Does Zoom use meeting content to train AI models?

Zoom excludes communications-like customer content from training its AI models or third-party AI models. Zoom AI still processes specified content to provide its function. Meeting Summary can use speech-to-text data and optional in-meeting chat. It can also use screen-share OCR. Review processing and storage alongside retention and access even when the training exclusion applies.

Can an administrator disable Zoom AI features for one group?

Zoom documents account and group controls for the availability of individual AI capabilities. Administrators can manage specific capabilities for a group, and selected settings can be locked. User controls depend on the capability and the administrator's configuration. Test the effective result with an account that belongs to the target group, because a configuration screenshot without a user-level execution leaves the enforcement outcome unproven.

Does zero data retention cover every Zoom transcript?

Zoom's documented zero data retention configuration concerns transcripts used for Meeting Summary. It requires the Meeting transcript setting to be disabled and locked. Cloud recording can create a transcript independently, while My Notes and archival have separate settings. Participants and external tools may also retain their own records. Review each transcript-producing path and inspect the resulting meeting assets after a controlled call.

Can DeepInspect inspect native Zoom AI Companion traffic?

Native Zoom-managed inference is outside DeepInspect unless a supported customer-controlled route explicitly sends the model-bound HTTP request through the proxy. A customer-owned service that forwards Zoom-derived content to an LLM can create such a path. It must supply the originating identity and relevant request context. Zoom settings and hosted inference are under their native controls. Meeting assets and third-party application actions are also under their native controls.

Zoom AI Companion Security: Separate Content, Models, and Controls

A meeting host turns on Meeting Summary. Zoom converts speech to text and may process screen-shared content through OCR. It can also include in-meeting chat. That data path is the starting point for Zoom AI Companion security. The name alone leaves out the content source and selected AI deployment model. It also leaves out the retention setting and administrator decision that made processing possible. I want to review those boundaries separately. The hard limit comes first: Zoom-managed AI traffic is inside Zoom's service architecture and cannot be assumed to cross a customer-operated HTTP enforcement point.

TL;DR

Zoom AI tools consume different content. Meeting Summary can use speech-to-text data and optional screen-share OCR. It can also use optional in-meeting chat.
Zoom documents three deployment options: Federated, Zoom-Hosted Models Plus, and Zoom-Hosted Models Only. Their routing differs, as does the availability of individual capabilities.
Account and group controls govern access to each capability alongside user and in-meeting controls. Transcript retention and cloud recording require separate checks.
DeepInspect applies only to authenticated LLM HTTP traffic explicitly routed through its proxy. Native Zoom-managed inference and third-party application actions are outside that boundary.

Zoom AI Companion security starts with the content map

Zoom's AI feature data-handling documentation lists the content used in each Zoom AI workflow. Meeting Summary can use speech-to-text data and optional OCR of shared screens. It can also use optional in-meeting chat messages. In-meeting Questions adds user prompts. Chat composition uses message text and participant names, along with user input. ZoomMate can draw on Zoom content and connected sources, as well as web content and local files.

Treat each workflow as a separate processing route. For Meeting Summary, record the meeting owner and enabled inputs. Add the group, transcript setting, and output location. For ZoomMate, add each connected source and its authorization. A single inventory row labeled "AI Companion" throws away the details a reviewer needs.

I would reject that row during production approval. One screenshot showing the Zoom AI settings pane beside a meeting transcript tells a sharper story: the control is named, and the content it governed is visible.

Three model deployments create different routes

The Zoom AI Whitepaper model architecture describes three service models. The Federated Approach can route work across Zoom-hosted models and third-party model providers. Zoom names Anthropic and OpenAI in that path, while particular workflows can use other providers. Zoom-Hosted Models Plus uses select models through Zoom-managed Amazon Bedrock instances. Zoom-Hosted Models Only keeps AI processing on models hosted within Zoom's infrastructure.

Those choices affect data routing and determine which capabilities are available. The whitepaper gives the Federated Approach the broadest support, while ZM+ and ZMO can carry limitations. Customers should contact Zoom support to confirm their default deployment and available alternatives.

Put the selected service model into the architecture record. Add the account region and enabled capability beside it. The useful artifact names the actual route rather than drawing one cloud marked "Zoom AI." Recheck it after Zoom rolls out a capability because Zoom says federation models can change as the service changes.

Administrative controls require executed tests

Zoom's administrative-controls documentation places management of individual capabilities at account and group levels. Selected settings can be locked so users cannot change them. Some capabilities also expose user controls. During meetings, hosts can manage supported AI tools. Participants receive in-product indicators and can request that Meeting Summary or Meeting Questions be disabled.

Test the hierarchy under a named group. Disable one capability at account level and lock it, then sign in as a group member. Preserve the setting and user-visible result with the timestamp. Repeat with an allowed capability during a synthetic meeting.

Custom AI Companion adds another control surface. Zoom documents an account or group setting that permits AI to perform tasks in third-party apps. That permission is part of the application-action register. The receiving app still controls its own credential and object permissions. It also controls the approval flow and transaction record. An inference policy cannot receive credit for a calendar update it never sees.

Retention and training statements answer different questions

Zoom's data-governance and privacy documentation says customer communications-like content is excluded from training Zoom's AI models and its third-party AI models. The same documentation explains that Zoom AI still uses selected customer content to provide the service. A training exclusion therefore leaves the processing and retention review active.

Meeting Summary provides a concrete example. Administrators can disable and lock the Meeting transcript setting so Zoom uses temporary speech-to-text data and discards it after successful summary generation. Zoom calls this a zero data retention approach for those transcripts. Cloud recording can independently create and retain a transcript. My Notes and archival use their own settings, and external participants or tools may keep separate records.

Review those switches together. Export the account and group settings, then open the cloud-recording configuration. Verify the assets after a synthetic meeting. A blank transcript folder beside a retained summary is physical evidence. A policy document alone proves intent.

Request policy has a narrow HTTP boundary

Native Zoom AI inference runs on a Zoom-managed service path. The three whitepaper deployment models describe provider selection inside that architecture, rather than a customer-configurable LLM endpoint. An external AI gateway therefore lacks an interception point for native Meeting Summary, ZoomMate, or Chat composition traffic unless Zoom exposes a supported customer-controlled route and the customer deliberately configures it.

A separate application may create such a route. A customer-owned service can receive an approved Zoom event and send its own authenticated HTTP request to an LLM endpoint. It must preserve the originating identity and classify the prompt assembled from Zoom content. It must also route the call through the policy point. AI policy enforcement at the HTTP layer applies to that request. The post-authentication gap explains why a service credential gives insufficient context for content-specific authority.

Zoom settings and notices are Zoom responsibilities. Third-party app actions are under Zoom and the receiving application. Keep the route diagram honest about each owner.

A production review produces five records

Use five executed records for the Zoom deployment.

Feature inventory: name Meeting Summary, ZoomMate, Chat composition, or another enabled capability and list the exact content inputs.
Model route: record the selected option, Federated, ZM+, or ZMO. Add the account region and confirm which capabilities are available.
Administrative decision: preserve the account or group setting and lock state. Include the administrator and observed user result.
Retention outcome: capture transcript and summary behavior under a synthetic meeting. Also capture cloud-recording, My Notes, and archival behavior.
Customer-owned inference: where a separate service sends Zoom-derived content to an LLM, retain caller identity and classification. Also retain the destination, policy version, and decision.

The fifth record is required only with that customer-controlled HTTP call. AI gateway architecture shows the placement. Zoom-native processing should retain its native evidence rather than borrowing claims from an external proxy.

DeepInspect

This is the gap DeepInspect closes on a supported customer-controlled model route. DeepInspect is inline as a stateless proxy between an authenticated user or agent and an LLM endpoint. The calling application supplies identity and classification context. DeepInspect evaluates that context with the destination and requested operation under a versioned policy before forwarding the HTTP request.

Each routed decision can allow, deny, or redact the request and produces a signed, tamper-evident record. Native Meeting Summary and ZoomMate processing are outside the proxy when Zoom manages the endpoint. The same boundary excludes Zoom settings and cloud recordings. It also excludes transcripts created by other Zoom tools and actions executed in third-party applications. Book a technical deep dive at deepinspect.ai.