
Before approving an AI meeting assistant, check 12 security criteria, then run 10 operational controls around the tool. Ask vendors for documents, not claims: a certificate, a signed DPA, a settings screenshot. The matrix and controls below are vendor-neutral, and each row shows how Noota answers it.
The 12 questions to ask every AI note taker vendor
- Where are recordings, transcripts and AI processing hosted, and who are the sub-processors?
- Is customer audio, transcript or summary used to train or fine-tune any model?
- Which independent certifications do you hold, and can we read the certificate and report?
- Is audio captured by a meeting bot, a desktop app, or both, selectable per team?
- Who sets retention, can we keep text only, and how fast is deletion?
- Do you provide an Article 28 GDPR DPA with sub-processor terms and records of processing?
- Which identity providers do you support, and is de-provisioning automatic and role-based?
- Are logins, views, exports and permission changes logged and exportable to a SIEM?
- What can we export, in which formats, and through which API?
- Does the tool score sentiment, stress, tone or engagement of participants?
- Can IT detect and consolidate accounts created with corporate emails, without viral invitations?
- What are your latest penetration test results, recovery plans, uptime commitments and status page?
Meetings hold some of the most sensitive speech in a company: HR reviews, customer negotiations, board prep, incident calls. An AI note taker records, transcribes and summarizes all of it, so it deserves the same review as any system that stores personal and confidential data. This guide is written for CISOs, IT directors and DPOs who need a rubric they can paste into a vendor questionnaire. Last verified: 16 November 2026. Noota is an AI meeting assistant hosted in the EU and is used here as the worked example in the last column.
What should IT check before approving an AI note taker? The 12-point security matrix
Use this table in security reviews and procurement. The last column shows where Noota publishes its answer; where a point is not documented publicly, it says so rather than guessing.
| Criterion | What to ask the vendor | What a good answer looks like | Where Noota answers it |
|---|---|---|---|
| 1. Hosting and sub-processors | Where are recordings, transcripts and AI processing located, and who are the sub-processors? | Named regions, a public sub-processor list, and contractual limits on transfers outside the EU. | Standard hosting is EU only: Google Cloud regions in Belgium and the Netherlands, plus MongoDB Atlas EU. AI models run through Vertex AI (Claude, Gemini, Mistral). Eight sub-processors are listed in the Trust Center. See the security page. |
| 2. Model training | Is customer audio, transcript or summary used to train or fine-tune any model? | A written commitment that applies to every plan, including free ones, and to sub-processors. | Customer data is never used to train Noota's or anyone else's models, on any plan including Free. |
| 3. Certifications | Which independent certifications do you hold, and can I read the certificate and report? | ISO/IEC 27001 with a stated scope and audit date, plus a SOC 2 Type II report, both available under NDA. | ISO/IEC 27001 certified in July 2026 with zero non-conformities at both audit stages. SOC 2 Type II certified. Certificate and report on request in the Trust Center. |
| 4. Capture mode | How is audio captured: a meeting bot, a desktop app, or both? Can we choose per team? | Both options, so IT can avoid bots in sensitive contexts and still cover every platform. | Noota records with or without a bot. Bot-free recording uses the desktop app on Mac and Windows, from the Pro plan; the Free plan uses the Noota bot. |
| 5. Retention and deletion | Who sets retention? Can we keep text and delete audio? How fast is deletion on request or account closure? Does deletion also reach backups and snapshots? | Customer-controlled retention, a text-only option, and a stated deletion delay. | Retention is a customer-controlled setting. A text-only option keeps transcript and summary without audio or video. Deleted recordings and account data are permanently removed within 24 hours. Backup handling is not described on the public security page: ask Noota for details. Contact security@noota.io. |
| 6. DPA | Do you provide an Article 28 GDPR DPA with sub-processor terms and records of processing? | A DPA you can sign, a sub-processor list, and support for data subject requests. | A DPA is available to customers, and a custom DPA is included in Enterprise contracts. The compliance register covers the DPA, records of processing and the sub-processor list. Request the DPA at dpo@noota.io. |
| 7. SSO and access control | Which identity providers do you support? Is provisioning and de-provisioning automatic? Are roles enforced? | SSO with your IdP, automatic de-provisioning at offboarding, role-based access and private-by-default content. | SSO with Microsoft Entra ID (Azure AD) or SAML, automatic provisioning and de-provisioning, and role-based access control. Workspaces are private by default. SSO is listed on the Enterprise plan. |
| 8. Audit logs | Are logins, views, exports and permission changes logged in a tamper-evident way, and can we export them to a SIEM? | Exportable logs with user, time, action and resource, with a stated retention period. | Not listed on the public security page: ask Noota for details. Contact security@noota.io. |
| 9. Data export | What can we export, in which formats, and through which API? | Self-serve export, documented formats, and an API so records can move to your own archive. | Exports are yours to take. API and webhooks are on the Business plan. Formats are not published: ask Noota for details. Contact security@noota.io. |
| 10. Emotion scoring | Does the tool score sentiment, stress, tone or engagement of participants? | No emotion inference on employees, stated in writing. | Noota has no emotion, sentiment or engagement scoring. It works from the transcript, the summary and the actions. See the Read AI comparison. |
| 11. Shadow AI controls | Can IT detect and consolidate accounts created with corporate emails? Does the product spread through invitations or recap links? | Domain-level controls, forced SSO for your domain, and no viral invitation mechanics. | Domain capture and shadow-account controls are not listed on the public security page: ask Noota for details. Contact security@noota.io. |
| 12. Incident record and availability | When were your last penetration tests, how are findings closed, and what are your recovery plans and uptime commitments? | Recent independent tests with verified remediation, tested recovery, a status page and a written SLA. | Two independent penetration tests in the last audit cycle, remediation verified at re-test, no open findings. Disaster recovery plans exist for standard and sovereign environments, and a point-in-time restore of the production database was tested in July 2026. A status page and an uptime SLA are not listed on the public security page: ask Noota for details. Contact security@noota.io. |
Need Noota's evidence for your review? Request the ISO 27001 certificate and the SOC 2 report in the Trust Center, or book a security review.
10 operational controls before and after purchase
A vendor can pass the matrix and still create risk if the rollout is careless. These ten controls sit on your side of the contract.
1. Set the legal basis, the notice and a sensitive-meeting blacklist
Recordings and transcripts are personal data under GDPR, so decide the legal basis before the first pilot. Tell participants before recording starts and give them a clear way to object. Write a list of meeting types where AI recording is not allowed, for example disciplinary hearings, works council talks, M&A deal rooms and security incident reviews. Our GDPR and AI guide covers the legal basis question in more depth. If you allow recording without a bot, no bot appears in the participant list to signal it, so your acceptable-use policy should say who may start a recording and require a spoken notice before capture wherever the law requires one.
2. Govern access through your identity provider
Connect the tool to your IdP and make SSO the only way in. Require automatic de-provisioning so a leaver loses access the day the directory account is closed. Keep workspaces private by default and share by exception, so a sales team does not see legal or finance meetings.
3. Check encryption and the hosting model
Ask for AES-256 at rest and TLS 1.2 or 1.3 in transit, and ask who holds the keys. Then separate two hosting models. Standard EU hosting keeps data in EU regions of a major cloud provider. Sovereign hosting narrows the perimeter further: provider, models and operator. Standard EU hosting still uses providers headquartered in the United States (for example Google Cloud, and the model providers reached through Vertex AI), so ask your legal team how US legal reach, such as the US Cloud Act, applies to your data. This is one reason sovereign hosting exists. Choose the one that matches the sensitivity of your meetings, and write the choice into your records of processing.
4. Minimise retention
Do not keep audio longer than you need it. A common pattern is to keep the transcript and summary and drop the audio and video once the text is checked. Choose a retention period you can justify and document, apply it per workspace, and test that deletion requests actually remove the data on the vendor side.
5. Forbid model training in the contract
A policy page can change; a contract clause is harder to change. Require a written no-training commitment that covers every plan and every sub-processor that touches the content. Ask how the AI providers behind the tool handle retention of prompts and outputs.
6. Rule out emotion inference
Article 5(1)(f) of the EU AI Act prohibits AI systems that infer the emotions of people in the workplace, except for medical or safety reasons. It has applied since 2 February 2025, and fines for prohibited practices reach EUR 35 million or 7% of worldwide annual turnover, whichever is higher. Ask every vendor in writing whether any feature scores tone, sentiment, stress or engagement, and switch such features off. For the wider obligations, see our EU AI Act guide.
7. Govern bots and shadow note takers
Decide where bots are allowed, who may invite them, and what happens when an unknown one appears. The next section covers the technical settings. Tell employees which tool is approved and why, because an official option is the best defence against unofficial ones.
8. Ask for proof, not logos
Request the ISO/IEC 27001 certificate and check its scope and date. Request the SOC 2 Type II report and read the auditor's opinion and exceptions. Ask for penetration test summaries and how findings were closed. A trust centre that hands these over on request is a good sign.
9. Classify the outputs
Summaries and transcripts become company records. Apply your confidentiality labels to them, decide which system is the master copy, and route notes into your document management or knowledge base through the API or exports. Check who can re-share a note outside the company.
10. Pilot first and plan the exit
Run a pilot with 20 to 50 users in two or three teams. During the pilot, test de-provisioning, a deletion request and a full export. Confirm what happens to data at contract end, how long deletion takes and what you can take with you. Only then widen the rollout.
How to govern Teams meeting AI note takers and block shadow bots
IT teams report a recurring pattern. An external attendee brings a note-taking bot to a call. After the meeting, the service emails a recap link to every invitee. An employee clicks it and signs in with Microsoft or Google, which grants calendar permissions. The bot then joins that employee's later meetings, and the cycle repeats with their guests. No one in IT approved the tool, yet it now holds internal conversations.
Three measures close the gap.
Set the Teams policy. Microsoft Teams admins control external meeting bots with the ExternalBotAccessMode parameter of the meeting policy (Set-CsTeamsMeetingPolicy). The values are AllowBots, RequireApprovalWhenDetected and BlockDetectedBots. By default, RequireApprovalWhenDetected applies, so detected bots wait in the lobby until someone admits them. The policy applies to external third-party bots only. Details are in Microsoft's documentation on external bots.
Restrict user consent in Microsoft Entra ID. Admins can limit user consent to unverified third-party apps, so an employee cannot hand calendar access to an unknown vendor with one click. Route consent requests to an admin review instead.
Offer an approved, bot-free option. When a team needs notes and has no sanctioned tool, it will find one. A desktop app that records locally, such as the Noota desktop app on Mac and Windows, avoids sending a bot into the call at all. Our guide on recording meetings without bots explains when this fits. If you choose a desktop app, check how it is deployed (signed installers, silent installation through your device-management tool, for example Intune or Jamf), which permissions it needs on the laptop, and whether audio is cached locally before upload and how that cache is protected.
How does Noota map to this checklist?
On the public record, Noota documents answers to most of the matrix: EU hosting on Google Cloud in Belgium and the Netherlands, a no-training commitment on every plan, ISO/IEC 27001 (July 2026) and SOC 2 Type II, a DPA, SSO with Entra ID or SAML, customer-controlled retention, and no emotion scoring. It also publishes two independent penetration tests and tested disaster recovery. Audit logs, export formats, an uptime SLA and domain capture controls are not listed on the public security page, so put them in your questionnaire. Send these questions to security@noota.io; data-protection and DPA requests go to dpo@noota.io. Noota records Microsoft Teams, Google Meet and Zoom meetings.
For stricter requirements, the Enterprise plan offers the Sovereign France environment: Scaleway in Paris, transcription by Gladia, AI models by Mistral, a capture bot built and run by Noota, and a text-only option that keeps no audio or video. Noota describes this as an architecture compatible with SecNumCloud requirements. An admin chooses the environment per workspace. Sovereign mode uses a Noota-run bot, so it is a different choice from bot-free desktop capture. More detail is on the Enterprise page, the security page and the Trust Center, where DPO and IT teams can request the certificate and reports.
Related reading: Best AI Meeting Minutes Software Tested & Compared and Meeting Notes vs. Minutes: Key Differences Explained.
Frequently asked questions
IT teams should check twelve criteria: hosting region and sub-processors, model training, certifications, capture mode, retention and deletion, the data processing agreement, SSO and role-based access, audit logs, data export, emotion scoring, shadow AI controls, and incident record with availability. Ask for evidence on each one, such as a certificate, a signed DPA or a documented setting, not a marketing statement. The 12-point matrix in this guide gives the exact question and the answer to expect for each criterion.
Yes, an IT department can validate a bot-based tool if it can control where the bot joins, who invites it and what it stores. Check that the bot is identifiable in the participant list, that admins can block or approve it, and that its data follows the same hosting, retention and DPA rules as any other processing. If the bot cannot be governed, prefer a tool that also offers bot-free capture, such as a desktop app.
A DPO can recommend an assistant that gives a signed Article 28 DPA, a public sub-processor list, EU hosting, a written no-training commitment, configurable retention and no emotion scoring. The DPO should also see an ISO/IEC 27001 certificate and a SOC 2 Type II report, not only a logo. Noota meets these points on its public security page and Trust Center, but any shortlist should be checked against the same twelve criteria.
Standard EU hosting keeps data in EU regions of a cloud provider, while sovereign hosting also restricts the provider, the models and the operator to a tighter European perimeter. At Noota, standard hosting runs on Google Cloud in Belgium and the Netherlands, and the Sovereign France option runs on Scaleway in Paris with Mistral AI models and a Noota-run capture bot. Noota describes that architecture as compatible with SecNumCloud requirements. Standard EU hosting can still involve US-headquartered providers, which is one reason sovereign hosting exists.
Yes, in the workplace it does, with narrow exceptions. Article 5(1)(f) of the EU AI Act prohibits AI systems that infer the emotions of a person in the workplace or in education, except for medical or safety reasons, and the ban has applied since 2 February 2025. Fines for prohibited practices can reach EUR 35 million or 7% of worldwide annual turnover, whichever is higher, so ask any vendor whether it scores sentiment, stress or engagement.
Combine three measures: set the Teams external bot policy, restrict user consent to unverified apps in Microsoft Entra ID, and give employees an approved tool. In Teams, the ExternalBotAccessMode setting can allow bots, hold detected bots in the lobby for approval, or block them. Restricting consent stops an employee from granting calendar access to an unknown app with one click, and an approved assistant removes the reason to try others.
An AI meeting note taker is safe to use in business only if the vendor can show evidence for the points that matter: no training on your data, a named hosting region and sub-processor list, independent ISO/IEC 27001 and SOC 2 Type II certification, controlled retention and deletion, and IT control over bots and access. Use the 12-point matrix in this guide to test any vendor against those points.
Written byAdesh Sonawane





