1. Scope
This notice explains how BotHub may use transcription, machine learning and generative artificial intelligence. It applies to messages and to voice calls placed through WhatsApp to a Client’s number. Features are enabled by a Client or White-Label Platform Operator and must be recorded in the Client’s Statement of Work or configuration schedule. This notice does not authorise a feature that is not contractually enabled.
2. Principles
- Purpose limitation — process only the information needed for the approved feature.
- Client control — the Client approves the feature, data sources, rules and human-review level.
- No general training by default — Client Data is not used to train Bvelaphanda’s general-purpose models.
- Provider restriction — external providers must be approved, contractually restricted and identified.
- Human accountability — AI assists; it does not transfer the Client’s business responsibility to Bvelaphanda.
- Transparency — material automated interactions should be disclosed where appropriate.
- Security and separation — prompts, documents and outputs must remain logically separated by Client.
3. Typical features
| Feature | Input | Output | Typical purpose | Required review |
|---|---|---|---|---|
| Voice-note transcription | Voice note sent as a message | Text transcript | Read, search, route or prepare response. Transcription is provided by an external provider identified in §4. | Review where accuracy matters; preserve audio only as required. |
| Voice calls | A call placed through WhatsApp to a Client’s number | Speech converted to text, and spoken replies | Answer a call with the Client’s assistant where the Client has enabled it, or route the call to a person. Speech-to-text and text-to-speech run on Bvelaphanda-controlled infrastructure; only the text of what was said is sent to the external provider that supports the Client’s assistant (§4), exactly as a typed message would be. The caller is told at the start of the call that the call is recorded and that an automated assistant may answer. | A caller may ask for a person at any time; the call is then transferred. Calls are recorded and may be reviewed by a person (§8). |
| Image reading | A photograph sent by a Messaging Participant (for example a utility meter, or a photograph attached to an order or complaint) | Structured values read from the image (for example a meter reading and serial number), or a description | Capture a value the participant would otherwise type; route a complaint. Provided by an external provider identified in §4. | The participant types the value; the value read from the image is used only to check it. Where the two disagree, or the image cannot be read, the reading is held for a person to review or referred back to the participant. |
| Classification / intent | Message text and context | Category, intent, tag or route | Assignment and workflow selection | Monitor false classifications and escalation. |
| Knowledge search | Question and approved Client knowledge source | Relevant passages or response context | Support response preparation | Verify source currency and access control. |
| Summarisation | Conversation or case history | Summary | Handover, case review or reporting | Human review for consequential use. |
| Suggested response | Message, context and instructions | Draft response | Agent assistance | Human approval by default unless low-risk automation is explicitly approved. |
| Automated response | Message, rules and knowledge | Response sent without pre-send review | Low-risk routine service | Requires explicit SOW approval, testing, escalation and monitoring. |
4. External providers
External AI and transcription providers are enabled on BotHub. This section identifies each one, the feature it supports, the data it receives, where it processes that data, how long it retains it, and whether that data is used to train the provider’s models.
| Provider | Feature | Data sent | Processing region | Retention | Used for model training? |
|---|---|---|---|---|---|
| OpenAI, L.L.C. | Voice-note transcription; image reading (photographs sent by Messaging Participants, including utility meter photographs and photographs attached to orders and complaints); and the assistant that composes replies on conversational services | Audio of voice notes; images sent by Messaging Participants; and, on conversational services, the text of the conversation | United States | Up to 30 days for abuse monitoring, then deleted (OpenAI API terms) | No — API data is not used to train OpenAI models under the standard API terms. |
| Google LLC (Gemini API) | Reading the content of documents — menu files supplied by a Client; and image reading of utility meter photographs sent by Messaging Participants on a meter-reading service, alongside OpenAI | Documents supplied by a Client; utility meter photographs on a meter-reading service. No conversation text is sent to this provider. | United States | 55 days | No — the account is on a paid tier, under which prompts and responses are not used to improve Google’s products and are covered by Google’s data processing terms. |
Which services use which provider. A meter-reading service sends the participant’s meter photographs, and nothing else, to OpenAI and to Google, each of which reads the meter independently; a reading is taken forward only when the service’s own checks pass, and the value is confirmed by the participant before it is recorded. A person reviews readings the rules hold back. A restaurant service sends voice notes, photographs and conversation text to OpenAI, and the Client’s menu documents to Google.
Voice calls. No audio from a call is sent to any external provider. Converting speech to text and text to speech for calls takes place on Bvelaphanda-controlled infrastructure. Where a Client’s assistant answers a call, the transcript of each turn is handled by the same external provider, and under the same terms, as that Client’s typed messages (OpenAI, L.L.C. for a conversational service). Where a call is routed to a person, no AI processing of the call takes place; the call may then be carried over a telephone network provider, which is recorded in the subprocessor register referred to in the Privacy Policy.
Scope of engagement. Each provider above is engaged for images, documents, audio and video, so that Bvelaphanda can select the most suitable service for a given task without interrupting the service. Anthropic PBC is engaged on the same basis. The table below records which are in active use at the date of this notice.
| Provider | Engaged for | In active use today |
|---|---|---|
| OpenAI, L.L.C. | images, documents, audio, video, conversation text | Yes — voice notes, images, and the assistant |
| Google LLC (Gemini API) | images, documents, audio, video, conversation text | Yes — documents, and meter photographs on a meter-reading service |
| Anthropic PBC | images, documents, audio, video, conversation text | No — no data is sent to it |
Where a task moves between these providers, the safeguards in this notice apply unchanged: each operates on a paid business plan under which Client and Messaging Participant data is not used to train their models.
Self-hosted components are not external providers. The Workflow Engine, database, media storage and web delivery run on Bvelaphanda-controlled infrastructure and are covered by the Privacy Policy rather than this section.
Client-selected providers. Where a Client selects and contracts an AI service directly, using their own account and API key, that provider is the Client’s processor rather than Bvelaphanda’s subprocessor. It is recorded in that Client’s Statement of Work. It is nonetheless listed in Bvelaphanda’s subprocessor register for completeness.
Commitment. Before any additional external provider is activated, this notice will be updated to identify the provider, feature, data sent, processing region, retention and training setting, and affected Clients will be notified in advance.
Note on this version. Version 1.0 of this notice stated that no external providers were enabled. That statement was incorrect at the time this version was issued: the providers listed above were already in use. This correction is retrospective and is published as such. Bvelaphanda did not notify Clients in advance as Version 1.0 committed; Clients are being notified on publication of this version. The technical safeguards described in §§2, 5 and 6 applied throughout.
Version 1.4 adds voice calls placed through WhatsApp (§§1, 3 and 4) and call recording (§8), and corrects two statements: the image-reading row of §3 now describes the check as it actually operates (the participant types the value; the image is used to verify it), and §8 now states the meter-photograph retention period in full. No external provider has been added. Clients whose service can receive calls are notified before calls are enabled on their number.
5. Accuracy and limitations
- Outputs may be incorrect, incomplete, outdated, biased, insecure or unsuitable for the context.
- A transcription may mishear names, numbers, languages, accents or background audio.
- A value read from an image may be wrong even when the model reports high confidence; that is why a participant confirms it and why held readings are reviewed by a person.
- A model may invent information or fail to follow an instruction.
- A knowledge source may be outdated or accessible to the wrong user if Client permissions are incorrect.
- The Client must define escalation when confidence is low, information conflicts or a person requests human assistance.
6. High-impact and prohibited uses
BotHub must not make a solely automated decision with legal or similarly significant effects unless the Client has completed a documented assessment, confirmed legality, implemented meaningful human intervention and obtained Bvelaphanda’s written approval. Health, financial, legal, employment, credit, insurance, biometric, children’s and emergency uses require heightened review and may be refused.
7. Responsibility
| Bvelaphanda | Client | White-Label Platform Operator |
|---|---|---|
| Implement the approved feature with reasonable skill and care; apply agreed security; follow documented instructions; correct reproducible platform defects. | Select lawful purposes; approve prompts, knowledge, rules and review; verify outputs; make business decisions; provide participant notices. | Flow down the same requirements, restrict support access, ensure downstream approvals and prevent cross-Client reuse. |
8. Retention and logging
The Statement of Work must specify whether audio, transcripts, prompts, outputs, embeddings, evaluation records and provider logs are retained, where they are stored and when they are deleted. Retention must be no longer than necessary. Debug logging must not silently become an indefinite copy of Client conversations.
Call recordings. Every call placed through WhatsApp to a Client’s number is recorded from the moment it is answered until it ends, including any part of the call handled by a person. The caller is told this at the start of the call. Recordings are stored only on Bvelaphanda-controlled infrastructure and in its encrypted backups; they are never sent to an external provider. They are kept for quality review and for evaluating the service. The retention period for call recordings and call records is under review and will be set no later than 1 October 2027; until then they are retained, and this notice will be updated when the period is set.
Meter photographs. On a meter-reading service, a photograph sent to read a meter is kept for three years from the date it is received, then deleted; the reading taken from it is kept with the account record.
9. Participant choices and complaints
A Messaging Participant may request human assistance, object to direct marketing and exercise applicable privacy rights through the Client. Where a materially consequential automated process applies, the Client must provide the information and safeguards required by law. Bvelaphanda will assist as Operator.
10. Feature approval checklist
- Purpose and lawful basis approved.
- Data fields minimised.
- Provider and location verified.
- Training disabled or separately authorised.
- Retention and deletion tested.
- Human-review level documented.
- Negative and edge cases tested.
- Escalation and opt-out implemented.
- Client notice updated.
- Production acceptance signed.