Meeting Copilot — private suggestions while you talk

Overview

Meeting Copilot is the copilot you already know, sitting in on a call. While a session runs it transcribes what is said, notices when someone asks a question or describes a problem, and answers privately, to you only — short bullets labelled Confirmed, Likely or Hypothesis, each citing the index, a document, a learning, a component or the transcript itself. Nothing is posted into the meeting and nothing is sent on anyone's behalf: the copilot drafts text for you to read, copy or ignore.

Open it under Overview → Meeting Copilot. The page has six tabs:

A session has a life of its own: Ready → Listening → Paused → Stopped → To review → Finalized (or Discarded). Audio is streamed, transcribed and dropped; the transcript text is stored — encrypted, unless the policy turns that off — for the retention period your policy sets. What outlasts that period is what you keep at the review — and, unless the session is ephemeral, the extracted items you leave undecided, which stay with the session in plain text until it is deleted (or until the policy's insight retention, when one is set). A discarded item keeps only its title.

Where the audio comes from:

• My microphone — you.
• Meeting audio — the other participants. In Chrome or Edge on a desktop you choose A browser tab and share the tab that holds the web meeting (tick Also share tab audio in the picker). On Windows and ChromeOS you can also choose The entire screen with system audio. For a native Zoom, Teams or Webex client, and on Safari or Firefox, use the Desktop companion.
• Microphone only — your side of the conversation, which is still enough for dictated notes and your own questions.

The copilot never sees a video frame: screen sharing is used for its audio track only.

Before you start

• Starting a session, importing a transcript and pairing a companion need Record meetings and dictate (Admin, Architect, Release Manager, DevOps and Developer by default). The Policy and Health tabs and deleting other people's sessions need Manage meeting policy and sessions (Admin, Architect). See Roles and permissions.
• Your workspace needs a speech-to-text provider configured by an administrator, and at least one working AI provider for suggestions. If no speech provider answers, the page says so and offers on-device dictation instead.
• Use a desktop browser. Chrome and Edge can share a meeting tab with audio. In Safari and Firefox — and in Chrome or Edge when your workspace policy turns meeting audio off — the set-up card pre-selects Microphone only and still offers the Desktop companion (when the policy allows it). A Mac browser shares tab audio only, never the whole screen. Phones and tablets get dictation only.
• Read the consent dialog the first time, and tell the other participants that the meeting is being transcribed — the dialog has a Let participants know tick that cannot be skipped when the policy requires it, and the server refuses to start without a current acknowledgement.
• Sharing items from the review needs Contribute org knowledge, like every other contribution to Org Knowledge.

Steps

Set up a live session

1. Open Overview → Meeting Copilot and the Live session tab.
2. Under Set up a live session, give the meeting a Title, choose the Org (evidence and scoping) the conversation is about — suggestions cite that org's index and knowledge — and a Profile:
   • Fast — interim words as early as possible; suggestions about 0.7 s after a question ends. A few more interruptions.
   • Balanced (default) — a suggestion when a question ends: its first lines within about a second, the full card within about two (on a fast AI provider).
   • Accurate — speaker labels and longer turns; suggestions take a second look at the evidence.
   • Economy — audio goes up in 15-second chunks to the cheapest batch model; suggestions only when you ask.
3. Choose the sources: My microphone (system default or a named device) and Meeting audio — A browser tab, The entire screen, Desktop companion or Microphone only. Options that are not available here say why — Turned off by your workspace policy, Linux browsers share tab audio only, A Mac browser can share tab audio only; for native meeting apps use the desktop companion, Desktop only. With Desktop companion you start capture from the companion itself.
4. Tick Ephemeral — keep nothing after the review if nothing from this meeting may be stored: transcript, suggestions and insights then live in the server's memory only, never in the database, and the questions you ask during the meeting are not recorded. Finalizing the review drops them, unless a save fails so you can retry (see Stop, review and decide). Closing the review without finalizing does not: the server keeps them — the items extracted from the meeting included — until two hours after the stop at the latest; a re-extraction late in the review does not extend that.
5. Click Review consent and continue. Nothing is captured yet.

The sources you choose here are part of what you consent to. A session set up for Microphone only stays microphone-only: the server refuses meeting audio for it, even if you confirm the consent again. To include the meeting audio, set up a new session.

Acknowledge what is captured

The consent dialog shows a table of What is captured, Where it goes and how long it is Kept — your microphone and the meeting audio go to the named speech-to-text provider and are never stored; transcript text is kept for the policy's retention period, encrypted unless the policy turns that off; suggestions and your questions go to the workspace's AI provider chain with transcript excerpts as evidence, suggestions are kept for the policy's suggestion retention and your questions and answers stay in your copilot history; and the insights found in the meeting stay with the session until you discard them (at most for the policy's insight retention, when one is set) — saved facts, risks and findings become your personal notes in Org Knowledge, shared items become candidates a curator verifies before colleagues see them, saved decisions and action items stay with the meeting, and personal notes and shared candidates are kept until retired. With Ephemeral session — keep nothing after the review switched on, the transcript and the suggestions and your questions read Not stored (ephemeral), the insights are held in memory until the review and afterwards only the personal notes and the shared candidates are kept, and the switch explains: Transcript, suggestions and insights stay in memory only, never in the database. Finalizing the review drops them (unless a save fails, so you can retry); otherwise they are dropped two hours after the session stops. Only personal notes you save and items you share outlast it.

1. Read the text; the vendor named is the one your policy prefers.
2. Tick Let participants know once you have told the other people on the call (required unless your policy says otherwise).
3. Tick I understand what is captured, where it goes and how long it is kept and confirm. Your acknowledgement is recorded with the policy version; if an administrator later changes the retention, the providers or the text, you will be asked again.

Start capture

1. Click Start capture. The browser's share picker opens: choose the tab (or screen) that holds the meeting and tick Also share tab audio at the bottom. The microphone permission follows.
2. The header pill turns to Listening with the elapsed time, source icons and a Stop button. It follows you to every page, so you can keep working while the meeting runs.

If the share came back without audio — the usual cause is the unticked box — the page shows The share came without audio with a Share again button that keeps everything else running. Once the tab's audio is attached, the page tells the server — over the live connection or, when the page has fallen back to HTTPS, over HTTPS — which records the added meeting audio in the audit log and shows the meeting-audio icon next to the controls.

If you open a session set up with meeting audio in a browser that cannot share it (Safari, Firefox, or a policy that turns meeting audio off), Start capture records your microphone instead, and the page says This browser cannot share meeting audio; using the microphone.

When that session was set up for meeting audio only, the consent you gave for it does not cover your microphone, so Start capture asks again first: the consent dialog opens for the microphone only, with the note This session was set up to capture meeting audio only, and this browser cannot share meeting audio, so only your microphone can be captured here. …. Acknowledging it sets up a new session for your microphone with the same title and settings — A new session was set up for your microphone; the earlier one stays under Sessions. Click Start capture to begin. — and you click Start capture again.

If the server refuses to start or resume, the page shows its reason in its own words — for example This session is being stopped; create a new session to capture again. when the session's stop is still being completed (see Troubleshooting). After a refused Resume the capture stays on hold, so the tab does not record audio the server will not take.

One tab captures one session at a time. Resume on Sessions, or Resume here on a session page, while this tab is capturing another session is refused with Stop the current capture before resuming another session.

During the meeting

• Transcript (left) shows speaker runs — me and them when two channels are captured, numbered speakers with the Accurate profile — with interim words in italics until the provider finalises them. A passage that sounded like an instruction to a system (ignore previous instructions…) carries the badge Instruction-like speech ignored: it is shown, never acted on, and excluded from every prompt. Click a line for Copy this passage or Ask the copilot about this.
• Suggestions (centre) light up a few seconds after a question, a described problem, a request or a topic change: a thinking row (Heard a question → Retrieving evidence… → Writing a suggestion… → Drafting…) that shows the card's title and each finished bullet as plain text while the copilot is still writing — usually within about a second — then, about a second later, the full card, which replaces the draft, with a mode chip (Answer, Explain, Clarify, Next steps, Risks, Patterns, Evidence), bullets labelled Confirmed (backed by a cited fact, document, learning or component), Likely or Hypothesis, and the citations — [T#] markers jump to the moment in the transcript. As in the chat, a chip says whether the card rests on Org knowledge, General AI knowledge, Web research or a mix, and the evidence meter shows how well supported it is. A card can appear first and be revised by a background web check: it shows Checking the web… meanwhile and a small Updated with web sources tag after, with the web sources ([W#]) under Evidence; if the check cannot finish it says Web check unavailable and the card stands. Each card offers Deeper analysis (the full copilot on the same question, with the transcript as evidence), Copy as text, Dismiss, Helpful and Not helpful. A card that arrived after the conversation moved on is greyed and marked Late; one replaced by a newer topic is Superseded. Nothing here is sent anywhere.
• Ask the copilot at the bottom of the suggestions pane takes any question — pick a mode or leave Auto — and shows the stages Gathering evidence, Planning, Running lookups, Writing the answer as they happen. While the answer is written its words appear as they arrive, as plain text without links; the finished answer, with its citations and links, then replaces them. If the AI provider fails part-way, the text starts over with the next provider. Save as learning under an answer drafts a learning for your team, exactly as in the chat. For follow-ups the copilot remembers your last few questions and answers in this session — yours only: on a session shared with the workspace, a colleague's question never carries yours into its prompt, nor yours theirs — and forgets them when the session stops (a question asked after the stop is answered without that memory). Each question and its answer is also recorded as a copilot run under your name, as a question asked in the chat is — except in an ephemeral session, where nothing is recorded, so those answers offer no Save as learning.
• Where a card comes from. The copilot looks in your org first. When the org evidence does not cover a Salesforce best-practice, how-to, limit or concept question, a card may carry general guidance bullets — general Salesforce knowledge, with no marker, set apart under General knowledge — not from your org and never presented as a fact about your org. Each card and answer shows its source (org, general guidance, web, or org plus either) and a confidence label with a one-line reason, worked out from the evidence it cites; a card resting only on general knowledge is never rated above medium. In interview mode no general bullets are written: your experience comes only from what you attached.
• Web research (only when the workspace policy turns on Allow web research, off by default, and your administrator has configured a web search provider): a thin card on a question that needs current or specific facts is shown at once, then updated in place a few seconds later with up to two Web bullets citing [W1], [W2] and their links. An in-meeting question's answer is updated the same way. Web research never delays the first card or answer. What leaves MetaLens360 is a short keyword search phrase — never transcript text, participants' or your name, the org's name, ids or record data — and nothing is stored with the provider by MetaLens360.
• Context (right) keeps the agenda, participants, topics, constraints, decisions, open questions, action items and unresolved points. Edit the agenda, participants and constraints yourself — the copilot never overwrites what you typed — and watch it add topics and decisions every few minutes. The Ask icon next to a section asks about it.
• Pause stops audio without ending the session; Resume continues on the same transcript. Stop is one action everywhere — the pane, the header pill, the companion's shortcut — and asks Stop capture already? only during the first ten seconds.
• If you reload the page, the session survives but a browser cannot resume a screen share without a click: the pill shows Share again, and the server pauses the session after 30 seconds of silence from your browser until you do.
• If the server restarted while nobody was connected, a session it no longer holds reads Paused everywhere — the list, the session page and the live view — until its owner resumes it.

Interview mode

Use Interview mode when you are the one being asked: a job interview, a client panel, a certification review. Start listening once and every interviewer question gets a card of short, first-person talking points you can say in your own words — drawn only from material you attach, never invented.

1. Put your material in Org Knowledge first: Org Knowledge → Add a note (paste your resume, the job description, notes on two or three projects) or Upload a document.
2. In Live session, set Assist mode to Interview and pick those sources under Your sources (resume, job description, notes) — up to ten. Only sources of your own workspace can be attached.
3. Review the consent, Start capture, and share the meeting tab so the interviewer is heard.

What changes in Interview mode:

• Questions are caught even when they are not questions. Tell me about…, Walk me through…, Describe a time…, How would you…, Why did you…, What was your role… trigger a card, and the bar for the other side of the call is slightly lower. Your own speech is held to the usual bar, so answering does not set off cards.
• Answers come from your sources. Your attached documents are searched first and cited as [S#]; a bullet about your experience is Confirmed only when one of them says it. When they say nothing relevant, the card says what is missing instead of making something up — for example No project with event streaming in your notes — bridge from a related one.
• Follow-ups build on what came before. The last five answers of the session are given back to the copilot, so "And what would you do differently?" continues the story you just told.
• A ten-second limit. If the AI provider has not answered within ten seconds, the thinking row says No AI answer within 10 s — showing any evidence found, and you get the most relevant passages from your sources instead.

You can switch between Meeting mode and Interview mode, and attach or remove sources under Your sources in the Context column, at any time during the session. Everything stays private to you: nothing is said, typed or sent into the call.

Example: Alex Example attaches Resume — Alex Example and Job description — Solution Architect, Contoso. The interviewer asks "Walk me through a migration you led." A card titled Leading the approval-process migration suggests: "At Example Corp I led the move of 40 approval processes to Flow in 2023 [S1]", "I ran it in three waves, retiring the old rules after each [S1]", and a Hypothesis bullet: "No figures on time saved in your notes — say what you measured, if you did."

Stop, review and decide

1. Click Stop. The transcript is sealed, the session moves to Stopped, and extraction runs over it (about a minute for a long meeting). The page opens the review automatically; you can also reach it later from Sessions or the session page's Review tab.
2. Each extracted item — Fact, Decision, Open question, Action item, Pattern, Risk, Learning, Knowledge gap — shows its confidence, the passage it came from (In transcript) and a novelty badge: new, duplicate of … (already known — safe to discard unless the wording here is better), updates … or contradicts … (disagrees with a verified item; one of the two is wrong or outdated). A Stale link badge means the related item was retired or superseded since.
3. Decide per item: Save (a fact, pattern, risk, learning, open question or knowledge gap becomes a personal learning only you see, retrievable in your own questions; a decision or action item creates no learning and is only kept with the meeting), Share (a candidate learning for This org only or A knowledge domain — decisions and action items too, with their owner, due date and context — that a curator verifies before colleagues see it) or Discard. A discarded item's statement, context, owner and due date are deleted; only its title stays, so its card shows the title alone and Re-extract does not offer the same item again, even reworded. Colleagues reading a shared session never see discarded items. Edit before saving lets you fix the title, the statement, the owner or the due date first. Save the rest privately and Discard the rest handle whatever you have not decided; undecided items stay pending — nothing is published or discarded implicitly.
4. Decide about the Transcript: Keep it until the retention date (greyed out for an ephemeral session, see below), Save it to Org Knowledge as a meeting document (choose Who may see it: this org, a domain or the whole workspace) or Delete it now. The Impact line says exactly what each choice deletes: deleting the transcript also deletes the session's suggestions and the decisions, action items, open questions and unresolved points captured from speech in the session context. Learnings and documents you already published are never touched by a transcript deletion.
5. Use Keep the session summary in the history to keep or clear the generated summary, click Finalize, read Finalize this session? (counts of saved, shared and discarded items, the transcript and summary decisions) and confirm. Finalize is the only call that publishes anything, and it is safe to repeat: items already decided are skipped.

For an ephemeral session the review carries the chip Ephemeral — unsaved items are dropped at finalize, or 2 h after the session stopped, and Keep it cannot be chosen for the transcript: it is greyed out and stays selected only as the default, so the options you can pick are Save it to Org Knowledge as a meeting document and Delete it now. A finalize that completes without a warning also drops everything the server held in memory for the session, whatever the transcript choice: Done. then says Transcript deleted. unless you saved the transcript to Org Knowledge, and a decision or action item marked Save is not kept anywhere — Share it to keep it. Facts, patterns and the other kinds marked Save still become personal learnings. If the transcript was not saved to Org Knowledge and nothing else was kept — nothing saved or shared (an ephemeral session stores no summary) — and the finalize reported no warning, the session becomes Discarded. When the finalize reports a warning (below), the server keeps what it holds in memory instead, so that a repeated finalize can still publish it, and Done. says Transcript kept. (unless you chose Delete it now); it is still dropped two hours after the stop.

If part of the finalize could not be done, a warning below Done. says so. The transcript was not saved to Org Knowledge. gives the reason — for example This transcript is already saved in Org Knowledge under a different scope. — and, when there is a retention date, It is still kept with the session until that date; the rest of the finalize still completes. Items that could not be published are listed as N item(s) could not be published: with the reasons.

After a finalize that reported a warning the review stays open instead of locking, and a warning says Some items could not be saved. They are still in the Review tab: finalize again to retry. on the session page, or Some items could not be saved. They are still in the review: finalize again to retry. on the Live session tab. Such a session is never marked Discarded, so the retry is always accepted. The items that failed are still pending: click Finalize again on the same screen to retry them — items already decided are skipped. Retry before you leave or reload the page; after that the page treats the session as finalized and no longer offers Finalize. A finalize without a warning locks the review.

Notes and export

On the session page's Notes & export tab, generate a Summary (what was discussed, decided and left open — one screen), a KT note, a Design note (options discussed, the decision, its rationale and consequences), a Follow-up message to the participants or an Action items table. Every sentence that states something from the meeting ends with a [T#] marker into the transcript; Regenerate and Copy markdown are there too. Export downloads the session as Markdown (.md), Word (.docx) or JSON (.json) — header, context, insights with their decisions, the kept summary, and the transcript only while it is still retained — never for an ephemeral session, whoever exports it. For an ephemeral session, notes and exports made during the review window include the decisions, action items and open questions heard in the meeting, read from memory and never stored; once the session has left memory — at a finalize that reports no warning, or two hours after the stop — only what you typed into the context (agenda, participants, constraints) is left.

Import a past meeting

1. On Sessions, click Import a meeting.
2. Drop a file: WebVTT (.vtt, what Zoom, Teams, Google Meet and most caption tools export), SubRip (.srt), Zoom or Teams JSON exports, plain text or Markdown ([00:12:34] Name: text lines or bare paragraphs), or a Word document — up to 15 MB, and a Word document's text may be at most 50 MB once uncompressed. Word files that are password-protected, damaged, or saved in a form that cannot be read (ZIP64, an unsupported compression method, more than 5000 internal parts, extra data in front of the document) are refused with a message that says why — a password-protected file, or an old .doc renamed to .docx, with This file is password-protected or in the old .doc/.xls format, which cannot be read. Remove the password (or save it as .docx/.xlsx) and upload again. Remove the password or re-save the file from Word — or save it as text — and import again. The first 20 000 transcript lines of a very long file are imported, with a warning. A recording (.wav, .m4a, .webm, .ogg, up to 25 MB) is accepted only when the policy allows Recording import; it is transcribed by the batch provider and the file is dropped.
3. Add a Title, the Org, When it took place and, optionally, the Participants, then click Import and extract. The session lands straight in review, with the same decisions as a live meeting. Imported text is treated exactly like spoken text: instruction-like lines are flagged and ignored.

Find something said in an earlier meeting

Click Search transcripts on Sessions. Titles, summaries, topics, constraints, decisions, action items, suggestion titles and insights of your sessions and the workspace-visible ones are searched at once — the 500 most recent sessions you can see, with at most 50 results. Tick Inside transcripts to also scan the text of the most recent retained transcripts you can see, newest first: at most 20 000 transcript lines in all, the first 2 000 of each, from the last 50 sessions. A transcript hit opens the session with that line highlighted.

• Searches are limited per person: 60 a minute, and 6 a minute with Inside transcripts (counted separately). Over the limit the box says Too many searches — try again in N seconds.
• A session whose transcript was deleted (retention or Delete transcript) no longer matches on its decisions, action items, open questions, unresolved points or the constraints the copilot added from speech; the constraints you typed still match. An ephemeral session matches only on its title and what you typed into its context (agenda, participants, constraints) — it stores no summary. Discarded items never match.

The Sessions list shows the 100 most recent sessions. When there are more, a line above the table says Showing the 100 most recent sessions. Filter by status or org, or search the transcripts, to find older ones.

The session page

Open any session from Sessions (/meetings/:id). Overview shows when, duration, segments, suggestions, insights, retention and which transcription providers were used; Transcript, Suggestions and Insights show what the meeting produced (Suggestions to you only: a colleague reading a shared session sees the tab without a count and the note Suggestions are private to the person who captured this session.); the Summary box on Overview says No summary yet. Generate one under Notes & export; the review lets you keep or drop it. while the session has none (An ephemeral session keeps no summary. for an ephemeral one); Review appears while decisions are pending; Notes & export is described above. Share with workspace lets colleagues read the session — its transcript, context and insights, never its suggestions — and Make private takes it back. Delete transcript or Delete session show the counts of what will go before anything is deleted — the transcript option also deletes the suggestions and the decisions, action items, open questions and unresolved points captured from speech in the session context; published learnings and documents stay.

Adjust your settings and the workspace policy

• Settings: Language, Latency profile, the suggestion kinds you want, Dictation in the chat (see Dictate instead of typing), and This device — what this browser can capture, with the reason when it cannot.
• Policy (manage): Transcripts kept (days), Suggestions kept (days), Insights kept (days, blank = forever), Encrypt transcripts and suggestions at rest, Ephemeral by default; the Speech-to-text providers — where audio goes table with Allowed ticks and a Preferred provider — it lists every configured provider, so one left out of a saved allow-list shows unticked and can be ticked again; ticking all of them allows every configured provider, including ones configured later (a policy with no provider allowed cannot be saved — tick at least one, or click Allow every configured provider); What people may capture (meeting audio, desktop companion, on-device and cloud browser speech, recording import, screenshot questions); Consent — Require a participant-notice tick before capture, Custom consent text; Suggestions — automatic or manual, and AI calls per session per hour; Allow web research (off by default; turning it on or off changes the consent text, so people are asked to acknowledge again). Saving is audited, and any change to what people were told bumps the consent version. Run retention sweep now deletes expired transcripts and suggestions of this workspace (up to 50 due sessions a run) without waiting for the nightly run. A changed Transcripts kept (days) applies to sessions captured afterwards; a session that already stopped keeps the retention date it got then.
• Health (manage): sessions, suggestions, cost estimate, socket connects, the targets (First interim, Final segment, Suggestion card, First lines, Ask: first words), and per AI provider the suggestion card and first-lines times — counted over individual cards, leaving out evidence-only and late cards.

Example

Priya, an architect at Acme, joins a Google Meet with the CPQ team in Chrome. She opens Overview → Meeting Copilot, titles the session Quote calculation incident, picks acme-prod, keeps Balanced, acknowledges the consent and tells the room the call is transcribed. Start capture, the Meet tab, Also share tab audio. Twelve minutes in, Sam says: "Why do quotes with more than two hundred lines time out on calculate?" About a second later the card's first lines appear for Priya only, and the full card is there within about two — Answer, two Confirmed bullets citing [L1] (a verified learning about the QCP timeout) and [C1] (the QuoteLineTriggerHandler class), one Hypothesis about a missing index. She reads it out in her own words. After Stop, the review lists a Decision (move the recalculation into a queueable), an Action item for Lee due Friday, a Learning marked duplicate of the existing one, and a Risk. She shares the decision with acme-prod, saves the action item (it stays with the meeting), discards the duplicate, keeps the transcript until its date and clicks Finalize. The next morning Arjun, a curator, finds the shared decision — with its owner and context — under To review on the Learnings tab of Org Knowledge and verifies it; only then do colleagues see it.

FAQs

Does anyone else see the suggestions?

No. Suggestions and your questions are yours, and so is the session until you click Share with workspace. Even then colleagues read its transcript, context and insights — never the items you discarded — but no suggestions, on the session page or live while you capture; the one trace is cross-meeting search, which still matches a shared session's suggestion titles. When a colleague asks the copilot during a shared session, the copilot keeps their recent questions and answers apart from yours, so neither of you gets the other's in a follow-up. They see shared insights in Org Knowledge only after a curator verifies them. A holder of Manage meeting policy and sessions can delete your sessions but cannot read the ones you have not shared.

Does the copilot speak, type or post into the meeting?

Never. It has no connection to the meeting tool. It drafts text on your screen; what you say or paste is up to you.

Can the copilot help me in a job interview?

Yes — set Assist mode to Interview and attach your resume and the job description from Org Knowledge. See Interview mode. It drafts from your documents only and never speaks for you.

Why did a suggestion arrive late, or not at all?

A suggestion is generated for a question after the turn ends and is dropped or marked Late if the conversation has moved on. With the Economy profile, or when Automatic suggestions is off in the policy, suggestions come only when you ask. The copilot also paces itself — the policy caps AI calls per session per hour, and the pane says pacing itself when the cap is reached.

How do I turn web research off for meetings?

Untick Allow web research on the Policy tab (it is off unless someone turned it on). A workspace admin can also switch web research off for the whole workspace, chat copilot included — see Use the AI assistant.

What does ephemeral mean exactly?

No transcript line, no suggestion and no record of the questions you ask during the meeting is written to the database (so their answers offer no Save as learning), and of the context only what you typed (agenda, participants, constraints) is stored — never the topics, decisions, action items, open questions or unresolved points taken from speech. The transcript, the insights and the full context are held in the server's memory for the review. Finalize drops all of it, whatever you choose for the transcript (Keep it is greyed out for an ephemeral session) — the result says Transcript deleted. unless you saved it to Org Knowledge, and a decision or action item you Save is not kept (Share it to keep it). Only a finalize that reports a warning keeps the memory, and the review stays open so a repeated finalize can still publish what failed — and then only until two hours after the stop. Without a finalize, the transcript, the full context and the extracted items are dropped two hours after the stop at the latest (a late re-extraction does not extend that) — so finalize (or choose Delete it now) when you are done. Notes and exports made while the session is in memory include the decisions and action items heard; the export never contains a transcript, whether you or a colleague exports it.

I am a platform admin — which workspace does my meeting go to?

Meetings, consents, the policy, your preferences, health metrics and paired devices are stored per workspace, and a platform admin's account may belong to none. A meeting with an org selected in Org (evidence and scoping) is stored in that org's workspace — the consent you give before capture is recorded there too, and so is the audit trail. Platform admins also get a Workspace selector in the page header: it follows the workspace of the selected org until you pick another, and it decides which workspace's Policy, Settings, Health and Devices you see, and where a meeting without an org goes. Sessions lists the sessions of every workspace with a Workspace column. Dictation in the Copilot chat works the same way: the microphone's acknowledgement and dictation session go to the workspace of the org selected in the chat (with All orgs, the org selected for the app, else the first org). Everyone else always works in their own workspace; the selector is not shown to them.

Can I use this on a phone?

Phones and tablets can dictate into the chat but cannot capture meeting audio. Use a desktop browser or the desktop companion.

What happens if my connection drops?

Up to twenty seconds of audio is buffered in the browser and replayed on reconnect; the transcript continues where it left off. If the server restarted, the page re-starts the capture on the same session and nothing already transcribed is lost. When the browser has fallen back to sending audio over HTTPS, a piece of audio the server asks it to slow down for is sent again unchanged after the wait the server names, up to five times; only then is it dropped and reported (Some audio could not be sent). Any other refusal or failed upload is reported at once and not sent again.

Troubleshooting

"The share came without audio"

Click Share again and tick Also share tab audio (or Also share system audio for a screen) at the bottom of the picker. Whole-screen audio needs Windows or ChromeOS. A Mac browser can share tab audio only: share the meeting tab, or use the desktop companion for a native client.

"This browser cannot capture this way" or "Safari cannot share meeting audio — your microphone only"

Use Chrome or Edge on a desktop to share the meeting tab, or install the desktop companion. See Set up the desktop companion.

"Choose an org (or a workspace) for this meeting"

Your account belongs to no workspace, and the call did not say which one the meeting is for. Pick an org in Org (evidence and scoping) — the meeting is stored in that org's workspace — or choose one in the Workspace selector in the page header, then try again.

"No speech-to-text provider is answering"

Every provider the policy allows is unavailable or resting after errors. The warning starts with the server's message — No speech-to-text provider is available right now (retrying); the last 15 seconds are buffered; older audio is not transcribed. The server retries every 10 seconds and keeps only the last 15 seconds of audio, which it transcribes when a provider answers again; anything said before that is not transcribed. Once the outage outlasts those 15 seconds, the speech-to-text chip next to the capture controls reads STT · closed, and for a moment its tooltip says audio older than the last 15 seconds is not being transcribed — only until the next retry, at most 10 seconds later, replaces that note. The lasting signals are this warning, which stays up while the outage lasts, and, once a provider answers again (or you pause or stop), About N seconds of audio were not transcribed (below). If it does not recover, stop and ask an administrator to check the providers on the Policy tab, which names the provider and the key it needs.

"About N seconds of audio were not transcribed"

When a provider answers again after an outage longer than those 15 seconds — or you pause or stop while it is still down — the server says how much was lost. This warning appears, next to the outage warning if that is still showing, for two minutes: Speech-to-text was unavailable for longer than the 15-second buffer, so that part of the conversation is missing from the transcript., followed by Transcription has resumed. when a provider is back. After a stop it is shown above the review. The speech-to-text chip reads degraded after a recovery, and its tooltip gives the server's own count (N s of audio were not transcribed: speech-to-text was unavailable for longer than the 15-second buffer).

"Live updates are having trouble"

The transcript, suggestions or insights could not be fetched, or the live connection failed. MetaLens360 keeps retrying on its own and the items may arrive late; the note gives the server's message and, after repeated failures, how many there were. It disappears 30 seconds after the last failure. If it does not clear within a minute, check your network connection.

"Some audio could not be sent"

A piece of audio sent over the HTTPS connection was not accepted. The note gives the server's message and says what happened: The server refused it (it is not sent again), The server was busy and still refused it after 5 retries (the server asked the page to slow down, and the piece was sent again up to five times), or The upload failed (no answer from the server). Either way that part of the conversation is not transcribed; capture continues. The note disappears 30 seconds after the last failure.

"Blocked by workspace policy", "Your workspace does not allow capturing meeting audio; microphone only." or "This session was set up for the microphone only; start a new session to include meeting audio."

The page shows the server's reason as given. Meeting audio, the companion or a provider is turned off under Policy — ask a holder of Manage meeting policy and sessions. A session that was set up for Microphone only never takes meeting audio: set up a new session with the meeting audio chosen. The workspace policy does not allow this capture. appears only when the server gives no reason.

"Consent is missing for this session; set up a new one", "A fresh consent is required" or "Your consent for this session …"

The policy changed since you acknowledged it, so the old acknowledgement is void. Set up a new session and read the dialog again. The consent belongs to the session, so confirming it again never widens it: Your consent for this session covers the microphone only; start a new session and accept capture of meeting audio to include it. and Your consent for this session does not cover the microphone; start a new session and accept microphone capture to include it. both need a new session with the sources you want (in a browser that cannot share meeting audio, Start capture offers that for the microphone — see Start capture). The server needs a fresh consent at the current policy version. appears only when the server gives no reason.

"This session is being stopped; create a new session to capture again."

You clicked Start capture or Resume while this session's stop was still being completed. A stop is final: set up a new session to capture again.

The pill says "Share again" after I reloaded

Browsers cannot resume a screen share without a click. Click Share again; the session, the transcript and the suggestions are unchanged.

"Stop the current capture before resuming another session."

This tab is already capturing another session, and a tab captures one at a time. Stop that capture with the header pill's Stop, then resume the other session.

"Your role does not have permission to do this (meetings.record)" or "Your role cannot record meetings (meetings.record)."

You need Record meetings and dictate. See Permission denied errors.

Related articles

• Dictate instead of typing
• Set up the desktop companion
• Meeting privacy, consent and retention
• Build org knowledge the copilot can cite
• Use the AI assistant
• Roles and permissions