AI Use Policy
Last updated 1 November 2026
Draft pending legal review. This describes how we use AI today. It has not yet been reviewed by a lawyer and will be replaced by a reviewed version before general release. Questions in the meantime: hello@velobiz.com.au.
Velobiz is an AI employee that answers calls and messages for a business. That makes AI our core product, so we hold ourselves to a clear standard. This policy follows the six essential practices in the Australian Government's Guidance for AI Adoption (National AI Centre, Department of Industry, Science and Resources, 21 October 2025) and says what we do for each one. It applies to Velobiz Pty Ltd (trading as Velobiz AI).
1. Decide who is accountable
Velobiz Pty Ltd has a named accountable owner for AI use: the "AI accountable owner" role. That role approves new AI features before release, keeps this policy up to date, and takes concerns raised under it. We review this policy every 6 months.
2. Understand impacts and plan accordingly
Before we release an AI feature we assess how it could affect the businesses that use it and their customers, across calls, messages, websites and content. We ask who could be affected, what could go wrong, and what a person on the other end would reasonably expect. We change the design, or hold the feature back, if the answer isn't good enough.
3. Measure and manage risks
- Grounded answers. Model outputs are grounded in the business's own knowledge: its website, documents and price list. The assistant doesn't make up prices.
- Claim checks. Checks stop replies and content that promise things the business hasn't said it will do.
- Moderation that fails closed. Website content is moderated before it is published. If the check can't run, the content is held, not published.
- Limits. Rate limits and cost caps limit how much the AI can do in a period, so a fault or misuse can't run away.
- Consent and Do Not Call. Consent and Do Not Call checks run before any outreach.
4. Share essential information
People should know when they're dealing with AI.
- Calls. Our default greeting is: "Thanks for calling {business}, you're speaking with their AI assistant — calls are recorded. How can I help?"
- If asked. The assistant never claims to be a person. If someone asks, it says: "I'm {business}'s AI assistant — I can take your details for the team."
- Website chat. The chat shows: "You're chatting with {business}'s AI assistant"
- Summaries. The written call and chat summaries in a business's dashboard are produced by our AI from the conversation, and can contain mistakes.
A business that writes its own greeting must keep both disclosures: that it is an AI assistant, and that calls are recorded. This is in the Terms of Service and the Acceptable Use Policy.
5. Test and monitor
- Automated regression suites run before every release.
- We run live call checks against the real phone network.
- We monitor for failures and complaints, and fix what we find.
6. Maintain human control
- A business can switch any agent or feature off at any time.
- A business approves drafts and content before they go out.
- A business can edit any record, and every edit is kept in an audit trail.
- A caller can always leave a message for the team.
We don't offer a live human takeover of calls or chats. If someone needs a person, the assistant takes their details and the team calls back.
AI providers and what they receive
We use the providers below to run the service. Each receives only what it needs for its job. Where each one is located is listed in our Privacy Policy.
| Provider | What it does | What it receives |
|---|---|---|
| Anthropic | AI language model that writes replies, summaries and drafts | Conversation text and the business knowledge needed for the task |
| Groq | AI language model for fast replies | Conversation text |
| VAPI, Deepgram, Microsoft Azure and Cartesia | Run the call, turn speech into text, and speak the replies | Call audio and the text of the conversation |
| Gemini AI models, search over business knowledge, and text-to-speech | Text for the task, and text to be spoken |
Velobiz does not use your data, or your customers' data, to train AI models. Anthropic and Google's paid API services don't use the data we send them to train their models under their commercial terms. We are confirming the equivalent settings for every other provider and will list them here.
How long we keep data
Calls, transcripts, summaries and lead records are kept and deleted as the Privacy Policy describes.
What the AI can't do
- It can make mistakes. Businesses should review what it does on their behalf.
- It isn't designed to give emergency, medical, legal or financial advice. In an emergency, call 000.
Raising a concern or asking for human review
If an AI decision or reply worries you, or you'd like a person to review it, email hello@velobiz.com.au, or privacy@velobiz.com.au for privacy matters. Tell us what happened and we'll look into it and reply.
Reviewing this policy
We review this policy every 6 months, and whenever we release a significant new AI feature.