“Does our data stay in Australia?” is the first question most clients ask when they start using AI for real work. The honest answer is that it depends on four choices: which product your staff use, which platform it runs on, which model you pick, and one setting most people never look at. Get any one of them wrong and your data can leave the country without anyone noticing.
This guide covers AI data residency in Australia across the tools businesses actually use in 2026, where each one sends your data, who keeps the logs, and the leak that residency settings do not fix at all.
| How you use AI | Who processes your data | Can it stay in Australia? | Whose logs |
|---|---|---|---|
| Free or personal plans (ChatGPT, Claude, Gemini) | The AI vendor | No guarantee | Nobody’s, from your point of view |
| Business apps (ChatGPT Enterprise, Claude Enterprise, Microsoft 365 Copilot) | The vendor, under a business contract | Varies by vendor and plan | The vendor’s admin console, or Microsoft Purview for Copilot |
| Microsoft Foundry (Azure) | Microsoft, or Anthropic for some Claude deployments | Only for some models and deployment types | Your Azure subscription |
| Amazon Bedrock (AWS) | AWS, inside your AWS account | Yes, if you pick the Australian profile | Your AWS account |
Free and personal plans are where most of the risk sits. They are also what staff reach for when the business has not given them anything better. Business apps exclude your data from training by contract, but the vendor still processes it and still holds the records. The two cloud platforms are the only options where you control the region, the access and the logs yourself, and even they have traps.
Amazon Bedrock runs models inside AWS, and the model providers cannot see your prompts. AWS states that Bedrock does not use your inputs or outputs to train models, and that each provider’s model runs in an AWS-owned account the provider cannot access.
Where the data is processed depends on the model ID you call. Current Claude models have an Australian profile, for example au.anthropic.claude-opus-5-5, that keeps processing within Australian regions. The same model also has a global. profile that routes requests to any AWS region worldwide for extra capacity, with no residency guarantee.
Both IDs work. Both return the same answers. Only one keeps your data in Australia. Availability also changes model by model: earlier this year some newer Claude models could only be reached from Sydney through the global profile, so choosing the most capable model meant quietly giving up residency.
The fix is simple and easy to forget. Write down the exact model ID for every workload that handles sensitive data, and check it again every time you switch to a new model.
Most of our clients run on Microsoft, so Foundry is the natural place to start. It is a good platform, but its residency story is uneven.
OpenAI models on Azure can be deployed as “Standard” regional deployments in Australia East, which process data in that region. The newest models are often not offered that way, though. The alternative Asia Pacific data zone means processing anywhere in the Asia Pacific region, not just Australia.
Claude models on Foundry have no Australian option at all today. They are offered in Global and US data zone deployments only. Microsoft’s own hosting comparison also notes that the Anthropic-hosted version can process data outside Azure entirely. If Australian processing is a hard requirement and you want Claude, Bedrock is currently the platform that offers it.
When an auditor, insurer or client asks “who accessed this data, and what did the AI do with it?”, you need evidence, not a vendor’s assurance. This is where the options really separate.
There is a catch on Bedrock. Its standard audit trail records that a call happened, who made it and when, but not what was in it. Capturing the actual prompts and responses needs model invocation logging, which is disabled by default. Turn it on for sensitive workloads, then protect those logs carefully, because they now contain the sensitive data too.
Picture a staff member asking an AI assistant to research treatment options for a named patient, or background on a confidential acquisition target. If the assistant can search the web, it may send a query containing those details to a public search engine. No system was breached and the model stayed in Australia, but the private information has left your business anyway.
Residency settings do nothing about this, because the leak happens through the tools the model is given, not where the model runs. The same is true on Bedrock: if you switch on a web search tool, the search goes to the public web with content drawn from the prompt.
The answer is the harness, the software layer that decides which tools and network destinations an AI agent can reach. Set it to block outbound internet access by default, and allow web search only for workloads that never touch sensitive data. This is the same layer the Australian Signals Directorate published guidance on in September, and the same failure behind the AI agent that reached a Medicare portal it was never meant to touch.
Australian residency means your data is processed and stored here. It does not change who owns the infrastructure. AWS, Microsoft and the AI labs are all US companies, subject to US law, including the US CLOUD Act, wherever their data centres sit.
For most businesses, Australian residency plus the provider’s contractual commitments is a sensible, defensible position. For Defence suppliers and some government work, contracts can demand more. If you are a DISP member, our DISP accreditation guide covers how AI fits within those obligations.
We run AI agents across our own ticketing, documentation, quoting and finance systems, and every one of them connects through a bridge we built. The model never holds a password: credentials stay on the bridge. Each role is granted specific tools and nothing else, and every call, allowed or refused, is logged. If a role is not granted a web search tool, no clever prompt can make the agent search the web.
That separation, where the platform decides what the model can reach, is what we build for clients through our managed AI service and AI agent development. Residency choices sit alongside it as part of our AI governance work.
Find out which AI tools your staff actually use. If the answer includes free or personal plans, that is your biggest residency and evidence gap, and it will not show up in any report. Give people an approved tool before you block the others.
Match sensitive workloads to a platform and model that stay in Australia. For health, legal, financial and Defence-adjacent data, pick a deployment with an Australian option, record the exact model ID, and recheck it whenever the model changes. Our Privacy Act 2026 checklist covers the disclosure obligations that sit alongside this.
Turn on logging and turn off outbound access by default. Make sure prompts and responses for sensitive workloads are logged in your own account, and that agents cannot reach the internet unless a workload genuinely needs it. Contact us on 1300 EPIC IT for a free AI data residency review and we will map where your AI data goes today.
Yes, but only with the right platform, model and settings. Amazon Bedrock offers Australian profiles for current Claude models, and Azure offers Australia East regional deployments for some OpenAI models. Free and personal AI plans offer no residency guarantee. AI data residency in Australia is a per-model setting, so check it every time you change models.
It does if you call the model through its Australian profile, which uses an “au.” model ID and processes requests within Australian AWS regions. The “global.” profile for the same model can route requests worldwide. AWS states Bedrock does not use your data to train models and that model providers cannot access it.
Not currently. Claude on Microsoft Foundry is offered in Global and US data zone deployments only, and the Anthropic-hosted option can process data outside Azure. Businesses that need Claude with Australian processing can use Amazon Bedrock’s Australian profile instead.
Business and enterprise plans from the major vendors exclude your data from training by contract, and so do Amazon Bedrock and Microsoft Foundry. Free and personal plans are different: depending on the vendor and your settings, conversations may be used to improve models. Check each plan’s data terms before staff use it for work.
No. Residency is about where data is stored and processed. Sovereignty is about which country’s laws can reach it. Australian regions run by US providers give you residency, but those providers remain subject to US law. For most businesses residency is enough; some Defence and government contracts require more.
