A meeting assistant is useful when it creates a reviewable record of decisions and owners. It becomes surveillance when capture expands, interpretation becomes behavioural scoring, and participants lose meaningful control.
Meetings create two real problems. People forget what was decided, and nobody remembers who owns the next action.
AI meeting assistants promise to fix both by transcribing conversation, generating notes and extracting tasks. They can. They also create a new record containing people’s words, speaker identity, attendance, screen content, chat, decisions, inferred tasks, and sometimes recordings and behavioural metadata.
The question is not whether the summary is convenient. It is whether the system produces accountability with proportionate capture, clear access, human correction and defined retention.
What the major platforms actually describe
Public documentation from Google, Microsoft and Zoom shows three different data paths.
Google Meet. “Take notes for me” creates organised notes in Google Docs, can identify decisions and next steps, informs participants when note-taking is active, allows configured recipients, can be stopped for sensitive parts, and stores notes in the organiser’s Drive under the organisation’s retention policy.
Google states administrators may require explicit participant consent, but that control is off by default. Hosts can configure sharing to hosts, internal invitees or all invited guests. Invited guests are not necessarily the same as actual attendees. Google warns that summaries can be incomplete or inaccurate.
Microsoft Teams. Intelligent recap depends on a transcript and, for some features, a recording. Different outputs are stored in different locations: transcripts in Exchange and sometimes OneDrive, AI notes and tasks in participants’ Exchange mailboxes, recording markers in OneDrive or SharePoint.
Microsoft states customer data is not logged or used for model training, and that the GPT-based features run through a dedicated Azure OpenAI service inside Microsoft 365 boundaries. That commitment matters. It does not remove the need to configure access, retention, correction and appropriate use.
Zoom. Meeting Summary may use speech-to-text data, optional screen-shared content through OCR, and optional in-meeting chat. Zoom warns outputs may be incorrect, inaccurate, misleading or offensive and should be checked.
Its wider AI documentation describes access to connected sources, local files and other Zoom content for different features. Meeting Summary should be assessed as one feature inside a broader administrative environment, not as an isolated button.
Product notice is not the complete consent analysis
A banner informs participants that capture is occurring. It does not answer whether participation is genuinely optional, whether a person can join without being captured, who required the assistant, which data is collected, who receives the output, how long it is kept, whether a participant can correct it, whether the recording is lawful in the relevant jurisdiction, or whether the employer’s use is necessary and proportionate.
Legal requirements vary by jurisdiction and context. Employment relationships complicate consent because refusal may not be genuinely free. The UK ICO’s guidance on monitoring workers is the clearest practical treatment of that problem.
The operating rule should be stronger than “the platform displayed a notice.”
Create an Accountability Receipt
Before the meeting, record the purpose, the reason for using the assistant, the person who enabled it, the data captured, intended recipients, retention period, applicable policy, and the participant notification method.
During the meeting, make visible the active capture indicator, the pause and stop control, who controls the assistant, and whether recording, transcription, screenshots or chat capture are active. Allow sensitive portions to occur without capture.
After the meeting, the receipt should contain the summary owner, decisions with timestamps or source links, actions with a named human owner and due date, open disagreements, unresolved points, a correction route and a deletion date.
The machine can propose the receipt. A named person approves it.
Do not let the summary become the decision
Meeting assistants compress. Compression removes uncertainty, disagreement, qualification, conditions, who proposed what, and whether a topic was decided or merely discussed.
Require separate states: proposed, discussed, agreed, disagreed, deferred, withdrawn. Do not turn every confident sentence into a decision.
For consequential meetings, keep source timestamps or a transcript reference so participants can inspect context.
Action items need human ownership
An extracted task such as “Nikola to prepare the proposal by Friday” may be wrong because Nikola offered to review rather than prepare, Friday was suggested rather than agreed, the task depended on information from someone else, or the assistant assigned the nearest name.
Every action item should require confirmation by the meeting owner or the assigned task owner. The assistant may draft tasks. It should not silently change workload.
Draw a hard line at behavioural scoring
Do not use meeting AI to infer emotion, attention, honesty, enthusiasm, productivity, leadership potential or employee value.
The EU AI Act prohibits emotion-recognition systems in workplaces except for narrow medical or safety purposes. Even outside that prohibition, conversational data is a weak basis for consequential judgments about people.
Counting speaking time is also misleading. More words are not more contribution.
An accountability system should track decisions and actions, not manufacture psychological profiles.
Seven controls
Necessity. Does this meeting need capture? Proportionality. Could decisions and actions be recorded instead of a full transcript? Transparency. Do participants understand what is active? Access. Are recipients limited to people who need the record? Correction. Can participants challenge inaccurate notes and tasks? Retention. Is deletion scheduled rather than indefinite? Human ownership. Who approves the final record?
A practical default
For ordinary internal meetings: use notes without recording where the platform allows it, share only with actual required recipients, name the human note owner, confirm decisions and tasks before closing, retain according to purpose, and prohibit behavioural analysis.
For confidential, employment, health, legal, disciplinary or sensitive commercial discussions: reassess whether automated capture should be used at all, obtain specific advice where required, and provide a manual alternative.
Not every meeting becomes more accountable when it becomes searchable.
How we approve one
We approve meeting assistants through a feature-specific review covering capture, notice, access, storage, retention, correction, task ownership and prohibited analysis. The selected platform matters less than the operating configuration, which is the same principle as the assistant operating manual and sits inside the AI operating model.
An assistant should make it easier to answer: what did we decide, and who owns the next move?
If it also tries to answer who looked engaged enough, turn it off.
If you are rolling one of these out across a team, tell us which platform and which meetings and we will tell you what to switch off first.
Sources
Google Meet, Take notes for me.
Microsoft Teams, data, privacy and security for intelligent recap.
Zoom, how AI features handle data.
UK ICO, data protection and monitoring workers.
European Commission, AI Act overview.
Product and policy documentation checked 29 to 30 July 2026. Features and settings change. This is an operational teardown, not legal advice.
