WORK SYSTEMS
A buyer-side checklist for tracing meeting data, testing admin controls, reviewing subprocessors, setting retention, and running a low-risk pilot before an AI notetaker reaches client calls.
Read the related Korean article →A vendor's ‘no model training’ statement answers only one question. Approval should depend on the complete meeting-data lifecycle: what the tool captures, where each artefact goes, who can share it, how long it remains, what connectors export, and whether an administrator can prove deletion and offboarding.
An AI notetaker can process far more than the final summary. Depending on the product and settings, the pipeline may include audio or video, speaker-labelled transcripts, participant identities, calendar metadata, chat messages, screen content, prompts, action items and copies sent to a CRM or another AI service.
That is why a security badge or a single ‘we do not train on your data’ sentence is not an approval decision. Buyers need a lifecycle view that identifies each artefact, storage location, subprocessor, access path, retention clock and deletion method.
Require evidence for the exact plan and configuration you intend to buy. A help-centre answer, a sales statement and a contract clause are not interchangeable; use the strongest available evidence and record any gap as an open condition rather than an assumed pass.
| Control | Question for the vendor | Evidence to retain |
|---|---|---|
| 1. Capture scope | Which of audio, video, transcript, chat, screen content and metadata are collected or stored? | Data-flow diagram and feature settings |
| 2. Storage | Where are primary data and backups stored, and can the region be contractually fixed? | Contract or security document naming regions |
| 3. Subprocessors | Which transcription and model providers receive each data type, and how are changes announced? | Current subprocessor list and notice terms |
| 4. Model training | Can the vendor or any subprocessor train or improve models with recordings, transcripts, summaries or feedback? | Contractual restriction, including feedback scope |
| 5. Identity and access | Can admins enforce SSO, MFA, roles, domain capture and least-privilege access? | Locked admin policy and access test |
| 6. Joining and sharing | Can admins block automatic joining and external or public-link sharing by default? | Workspace policy and negative test |
| 7. Retention and deletion | When is each artefact removed from active systems, trash and backups? | Retention schedule and deletion test result |
| 8. Export and exit | Can admins export records in usable formats and reassign or remove data during offboarding? | Bulk export and deprovisioning test |
| 9. Auditability | Who can view, share, export or delete meeting records, and how long are audit events retained? | Audit log sample and retention statement |
| 10. Incident response | What are the notification timeline, investigation support and contractual remedies after an incident? | DPA, security addendum and response contacts |
The examples below were rechecked against public vendor documentation on August 29, 2026. They illustrate available controls and stated policies; they do not certify a product as safe, prove that every control is included in your plan, or replace the signed agreement and configuration used by your organisation.
| Vendor documentation | Publicly documented control or policy | Buyer follow-up |
|---|---|---|
| Microsoft 365 Copilot | Prompts, responses and Microsoft Graph data are not used to train foundation models; access follows existing Microsoft 365 permissions and stored interactions can be governed with Purview | Check meeting recording ownership, guest access, retention labels and any optional third-party models or agents |
| Zoom AI Companion | Meeting Summary can use speech-to-text, optional screen OCR and meeting chat; admins can restrict external sharing, lock settings, show a notice and auto-delete summaries | Confirm which models process each enabled feature and separately test transcript, recording, summary and feedback handling |
| Otter Enterprise | Admins can remove stored audio, set custom retention, lock auto-join behaviour and block Notetaker from meetings whose titles contain selected keywords | Confirm the purchased plan, storage region, export coverage, offboarding and how deletion reaches backups |
| Fireflies | The vendor states that core meeting content is not used for AI training and documents retention and access controls; its MCP guidance says external connectors fall under the external platform's terms | Map every CRM, automation and AI connector because the core-service promise may stop at the integration boundary |
The safer default is to disable organisation-wide auto-join and permit recording by meeting class. Low-risk internal operations can use an approved configuration, while customer-confidential work, pricing negotiations, product roadmaps and employee discussions need tighter sharing and case-by-case approval.
Highly sensitive meetings—such as legal advice, disciplinary matters, credentials, health information or regulated transactions—should normally stay outside automatic capture unless specialist review has approved the exact use case. A convenient summary is not worth creating an uncontrolled copy of the most sensitive conversation in the business.
| Meeting class | Default route | Examples |
|---|---|---|
| Low risk | Approved tool and standard retention | Routine internal operations, public-information training |
| Medium risk | Meeting-level approval and restricted sharing | Client calls, pricing, non-public roadmap discussions |
| High risk | No automatic recording without specialist approval | HR cases, legal advice, health or credential data, sensitive board matters |
Recording and privacy requirements vary by country, industry, contract and participant location. As a practical baseline, tell participants before capture starts why the meeting is being recorded or transcribed, what outputs will be created, who will receive them, how long they will be kept and what alternative is available if someone objects.
A notice is not a substitute for identifying the appropriate legal basis or obtaining consent where required. It also creates risk if it promises 30-day deletion while a connector, backup or downstream workspace keeps the transcript longer.
Run the pilot on five low-risk internal meetings, not on a customer call or an HR case. Include a failed-path test: block the bot from one meeting, attempt an external share, deactivate a test user, export the record, then delete the audio, transcript and summary and search for each artefact again.
Do not score only summary accuracy or time saved. Expansion should require 100% recorded notices, zero unintended external shares, successful enforcement of join and sharing policies, a documented owner for administration, and evidence that export, offboarding and deletion behaved as promised.
| Pilot gate | Pass condition |
|---|---|
| Notice | All five meetings retain evidence of pre-meeting and in-meeting notice |
| Access | No public link or unauthorised external recipient can open a record |
| Policy | A blocked meeting remains unrecorded and users cannot override locked controls |
| Offboarding | A deactivated user's records are reassigned, retained or deleted according to policy |
| Exit | The team can export required records and verify deletion of each artefact |
| Ownership | A named administrator owns quarterly control and subprocessor reviews |
AI meeting notes can remove low-value administration, but the demo rarely shows how data behaves after the call ends. A defensible buying decision ties every vendor promise to contract language, an administrator setting or a reproducible test.
Keep the completed 10-question scorecard with the approval record and reopen it when the vendor changes models, subprocessors, connectors or retention features. If a material answer remains unknown, limit the use case rather than treating the unknown as a pass.
Use the NIST Generative AI Profile to document ownership, test evidence, unresolved risks and the conditions attached to rollout.
Open the NIST GenAI Profile ↗Official sources checked as of the publication date. Products and policies can change.
Editorial note: This buyer checklist was reviewed against public vendor and regulator documentation on August 29, 2026. It does not claim hands-on testing of the products, does not certify any vendor or plan, and is general operational guidance rather than legal or compliance advice. Requirements vary by jurisdiction, sector, contract and configuration.