AI in Australia
Where your AI requests go, in plain words.
AI features send text to a model. Customers and their families are entitled to know where that text goes and what is in it. This is how XTutio, our own product, handles it.
The route a request takes
We tell you exactly where your data and AI requests go.
Model calls can be served by Claude on Amazon Bedrock in Australia: Sydney first, with Melbourne as an in-region fallback, and Anthropic direct only as a last fallback. The route is set for each customer. Customers on Australian hosting stay in Australia for the models Bedrock serves there.
STEP 1
Amazon Bedrock, Sydney
The first stop for a customer on Australian hosting.
STEP 2
Amazon Bedrock, Melbourne
An in-region fallback if Sydney cannot answer.
STEP 3
Anthropic, direct
The last fallback only. This is outside Australia.
What the model is sent
Before a prompt is built, names are replaced with pseudonyms such as "Student A", and the real name is restored only for a person entitled to see it. Contact details are scrubbed. A check in the build fails any code that sends personal fields to a model without the pseudonym layer.
On the screen (made-up)
Ava wants help with her persuasive essay. Her mum Mia is on 0400 000 000.
What the model receives
Student A wants help with her persuasive essay. Her mum Parent of Student A is on [phone number removed].
Example text from XTutio's public security page, which uses made-up names.
Controls around every AI call
A person asks for it
AI runs when a person presses a button. There is no automatic AI, apart from three stated exceptions: enquiry triage on a new enquiry, phone call notes and live call suggestions, and the child-safety checks.
Every call is logged
Each AI call is recorded with its cost, so usage can be seen and questioned.
Children are protected
Safety classifiers check what children send and what they are shown.
A parent can switch it off
A parent can turn AI study tools off for their family, and the tools check that switch first.
Checked in the build
A continuous-integration check fails code that sends personal fields to a model without the pseudonym layer.
No promises about answers
AI can be wrong. It is built to coach and to hold a reply back when a check cannot answer, not to be trusted blindly.
These are the practices behind XTutio, our own product, and the foundation we start client builds from. What applies to a particular build is agreed in its scope, and we say so in writing.