Trust Centre
How AI is used.
Which models do what, what they see, what we won’t do with it, and our practice statement.
Which models do what
Command sends the text needed for each task to an AI model provider. Here is what each one does.
| Provider | What it does | What it sees | Where it processes it |
|---|---|---|---|
| OpenAI | Language models. Voice and transcription features, which aren’t live for every client yet. | The text, or the voice, sent for each task | United States |
| Anthropic | Language models | The text sent for each task | United States |
A model-based second check runs on Gmail replies before they’re sent. It can hold a reply, and you can overrule the hold.
Which provider handles which task, and what happens when one is unavailable, will be listed on this page before launch.
No training on your data
In progress · target Nov 2026
We use AI providers’ business services. We’re checking that their terms don’t let our clients’ data be used to train their models, and we’ll confirm each provider’s terms on this page before launch.
For each provider, we’ll publish the date we read its terms, a link to them, how long it keeps data, and whether zero data retention is switched on.
What we won’t do with it
- We won’t make calls to people until consent, do-not-call checks and disclosure are in place.
- We won’t use one client’s records to improve another client’s answers.
- We won’t say answers are scored for quality until there is a test suite that scores model answers.
Guidance for AI Adoption: our practice statement
The Australian Government’s Guidance for AI Adoption (2025) sets out six practices. This is what we do against each one, and what is still being set up.
In progress · target Nov 2026
1. Accountability
Camber & Casper is accountable for what Command does for each client, including which AI providers it uses. Each client will sign a written agreement before we connect their records. We’ll name the role that holds this accountability, and the person who signs off each new AI provider, before we add one.
2. Planning around AI’s impacts
Before a new use of AI goes live, we’ll write down who it touches and what could go wrong for them. Voice calls to people won’t start until consent and disclosure controls are in place. Film and images made with AI are labelled as AI-made.
3. Measuring and managing risk
Command is built to say when it can’t see something, rather than guess, and to show where an answer came from. We’re confirming that every answer does. Forecasts are graded and misses go on the record; the grading is built and tested, but not yet proven with a live client. We’ll keep a written risk register and review it on a set date.
4. Sharing information
We’ll publish the list of AI providers, where each one processes data, and the date it last changed, and we’ll give notice before we add one. We’ll tell each client which AI providers their text goes to and what each one does with it. When something goes wrong, we’ll publish a short note on what happened and what we changed.
5. Testing and monitoring
We’ll test every client-facing change before it’s released. The local test suite passed on the candidate build on 5 October 2026, with 643 server checks. Repairs are checked by a separate audit, not by the person who made them, and a repair isn’t called proven until it is. No test suite yet scores model answers, and we won’t claim one until it exists.
6. Keeping humans in control
Command sends nothing without your OK. A Gmail reply goes out only when you confirm it on screen, the server refuses the send otherwise, and no more than 30 a day are sent. A second check can hold a reply, and you can overrule the hold. We won’t have Command sign, pay or agree to anything for you, and we’ll put your right to switch it off in the client agreement.
This is our own statement against the Australian Government’s voluntary Guidance for AI Adoption (2025). It is not a certification.
Last reviewed 15 October 2026.