is published by Recapro, one of the vendors compared. The facts come from the public sources cited below; keep that context in mind.

Last checked: October 16, 2026. Review: quarterly.

Scope
Fourteen AI transcription and meeting-notes tools available in France, Recapro included. The list is not exhaustive.
Sources and date consulted
Each cell draws on the vendor’s public documents: privacy policy, data processing agreement, subprocessor list, trust center or pricing page. All were consulted on October 16, 2026 and are linked at the bottom of the page with their archived copy on the Wayback Machine.
Criteria
Hosting: provider and storage region as published. Third-party LLM: external model provider the vendor names for transcription or summaries. On-premise: a published offer that runs the service or stores data on your infrastructure. Public on-premise pricing: the price of that offer shown on the vendor’s site. The table does not score or rank the tools; they are listed alphabetically.
“Not published” and “Not applicable”
“Not published by the vendor (2026-10-16)” means the information does not appear in the vendor’s public sources on that date. It does not prejudge the vendor’s actual practices. “Not applicable” means no on-premise offer is published, so no price can be.
Corrections and right of reply
A vendor can report an error or an update, or exercise its right of reply, at [email protected]. We correct the table and its verification date.

Dicte.ai (France)

Published hosting
Europe; storage region France or EU depending on the plan
Third-party LLM at inference
Not published by the vendor (2026-10-16)
On-premise
Yes, DicteBOX (Enterprise plan)
Public on-premise pricing
No (on quote, 10 licences minimum)

Fathom (United States)

Published hosting
Google Cloud (United States)
Third-party LLM at inference
Anthropic, OpenAI and other listed AI providers
On-premise
Not published by the vendor (2026-10-16)
Public on-premise pricing
Not applicable

Fireflies.ai (United States)

Published hosting
Google Cloud and AWS (United States)
Third-party LLM at inference
OpenAI, Anthropic
On-premise
Not published by the vendor (2026-10-16); private storage on the Enterprise plan
Public on-premise pricing
Not applicable

Gladia (France)

Published hosting
OVH and Shadow (France); AWS and Google Cloud (EU and United States) depending on the service
Third-party LLM at inference
OpenAI, Together AI, Google Cloud (add-ons)
On-premise
Not published by the vendor (2026-10-16); custom hosting on the Enterprise plan
Public on-premise pricing
No (Enterprise on quote)

Granola (United States / United Kingdom)

Published hosting
AWS (United States)
Third-party LLM at inference
OpenAI, Anthropic
On-premise
Not published by the vendor (2026-10-16)
Public on-premise pricing
Not applicable

Leexi (Belgium)

Published hosting
AWS (France); Scaleway storage (France)
Third-party LLM at inference
Microsoft Azure, Mistral AI
On-premise
Not published by the vendor (2026-10-16)
Public on-premise pricing
Not applicable

MeetGeek (Romania)

Published hosting
AWS (EU for the European instance, United States for the US instance)
Third-party LLM at inference
OpenAI
On-premise
Optional storage of recordings on your infrastructure (Enterprise plan)
Public on-premise pricing
No (Enterprise on quote)

Modjo (France)

Published hosting
AWS (European Economic Area)
Third-party LLM at inference
Microsoft, OpenAI, Google
On-premise
Not published by the vendor (2026-10-16)
Public on-premise pricing
Not applicable

Noota (France)

Published hosting
Google Cloud (France, Belgium, Netherlands)
Third-party LLM at inference
Not published by the vendor (2026-10-16)
On-premise
“Privately hosted AI & Data” option (Enterprise plan, from 50 licences), deployment model not specified
Public on-premise pricing
No (on quote)

Otter.ai (United States)

Published hosting
AWS (United States)
Third-party LLM at inference
Anthropic, OpenAI
On-premise
Not published by the vendor (2026-10-16)
Public on-premise pricing
Not applicable

Plaud (United States)

Published hosting
On the device by default; optional cloud sync in the user’s region (United States, Germany, Japan or Singapore)
Third-party LLM at inference
GPT, Claude and Gemini, according to its Trust Center
On-premise
Not published by the vendor (2026-10-16)
Public on-premise pricing
Not applicable

Pocket (United States)

Published hosting
United States (“US-based hosting”), provider and region not published
Third-party LLM at inference
GPT-5 (OpenAI), Claude (Anthropic) and other models, according to its home page
On-premise
Not published by the vendor (2026-10-16)
Public on-premise pricing
Not applicable

Recapro (France)

Publisher of this barometer

Published hosting
Scaleway, fr-par region (France); TLS terminated by Cloudflare
Third-party LLM at inference
No third-party LLM (open-weight models running on our servers)
On-premise
Yes, Fortress On-Premise (can be deployed air-gapped)
Public on-premise pricing
Yes (€19.90 per seat per month, 10 seats minimum, hardware extra)

tl;dv (Germany)

Published hosting
Google Cloud, Hetzner and Wasabi (European Economic Area)
Third-party LLM at inference
Anthropic, via Google Cloud Vertex AI
On-premise
Not published by the vendor (2026-10-16)
Public on-premise pricing
Not applicable

Why hosting alone does not guarantee confidentiality: see the AI sovereignty & inference explainer.

Sources consulted on October 16, 2026

Frequently asked questions

01

Which AI meeting-notes tools host your meetings outside the European Union?
According to their public sources as of October 16, 2026, Fathom, Fireflies.ai, Granola and Otter.ai host in the United States. Gladia and MeetGeek combine European and US infrastructure depending on the service or the instance chosen. Plaud stores on the device by default and, if you turn on sync, in the user’s region. Pocket states US-based hosting without naming its provider or region. The other tools publish EU hosting; at Recapro, data at rest sits with Scaleway (fr-par region) and Cloudflare, a US company, terminates TLS in transit. Where data is hosted says nothing about whether a third-party model is called at inference time.
02

Which tools send inference to a third-party language model?
Eleven vendors in the table name at least one third-party model provider in their public sources, among OpenAI, Anthropic, Google, Microsoft, Mistral AI and Together AI. Dicte.ai and Noota do not publish the information, which does not prejudge their practices. Recapro calls no third-party LLM: its open-weight models run on its own servers.
03

Which AI transcription tools offer an on-premise option?
As of October 16, 2026, two vendors publish an on-premise offer: Dicte.ai (DicteBOX, Enterprise plan on quote) and Recapro (Fortress On-Premise, public price). MeetGeek offers, as an Enterprise option, storage of recordings on your infrastructure. MeetGeek (recording storage on your infrastructure) and Noota (“Privately hosted AI & Data”) offer a private hosting option on their Enterprise plan, without stating whether processing runs at the customer’s site. The other vendors publish no on-premise offer, which does not rule out one on request.
04

What is the best AI meeting-notes tool in France?
The barometer does not name a best tool. It describes four criteria published by the vendors, and the right choice depends on your requirements for each, and on criteria it does not cover: features, ease of use, integrations, budget. Every row links to the vendor’s source so you can check for yourself.
05

Which tools in the barometer are headquartered in France?
Five vendors in the table are headquartered in France: Dicte.ai, Gladia, Modjo, Noota and Recapro. The country of the headquarters does not tell you where data is processed. For that, read the hosting and third-party LLM columns, which reflect what each vendor publishes.
06

Which AI meeting-notes tool suits a regulated sector such as banking or legal?
Two questions come first: where the content is processed and who can technically access it. An on-premise offer keeps processing on your servers; only a deployment isolated from the network (air-gap) rules out any outside access. Depending on your security policy, other criteria matter too, such as encryption, certifications or processing agreements. The table does not cover certifications.
07

How is this barometer compiled?
Each cell draws on the vendor’s public documents (privacy policy, data processing agreement, subprocessor list, security or pricing page), consulted on October 16, 2026, with a link to the source and its archive. “Not published” means the information is missing from those sources on that date, without prejudging actual practices. The table is reviewed every quarter. A vendor can report an error or exercise its right of reply at [email protected].
08

Which tool should a council with sovereignty requirements choose?
Start with where processing happens. A tool whose transcription and summaries run on your servers keeps the session audio on your network; a cloud tool sends it to the vendor and, where applicable, to its model providers. The table shows, for 14 tools, what each vendor publishes on these points.

See the Recapro plans

Recapro Cloud: data at rest in France at Scaleway (Postgres and object storage, fr-par region); in transit, Cloudflare terminates TLS. Fortress On-Premise: processing stays on your servers. The models are open-weight, with no third-party LLM.

See plans and pricing Read the explainer

From the barometer to the build

The barometer measures the gap between marketing claims and terms of service. These pages say what we do about it.