AI Companion Privacy & Security: Is Meeting Data Used to Train Models? Retention Rules Explained (2026)

Published October 28, 2026

Published October 28, 2026 · 11 min read

If you are a lawyer, compliance officer, or executive evaluating whether the built-in AI Companion in Zoom or the Copilot meeting assistant in Microsoft Teams is private enough for sensitive client calls, the short answer is this: both vendors state that your meeting content is not used to train their own or third-party models, but "not training" is only one of three questions you must answer. You also need to know (1) where the transcripts, summaries, and recordings are stored, (2) how long they are retained — including by third-party model providers like OpenAI and Anthropic — and (3) whether those artifacts become discoverable in litigation. This guide answers all three, names the real options, and shows where an on-device notetaker like Basil AI fits.

What "AI Companion" actually means in 2026

The phrase "AI Companion" is overloaded. In most buyer searches it refers to Zoom AI Companion, Zoom's built-in generative-AI assistant that produces meeting summaries, next steps, and in-meeting Q&A. But the same buyer question — "is the AI meeting summary private?" — also applies to Microsoft 365 Copilot in Teams, Google's Gemini in Meet, and standalone bots like Otter, Fireflies, and Granola. This page focuses on the two the models cited most heavily in buyer research — Zoom AI Companion and Teams Copilot — and then contrasts them with on-device tools.

According to Zoom's AI Companion 3.0 launch announcement, the product uses a "federated" architecture that routes work across Zoom's own LLMs and SLMs plus third-party models from OpenAI, Anthropic, and NVIDIA's open Nemotron family. That architecture is central to understanding the privacy trade-offs — because when third parties are in the data path, the retention story changes.

Does Zoom AI Companion use my meeting data to train models?

Zoom's position is unambiguous. In its own documentation, the Zoom newsroom states that Zoom "does not use communications-like customer content to train its own or third-party models." That covers audio, video, chat, screen sharing, attachments, poll results, whiteboard content, and reactions. The commitment has been echoed in every university IT review of the product — for example, Tufts Technology Services and the University of Pennsylvania Law School both confirm the no-training policy for enrolled tenants.

So on the training question, Zoom's answer is "no." But training is only one risk vector. The other two — retention and third-party exposure — are where the subtleties live.

Third-party model providers: what OpenAI and Anthropic may retain

Here is the clause most buyers miss. Zoom's own AI Companion bluepaper explains that when a feature relies on a third-party AI model, "the third-party model provider may retain the content used to provide the service for trust and safety purposes, within the U.S., for up to 30 days," unless a longer period is required by law. That means your meeting transcript — the one generated to produce the summary — can briefly sit on Anthropic or OpenAI infrastructure. Zoom itself may also retain customer content for "up to 30 days for support and debugging purposes."

The Claude-OpenAI data path is confirmed in Anthropic's own Zoom case study, which describes Claude as the model behind AI Companion's meeting summaries. For a buyer asking "who sees my meeting content?" the correct answer is: Zoom, the model provider that generated the specific summary, and anyone your admin shares it with — not just Zoom.

How long does Zoom AI Companion keep summaries, transcripts, and recordings?

Zoom's native retention rule is counterintuitive: Meeting Summaries persist until a human deletes them. The official Zoom-supplied text republished by St. Edwards University states that "Zoom's AI Companion Meeting Summary is retained until the meeting summary is deleted by the user, account administrator, or until the user's or customer account is terminated." There are no custom retention settings for the summary itself — it is on the host, the admin, or an institutional overlay policy.

In practice, large institutions layer their own retention on top. Johns Hopkins IT deletes AI Companion summaries and Smart Recordings after 180 days. Tufts auto-deletes after 730 days, and the University of Pennsylvania Law School runs a 120-day retention period. Your firm's default is "never" unless someone turns on a policy.

Zoom's optional "Zero Data Retention" setting

Buyers who need stronger guarantees should know about the opt-in feature. Per the Zoom AI Companion bluepaper, Zoom offers a Zero Data Retention option that deletes the temporary transcript, OCR of shared screens, and in-meeting chat used to generate a Meeting Summary "immediately after the summary is created." If summary generation fails, inputs are kept up to 24 hours for retries. The catch: it is not on by default. Firms must open a support ticket or contact their account team to enable it.

Does Microsoft Copilot in Teams use meeting transcripts to train models?

Microsoft's answer is also "no." According to a Microsoft Learn Q&A answer, Microsoft does not use Teams meeting transcripts to train AI models, and customer data — including prompts, responses, and transcripts — is not used to train the foundation large language models or for general model improvement outside your organization. Independent analysis at Sagnik Bhattacharya's privacy write-up confirms Microsoft's stated position that intelligent-recap customer data "isn't used to train or test AI models."

But Copilot prompts and responses still live in the Microsoft 365 compliance substrate

This is the quiet detail every lawyer should know. Microsoft's own support page, Use Copilot without transcribing or recording a Teams meeting or call, is explicit: even when transcription is off, "Copilot prompts and responses may be retained in the Microsoft 365 compliance substrate and are subject to your org's Microsoft Purview retention policies." A Microsoft Tech Community healthcare post adds that those interactions may also be discoverable through Microsoft Purview eDiscovery. In other words: turning off the transcript does not turn off the record.

A June 2026 Microsoft roadmap entry, RM558933, goes further — admins can enable a tenant-wide policy that lets Copilot (and Facilitator) retain key insights from meetings beyond transcript retention periods, so Copilot can continue answering questions about past meetings even after the underlying transcripts expire. If your firm hasn't audited this toggle, your "deleted" meetings may not be as deleted as you think.

Comparison: AI Companion data handling across the main options

Here is how the dominant AI meeting assistants compare on the dimensions that matter for a privacy and compliance review. All claims are drawn from the vendor's own documentation.

What the ABA says about using cloud AI notetakers on client calls

The authoritative legal-ethics framework is ABA Formal Opinion 512, issued July 29, 2024. According to the National Conference of Bar Examiners' explainer, Opinion 512 lays out six duties lawyers must apply to generative AI — competence, confidentiality, communication, candor, supervision, and reasonable fees — and tells lawyers to read and understand the Terms of Use and privacy policies before putting client information into any GAI tool.

The confidentiality analysis is the one that matters here. A detailed rule-by-rule review at The Legal Prompts captures the key sentence: before inputting client information, lawyers "must evaluate the risks that the information will be disclosed to or accessed by others outside the firm." For self-learning tools, informed consent is required, and GC AI's analysis of Opinion 512 notes that the opinion specifically rejects relying on generic boilerplate buried in an engagement letter. The practical consequence: a lawyer using Zoom AI Companion on a privileged call needs to be able to explain, to a client, exactly which third-party model providers see the transcript and for how long.

Are AI Companion meeting summaries discoverable in litigation?

Yes. Both Zoom AI Companion summaries and Microsoft Copilot interactions are electronically stored information (ESI) subject to the Federal Rules of Civil Procedure. A detailed overview from the law firm Michael Best warns that AI-generated transcripts can become discoverable evidence and that deleting them after a litigation hold attaches may itself violate preservation duties.

The Karta Legal analysis of AI meeting tools and eDiscovery frames the core problem: what was once ephemeral dialogue is now permanent, indexed, and easily shared — turning casual chatter into evidence and internal strategy into discoverable material. For regulated industries, SteelEye's record-keeping guide for SEC-registered advisers notes that even if AI transcripts aren't explicitly required by SEC rules, existing transcripts can be subpoenaed, and the SEC may consider summaries as "written communications regarding investment advice." We go deeper on this problem in our guide to AI meeting transcripts as discoverable evidence.

What this means for different buyer roles

For attorneys evaluating Zoom AI Companion for privileged client calls

• Document which third-party model provider (OpenAI, Anthropic, Nemotron) will see the transcript for each AI Companion feature you enable.
• Open a Zoom support ticket to turn on the Zero Data Retention option before using AI Companion on any privileged call.
• Set an institutional retention policy for Meeting Summaries — Zoom's default is "until manually deleted," which likely conflicts with your records-retention schedule.
• Follow ABA Formal Opinion 512's confidentiality framework and obtain informed client consent in the engagement letter, with specifics, not boilerplate.

For IT and compliance teams managing Microsoft Copilot in Teams

• Audit the Microsoft 365 roadmap item RM558933: the "AI-generated meeting archive" can retain insights beyond your transcript retention policy.
• Understand that turning off transcription does not stop Copilot prompts and responses from being retained in the Microsoft 365 compliance substrate.
• Apply a specific Purview retention policy to Copilot interactions, separate from your general Teams policy.
• Use Teams Premium sensitivity labels to force transcription off for sensitive categories of meetings.

For executives and boards handling market-moving conversations

• Default to no bot in the room for board sessions; see our analysis of AI notetakers in board meetings and privilege risk.
• If capture is needed, prefer an on-device tool that produces no vendor-side artifact.
• Separate verbatim transcripts (short retention) from approved minutes (long retention).

How Basil AI solves this: on-device transcription with no cloud, no retention, no third-party models

Basil AI takes a different architectural approach that answers the AI Companion privacy question structurally rather than contractually. Instead of sending audio to a cloud service that promises not to train on it, Basil runs transcription and summarization entirely on your Mac or iPhone using Apple's on-device speech recognition and the Apple Neural Engine. Nothing is uploaded. There is no vendor server that holds your transcript. There is no third-party model provider in the data path — because there is no network call.

That design flips every row in the comparison table above:

• Training risk: impossible — content never leaves the device, so no vendor could train on it even if they wanted to.
• Retention risk: you control it. Delete a recording and it is gone, locally, immediately. No 30-day debugging window, no model-provider trust-and-safety retention, no Purview substrate.
• Subpoena risk: there is nothing a vendor can produce in response to a third-party subpoena, because the vendor has nothing. The record exists only where you keep it.
• Discovery: you still have the same ESI obligations once litigation attaches — but the artifact lives in your control, with the retention window you choose.

For the technical details on how Apple's Neural Engine makes this possible, see Apple's privacy overview and the Speech framework documentation. For a deeper look at how this architecture compares to cloud notetakers, read our AI meeting notetaker comparison guide and our breakdown of what compliant AI meeting notes actually means.

The honest trade-offs

An on-device notetaker is not a drop-in replacement for every Zoom AI Companion or Copilot workflow. Three honest caveats:

• No tenant-wide search. Copilot's agentic retrieval across Google Drive, OneDrive, and Salesforce — described in this 2026 Zoom AI Companion review — is a legitimate productivity story. On-device tools don't do that; they capture one meeting at a time.
• No in-meeting "catch me up" Q&A. Zoom AI Companion's in-meeting questions feature requires the cloud. On-device tools generate the summary after the meeting ends.
• You still own the ESI. If you record a meeting locally and litigation attaches, you must preserve it — the on-device model moves the artifact into your control, it doesn't make the preservation duty disappear.

For regulated workflows — attorney-client calls, healthcare visits, financial-advisor conversations, M&A discussions — the trade-offs land firmly in favor of the on-device approach. For a general all-hands with no sensitive content, Copilot or AI Companion with proper retention controls may be perfectly appropriate.

Bottom line for buyers

The honest, buyer-fair answer to the target question — "AI Companion privacy, security, meeting data, not used to train models, retention" — is this:

1. Training: Both Zoom AI Companion and Microsoft Copilot say they do not train on your meeting content. Credible so far.
2. Retention: Both retain meaningful artifacts. Zoom summaries persist until deleted; Copilot prompts live in the Purview substrate even without transcription.
3. Third-party exposure: Zoom's federated architecture puts OpenAI and Anthropic in the data path with up to 30-day trust-and-safety retention.
4. Discovery: Both are ESI. Both are subpoenaable. The ABA's Formal Opinion 512 requires lawyers to evaluate every one of these risks before use.
5. On-device alternative: Tools like Basil AI eliminate the retention, third-party-model, and vendor-subpoena vectors entirely by never transmitting audio off the device.

Capture meetings without the retention headache

Basil AI runs 100% on your Mac and iPhone. No cloud, no third-party models, no vendor retention — just your meeting, your notes, your control.