BranchSiteBack to Home

Office dictation, Talk, and privacy

Record an editable assistant draft, talk with the assistant hands-free, and understand what is retained.

Office dictation is an optional preview for enabled accounts. Use Dictate in the office assistant composer, then choose Start. Recording begins only after that explicit interaction and microphone permission. The recording indicator, elapsed time, Stop, and Cancel stay available. Ordinary typing works when dictation is unavailable.

Review before sending

Stop sends the clip for transcription and inserts an editable draft when it can do so safely. If you typed, moved the selection, or changed context while waiting, review the proposed insertion first. Dictation never sends a chat message automatically. Read the draft, correct names, numbers and dates, then press Send.

On an existing client or job, Wrap up a customer call prepares an editable instruction for your own recap. It does not start a microphone or submit anything. Speak your recap after the call; this feature does not record customer calls or third-party conversations.

A note and a follow-up task have separate proposal cards and separate approvals. Check each saved receipt. A saved note does not mean its task also saved. Edit first opens the real form; finish and save there. Sending messages to customers and other manual actions still use their existing forms and permissions.

Recording limits and recovery

A normal recording stops at three minutes. The upload also has a size limit, so a clip can be rejected before three minutes; split a long recap or type the rest. The browser timer does not prove the duration of an arbitrary uploaded file.

Cancel stops local capture and ignores obsolete results. A request already dispatched may still incur cost. Cancel does not undo a note or task you already approved and saved.

A failed transcription may be retried while its audio remains in this tab's memory, for up to five minutes from the first failure. Some failures immediately discard audio. Closing the tab, changing account or thread, or canceling removes local audio; it is not a recoverable saved recording. If recovery expires, record again or type. Your existing typed draft is preserved during normal recording and transcription recovery.

A draft retains its original destination. If the page, account calendar date, or account time zone changes, review the destination and dates before sending. Relative phrases such as “tomorrow” need that review; the proposal shows the resolved date before approval.

What is retained

BranchSite handles office audio transiently for transcription. It creates no office recording bucket and stores no raw audio or transcript in the attempt ledger, routine logs, or analytics before you submit the draft. Once you press Send, the submitted text becomes an ordinary assistant conversation message, with the account's conversation retention policy. Saved notes and tasks have their normal record policies.

Access-controlled attempt metadata records operational outcomes and available usage information. Cleanup selects metadata older than 30 days and runs hourly. Deletion can lag the threshold by up to an hour during normal operation, and longer during failures or a backlog. This is a cleanup threshold, not an instantaneous deletion guarantee. Historical usage aggregates remain after attempt metadata is removed.

Audio is sent through the configured AI service. Provider retention and training policies depend on the approved model and account configuration. A no-training promise is different from zero data retention. Where Settings → Assistant has a Voice section, its Voice privacy choice decides which models hear audio for dictation and Talk; Require zero-data-retention models applies to text chat. Otherwise, accounts requiring zero data retention can use speech only when an eligible model is approved, and speech fails closed. Dictation does not promise universal zero data retention for every provider or model.

Hands-free Talk

Talk is a spoken conversation with the office assistant. Tap Start Talk in the assistant header, or on AI Home, once. The assistant listens and shows your words as you speak. When you pause, it sends your turn, reads its reply aloud, and then listens again on its own. Cards and proposals appear in the thread just as they do for typed messages. Tap End Talk when you're done.

  • Your words appear as you speak. They become your message when you stop talking. How long a pause counts as finished depends on the workspace's When should it answer? setting. Tap Send now to send straight away.
  • Interrupting. When Let me interrupt is on, start talking while the assistant speaks and it stops to listen. Say at least two words together. A single word, a cough or a tap on the desk does not interrupt. Say the whole phrase in one breath, such as "Stop, just the first job". A first word followed by a pause of more than about a second and a half is treated as background noise and dropped. A reply cut off before its text finished is marked Interrupted in the thread.
  • Mute keeps Talk open but sends no audio. Nothing you say while muted is heard or sent.
  • Skip reply stops the spoken reply. Its text still finishes in the thread.
  • Typing during Talk sends an ordinary message, and the conversation carries on.
  • Approvals still need a tap. Saying "yes" or "approve" never applies a proposal. The spoken reply tells you when a card is waiting for you.

Talk ends on its own:

  • after the silence period an admin chose (30 seconds to 2 minutes);
  • after three turns in a row where nothing was heard;
  • when the tab stays in the background for 30 seconds;
  • when you leave the page, switch threads or change workspace.

The bar says why it ended.

Speakers, headphones and noisy rooms

Talk uses your browser's echo cancellation so the assistant doesn't hear itself. On laptop speakers or a phone's loudspeaker, a TV or another person talking nearby can still be heard as you. The assistant may then stop mid-reply, or answer something you didn't say. Headphones avoid this. In a shared or noisy room, an admin can turn off Let me interrupt. Talk then waits for each reply to finish before it listens.

What Talk sends and keeps

  • While Talk is listening, your audio streams from your browser to the AI service chosen by Voice privacy. With Let me interrupt on, that includes the time while the assistant speaks. BranchSite stores no Talk audio. What you said is kept only as the text of your message, like a typed message, under the account's thread retention.
  • Best quality uses OpenAI voice models. Audio sent to them is not covered by zero data retention, even when the workspace requires zero data retention for text chat. The assistant shows a notice about this the first time Talk starts in your browser.
  • Zero data retention uses xAI Grok voice models under zero-data-retention terms. Your words appear in larger steps. Office dictation may be unavailable with this choice; Talk and typing still work.
  • Spoken replies are generated by the same AI service and are not stored.

Talk limits

A single turn can last up to one minute; at one minute it is sent as it stands. Each workspace has a daily listening allowance and a daily speaking allowance, which reset at midnight UTC. When listening runs out, Talk ends and typing still works. When speaking runs out, replies appear as text only. Some workspaces still have the earlier Talk, where you tap to speak each turn; everything above about privacy applies to it too.

Usage and controls

An owner or settings manager can see office dictation separately in Settings → Assistant → Usage. The recent card covers seven account calendar dates, including today. It shows admitted and dispatched attempts, outcomes, known seconds, missing duration, reported cost, and unreported cost. An unknown cost is not zero. Canceled or failed requests can still cost money.

Talk has its own Talk this month card on the same page: sessions, listening minutes, spoken characters and costs. See Assistant settings and usage.

The monthly feature table and recent card are two views of reported speech spending; their values must not be added together. Older aggregates cannot prove complete historical cost coverage after attempt metadata expires. The account calendar is fixed for each attempt when it begins; a later time-zone change does not move it. The speech daily allowance resets at midnight UTC independently of those reporting dates.

Office dictation needs BranchSite preview access and the account's office voice toggle. Disabling it leaves ordinary typed chat and previously approved actions subject to their normal assistant controls. Field voice notes have separate controls.