Huddleboard — optional voice disclosure Version: voice-2026-11-02.2 | English

1. Availability and scope

Voice is an optional huddle feature. When enabled, it uses the workspace OpenAI key for live conversation. If no OpenAI key is configured, an enabled OpenRouter fallback uses the workspace OpenRouter key for one recorded question and one spoken answer at a time. An invalid or exhausted OpenAI key does not silently switch providers. This notice does not amend agreements already accepted. The applicable terms, privacy notice and business DPA remain available at /legal and /dpa. Voice is intended for adults. The voice assistant is AI, not a human colleague; its speech and transcript can be inaccurate and should be checked before use.

AI usage and tokens: voice requests consume billable AI processing. Depending on the model and stage, providers meter input/output text or audio tokens, recording duration, or characters used for speech synthesis. Tokens are units of model input/output, not a Huddleboard subscription or a fixed price per message. Billable input can include relevant project/agent context as well as your question; output includes the generated answer. OpenRouter fallback has three billable stages: transcription, the model answer and speech generation. Check the connected provider account for current prices and actual charges. Huddleboard’s displayed usage is an estimate and can include conservative reservations; it is not an exact token counter or provider invoice.

In the current implementation, opening a huddle, viewing updates derived from existing runs, posting text notes and recording a typed @agent question do not themselves make a new model call. Typed agent questions are not yet answered automatically. Any charges for the underlying agent runs are separate. Voice starts provider processing when a live connection is established or a recorded question is submitted. Playing the prerecorded public website demo does not use your provider key or consume your AI credits.

2. Customer account and charges

For live conversation, your workspace administrator connects an OpenAI API account they are authorised to use for the workspace. One workspace connection can fund use by its authorised members; it is not limited to the administrator who entered the key. OpenAI bills the connected account under its applicable agreement and pricing. In fallback mode OpenRouter bills the connected OpenRouter account for transcription, the answer and speech generation; no separate OpenAI key is required. Huddleboard does not resell keys or give members the provider key or OpenAI account login access. Do not share personal login credentials or distribute another person's personal API key. For a team integration, use a dedicated project and an appropriately restricted service-account key where available. This security recommendation does not itself establish contractual permission for third-party custody.

3. What is sent and stored

In OpenAI live mode, microphone audio goes directly from your browser to OpenAI for speech processing. In OpenRouter mode, a recording of up to 30 seconds passes through Huddleboard in memory to OpenRouter and the speech provider for transcription. The transcript and relevant project/agent context go through OpenRouter to the answering model provider, and the answer goes to the speech provider to generate synthetic audio. The current fallback uses Mistral Voxtral transcription and speech models and GPT-4o mini for answers. Huddleboard temporarily handles both audio directions but does not persist audio recordings. Huddleboard handles connection signalling, relevant workspace/agent context and transcript text. Signalling may contain network addresses; Huddleboard does not persist the signalling payload or audio recordings. Huddleboard stores transcripts as workspace content accessible according to workspace permissions. OpenAI has its own processing and retention obligations; our decision not to store audio does not mean OpenAI retains nothing.

4. Privacy settings and international processing

OpenRouter fallback uses its global endpoint and requests no-collection and zero-retention routes for all three stages; if no eligible route exists, the turn fails without relaxing those filters. These filters do not guarantee that every service record is unretained, establish EU-only processing, or replace appropriate contracts and transfer safeguards. Provider exceptions, security/legal obligations and account logging settings still apply. The applicable OpenRouter and model-provider agreements must cover the actual roles and data, including any processor-to-processor transfers. Do not submit sensitive or business-controlled personal data without the required arrangements.

The direct OpenAI integration uses OpenAI's global API endpoint. It does not promise EU-only processing or zero data retention. OpenAI's current API documentation lists a default 30-day abuse-monitoring retention period for Realtime and eligibility for approved zero-retention controls. Account/project settings, approved exceptions and legal obligations can affect processing. API data is not used for model training by default; optional sharing choices can change this. Review the actual connected account before use.

OpenAI's standard DPA provides for OpenAI Ireland to process EEA/Swiss customer data and for onward transfers under SCCs or an adequacy decision. That framework does not establish that every account, recipient or customer workflow is compliant. For business-controlled content, Huddleboard processes workspace data under its DPA and customer instructions. The customer-selected OpenAI account has its own agreement; paying OpenAI directly does not remove Huddleboard's data-protection responsibilities. Account administration and billing data may be governed separately from API content.

5. Participant notice and permitted use

Before starting, inform everyone whose speech may be captured whether audio will be sent directly to OpenAI or through Huddleboard to OpenRouter and its speech/model providers and that a transcript will be saved for authorised workspace members. Establish the lawful basis and any permissions required for your situation; browser microphone permission alone is not participant consent. Do not capture people unaware of transcription. Avoid sensitive personal data or confidential information unless the required authority, safeguards and provider arrangements are in place. Huddleboard does not provide speaker identification, voice cloning or biometric identity verification through this feature.

OpenRouter mode allows one question per session. Before provider calls, Huddleboard reserves $0.03 against the huddle’s estimated voice allowance. This is a conservative usage estimate, not a Huddleboard fee or the actual provider price. The reservation remains after a failed or interrupted provider attempt, because charges may already have occurred. The upload is limited to 30 seconds, provider work to 45 seconds, and concurrent/rate limits can reject a busy request. Stopping playback or closing the page cannot recall data already sent or cancel charges already incurred. There is no automatic retry of a submitted turn. Technical admission records contain session/workspace identifiers and timestamps, no audio or key; records older than one day are removed on the next voice admission.

6. Allowances and stopping calls

Time and cost allowances are estimates, not guaranteed invoice caps. The backend requests provider hangup at the deadline and on relevant stop events, and retries failed requests. The browser also stops its microphone. Network, database or provider outages, revoked credentials and retry delays can postpone provider termination and cause additional charges. Check your provider's usage and available controls. These explanations do not exclude Huddleboard's responsibility for its own obligations or restrict mandatory consumer rights.

7. Keys, transcripts and deletion

The configured key is encrypted at rest and decrypted by authorised backend services when needed. Removing it requests shutdown of active calls. A separate encrypted copy is retained solely to finish pending shutdown, including after workspace deletion, and cleared from the active database when termination is confirmed. Persistent failures require restricted operational follow-up; we cannot promise immediate removal of a key still needed to close a call. Completed shutdown records, without the key, are scheduled for removal after one day. Backup expiry follows the main privacy notice.

Transcripts follow the workspace retention, access, export and deletion arrangements in the main privacy notice. Removing a key does not erase transcripts or data already processed by OpenAI. Contact contact@huddleboard.ai for rights requests when the service launches; this mailbox must be operational before launch. The main privacy notice identifies the operator and explains complaints and data-subject rights.