We are very interested in using the new Realtime Voice capability with Glean Assistant for our support operations. It is a great feature, and hands-free, real-time conversation would be very valuable in our environment.
However, in a support or call center setting, the microphone can easily pick up background conversations and sensitive information (customer names, contract details, internal codes, etc.). As far as we understand the current behavior, the recognized speech is directly sent as a prompt to Glean, which means there is a risk that confidential information is unintentionally sent to the assistant.
From a security and compliance perspective, we would like to avoid sending such information to Glean without explicit user control. To address this, we would like to request the following enhancements:
- Post-transcription review and edit step (approval before sending)
- After speech is converted to text, the content should first be shown to the user as a draft instead of being immediately sent to Glean.
- The user should be able to review and edit the text (e.g., remove or mask sensitive parts, adjust wording) before it is submitted.
- Only when the user explicitly clicks a “Send” or “Use this prompt” button should the text be sent to Glean Assistant.
- Whether this confirmation step is required should be controllable as an organization-level policy in the Admin Console (Assistant → Realtime Voice).
- Policy to enforce “manual send” instead of automatic submission
- In addition to a fully automatic mode, we would like an option where:
- Speech-to-text runs automatically, but
- Submission to Glean is never automatic and always requires an explicit user action.
- Admins should be able to enforce this “manual send required” mode for all users (or specific groups) as part of the Realtime Voice settings.
- Simple redaction / masking support (optional but very helpful)
- In the review UI, it would be helpful to have an easy way to quickly mask selected text (e.g., replace with
***) before sending. - Alternatively or additionally, if there is a list of sensitive keywords managed by admins (such as known customer names), the system could highlight these in the draft and warn the user before sending.
Example usage flow we would like to achieve:
- An operator speaks to Glean Assistant using Realtime Voice.
- The speech is transcribed to text, which appears as a draft on the screen but is not yet sent.
- The operator reviews the text and removes or masks any background conversation or customer-sensitive information.
- The operator clicks “Send to Glean” to confirm.
- Only then is the text sent as a prompt to Glean Assistant, and the assistant’s response is returned.
With this flow, we believe we can benefit from the productivity of Realtime Voice while maintaining our security and data protection requirements in a support environment.
Thank you very much for considering this feature request, and we would appreciate it if you could take this into account for future improvements to Realtime Voice.
Best regards,
