Meeting AI on your own cloud gateway: how recaps, summaries and the in-meeting assistant stay inside your Azure, Google or AWS perimeter

Every meeting platform now ships AI features, and every AI feature adds at least one company to the list of who sees your meeting content. For most organizations that list already has a shape: the language-model gateway they use for their own AI — Azure OpenAI with their Microsoft 365 tenant, Google Vertex AI with Google Workspace, Amazon Bedrock with their AWS estate. Their compliance team has reviewed that gateway once, with its DPA, its region and its zero-training terms.

InterMIND runs its AI features on that same gateway. This post explains what that means in practice, what leaves your organization for an AI call and what does not, how the setting works, how other meeting tools handle the same question in their public documentation, and what a procurement reviewer can check without taking our word for it.

What "the same gateway" means

An organization's AI features in InterMIND — meeting recaps, document summaries, the writing assistant in notes, Ask AI, and Mia, the AI participant you address by name in a meeting — all run through one language-model gateway per organization. An organization admin picks the gateway on the Billing page, and the choice applies to every AI feature at once.

The consequence for a compliance review is simple: no new AI sub-processor enters the perimeter. An organization on Microsoft 365 whose meeting recaps run on Azure OpenAI is dealing with a party it has already approved, under the Microsoft Products and Services DPA and the EU Data Boundary it already relies on. The same holds for Google Workspace customers on Vertex AI. An organization on AWS goes one step further: with Bedrock in its own account, the meeting content never even passes through InterMIND's cloud tenant on the way to the model.

The gateways InterMIND uses are configured with zero data retention and no training on customer content — settings that are Azure OpenAI's default for the EU Data Zone and Google Cloud's contractual position for Vertex AI (see the sources at the end). InterMIND's own subprocessors page lists Microsoft and Google with region and purpose; it lists Google only for organizations that selected Vertex AI.

What leaves the organization for an AI call, and what does not

Each AI feature sends only what it needs, once:

• Meeting recap: the meeting transcript and the chat window of that session, at meeting end.
• Document summary: the extracted text of a document shared in the meeting chat.
• Writing assistant: the text of the note being edited.
• Ask AI: the question and the documentation it searches.
• Mia in the meeting: the question addressed to her, the transcript so far, the chat and shared documents of the session, and the recaps of the room's previous sessions.

What never goes to a language model: audio and video streams, recordings, and speech recognition. Live translation of voice, chat and notes runs on InterMIND's own engine in the EU — it is a separate pipeline from the AI features and does not use any hyperscaler gateway. The Security page has the full table of what each party sees.

The off switch

Some organizations do not want any language model near their meetings, whichever gateway it is. The AI features switch on the Billing page turns every AI feature off for the whole organization: recaps arrive without a summary, documents are shared without a summary, the writing assistant and Ask AI are hidden, and Mia does not listen. Transcription and translation keep working, because they do not depend on a language model. Turning the switch back on restores everything.

This is the setting procurement questionnaires increasingly ask about first — "can AI features be disabled at the tenant level?" — and the answer is yes, per organization, by its own admin.

Mia: an AI participant that speaks every participant's language

Mia is the part of this that people notice in a meeting. Say her name — "Mia, what did we agree on last time?" — or type @Mia in the meeting chat, and she answers from this meeting's transcript, chat and shared documents, and from the recaps of the previous sessions of the same room.

Two design decisions matter for the perimeter question:

• Mia runs on the organization's gateway, like every other AI feature, and her voice is synthesized by the same vendor (Azure AI Speech or Google Cloud Text-to-Speech, EU regions). On an organization's own endpoint she answers in text.
• Mia is not a participant of the call. Her answer is written natively in each language present in the room and voiced on each participant's own device. Nothing of hers enters the audio stream, the recording or the transcript, and every participant hears her in their own language.

How meeting tools handle the same question today

The table states what each vendor's public documentation says about where its AI runs and what the customer can choose. It is a list of documented facts, not a ranking.

Two things stand out in that table without any verdict. Region choice for AI processing exists in the category (tl;dv documents EU or US). A choice of the gateway itself — running the vendor's AI on the customer's own hyperscaler tenant, or in the customer's own AWS account — is not documented by the vendors above; it is the axis this post is about.

What a procurement reviewer can verify

The questions below are the ones that recur in AI vendor security reviews in 2026 — a vendor-side guide to those reviews lists written zero-training policies extending to third-party model providers, sub-processor transparency naming the model provider and region, and the ability to disable AI as consistently requested items (see the sources). Here is where each answer is verifiable for InterMIND:

• Which model provider, in which region? The subprocessors page names Microsoft (Azure OpenAI, EU Data Zone) and, for organizations that selected it, Google (Vertex AI, EU multi-region). The gateway an organization uses is visible to its admin on the Billing page.
• Does the model provider train on our content? No. Zero data retention and no training are the configured terms on both gateways; the provider's own documentation for those terms is linked in the sources.
• Can we keep our content out of your AI vendor entirely? Yes: bring your own endpoint, or, on AWS, your own Bedrock account (in preparation). The key you enter is stored encrypted and never displayed again.
• Can we turn AI off? Yes, per organization, by its admin; transcription and translation are unaffected.
• What does an AI call contain? Listed feature by feature above and on the Security page — never audio, never video, never recordings.

FAQ

Which AI providers does InterMIND use for meeting recaps? By default, Azure OpenAI in Microsoft's EU Data Zone, from InterMIND's own tenant. An organization admin can switch the organization to Google Vertex AI on Google's EU multi-region endpoint, to Amazon Bedrock in the organization's own AWS account (in preparation), or to the organization's own OpenAI-compatible endpoint. The choice applies to every AI feature of the organization.

Can we run InterMIND's AI features on our own Azure or Google Cloud tenant? Not on your tenant — on ours, on the same gateway your tenant uses: Azure OpenAI or Vertex AI. For your own tenant or your own hardware, choose Your own endpoint and point InterMIND at any OpenAI-compatible server, for example vLLM in your data centre or a model deployment in your own cloud project. For AWS, the Bedrock option works in your own account by design.

Can AI features be disabled for the whole organization? Yes. The AI features switch on the Billing page turns off recaps, document summaries, the writing assistant, Ask AI and Mia for every meeting of the organization. Transcription and live translation keep working, because they run on InterMIND's own engine, not on a language model.

Does the meeting audio go to the AI provider? No. Audio and video streams, recordings and speech recognition never reach a language model. The AI features receive text: the transcript and chat window at meeting end for a recap, a document's extracted text for its summary, a note's text for the writing assistant, a question for Ask AI or Mia.

Is meeting content used to train the AI models? No. Azure OpenAI in the EU Data Zone and Vertex AI are configured with zero data retention and no training on customer content; the providers' documentation of those terms is linked in the sources below. On your own endpoint, retention and training are whatever you configure.

Where does Mia's voice come from, and is she in the recording? Mia's voice is synthesized by the same vendor as the organization's gateway — Azure AI Speech or Google Cloud Text-to-Speech, in EU regions. She is not a participant of the call: her answer is voiced on each participant's device in that participant's language, so it is not in the audio stream, the recording or the transcript.

What happens to an organization's AI settings for guests in a meeting? The meeting host's organization decides. Guests use the host's AI features — including addressing Mia — under the host organization's gateway and its switch; a guest cannot change either.

See it for yourself

• Read the Security page — the AI gateway table, what each party sees, and how to switch AI off.
• Try the live demo — the production translation pipeline on your own audio, no signup.
• See the subprocessors list — vendor, country, purpose, data, location and safeguards for every party.

Sources: Zoom sub-processors (zoom.com/en/trust/subprocessors, page updated 2026-04-29); Otter.ai (Claude customer story, Otter Enterprise); Fireflies.ai (security, Private Storage, 2023-08); tl;dv (security commitment, GDPR-compliant meeting assistants, 2025-12-08); Azure OpenAI data handling and EU Data Zone (Data, privacy, and security for Azure OpenAI); Google Cloud generative AI data governance for Vertex AI (Generative AI and data governance); Amazon Bedrock data protection (Data protection in Amazon Bedrock); AI vendor security review items (Merciv, AI vendor security review process, 2026-07-07); InterMIND statements verified against the shipped code and /legal/subprocessors; all checked September 2026.