Which AI tools keep your data in Australia?
You can keep AI prompts in Australia today, but only on business plans, and only if someone switches the right setting on. Microsoft 365 Copilot stores prompts and responses in Australia for Australian tenants. Azure OpenAI and Amazon Bedrock can also keep the model processing onshore. ChatGPT Enterprise and the OpenAI API can store data in Australia, but the model may still run offshore. Gemini for Workspace offers no Australian data region. The Claude API offers only US or global processing.
The question reaches IT in different words. Half the team is already on free ChatGPT, a client contract says "we can't send that data offshore", or a manager wants to put client files into Copilot.
What is AI data residency? It is a vendor's commitment about where your prompts, files and outputs are kept. Storage residency covers data at rest. Inference residency covers where the model processes your request. They are separate commitments, and many plans offer one without the other.
As checked 4 November 2026. Every residency fact below comes from the vendor's own documentation, linked in each section and listed under Sources. Vendors change these settings often. Where we could not confirm a fact, we say "check with the vendor". Use it as a starting list for your own check. It is not legal advice.
Why this question has a deadline now
Two Privacy Act obligations make the answer matter this quarter. The OAIC explains both.
The first is APP 8, cross-border disclosure of personal information. Before personal information goes to an overseas recipient, your organisation generally has to take reasonable steps to protect it. Whether a given AI service counts as a disclosure depends on the facts. The OAIC's APP 8 guidelines set out the test.
The second is new. From 10 December 2026, APPs 1.7 to 1.9 require your privacy policy to describe certain automated decisions that use personal information. The OAIC confirmed the date when it released its transparency fact sheets and flowchart on 30 September 2026. If an AI tool helps make decisions about customers or staff, you will need to know what it uses and where. Our guide to Privacy Act automated decision making covers that obligation in detail.
The OAIC's guidance on commercially available AI products (21 October 2024) recommends that organisations do not enter personal information, and particularly sensitive information, into publicly available generative AI tools. It names APP 8 as one of the obligations to weigh.
The one-table answer, product by product
Each row is unpacked below, with the vendor link.
| Tool and plan | Prompts stored in Australia? | Model runs in Australia? | Trains on your data by default? | What you switch on |
|---|---|---|---|---|
| Microsoft 365 Copilot and Copilot Chat (work account) | Yes, for tenants provisioned in Australia | Not guaranteed. Check with Microsoft | No | Confirm data location; set the web search policy |
| Azure OpenAI in Microsoft Foundry | Yes, in the resource's geography | Yes, with Standard or Regional Provisioned deployments | No | Deploy in an Australian region; block Global types |
| ChatGPT Free, Plus, Pro | No residency option | No | May, unless switched off | Turn off "Improve the model for everyone", or block it |
| ChatGPT Business | Not listed as eligible. Check with OpenAI | No | No | Admin controls; check residency with OpenAI |
| ChatGPT Enterprise and Edu | Yes, for new workspaces | No (EU, US, UAE only) | No | Choose Australia when the workspace is created |
| OpenAI API | Yes, per project | No | No | Australian project region; approval required |
| Gemini for Google Workspace | No (US or Europe only) | No | No | Nothing onshore to switch on today |
| Gemini on Vertex AI | Yes, at rest in your chosen location | Yes, in australia-southeast1, for listed models | No | Use the Sydney regional endpoint, never global |
| Claude Free, Pro, Max | Not documented. Check with Anthropic | Not documented | Only if the model improvement setting is on | Check the setting, or block it |
| Claude Team and Enterprise | Not documented for Australia. Check with Anthropic | Not documented | No | Ask Anthropic before relying on it |
| Claude API (direct) | No (workspace geo is US only) | No (US or global only) | No | Not an onshore route today |
| Claude on Amazon Bedrock | Yes, in your AWS account | Yes, with the AU inference profile (Sydney and Melbourne) | Model providers cannot see prompts (AWS) | Use the au. profile; restrict regions with SCPs |
Copilot answers stay in your tenant, with one processing caveat
Where prompts are stored. Microsoft's Product Terms commit to store "any stored content of interactions with Microsoft 365 Copilot or Microsoft 365 Copilot Chat" at rest in the tenant's geography, and Australia is on the list. Copilot was added on 1 March 2024 and Copilot Chat on 1 September 2025, per Microsoft's Product Terms data residency page.
Where processing happens. Microsoft's Copilot privacy documentation says calls to the model are routed to the closest data centres, but can go to other regions during high demand. It also says "customers outside the EU may have their queries processed in the US, EU, or other regions." Microsoft announced plans in November 2025 to offer in-country processing of Copilot interactions in Australia. Whether that is active for your tenant today is a question to put to Microsoft.
Training. Microsoft states that prompts, responses and data accessed through Microsoft Graph are not used to train foundation models.
What to switch on.
- Confirm your tenant's data location under Settings > Org settings > Organization profile > Data location in the Microsoft 365 admin center.
- Decide on web search. When it is on, Copilot sends short generated queries to Bing. Microsoft says its Data Protection Addendum does not apply to those queries. Admins control this with the Allow web search in Copilot policy in Cloud Policy service.
- Decide whether to allow third-party models. Microsoft offers Anthropic and OpenAI models inside Copilot as subprocessors, with additional terms. Read those terms before you enable them.
- Make sure staff use Copilot Chat signed in with their work account. The commitments above attach to your tenant.
Residency is only half of Copilot readiness. Copilot surfaces anything a user can already open, so overshared SharePoint sites become a bigger risk. Our Microsoft 365 Copilot readiness guide covers the permissions audit.
Azure OpenAI in Australia East keeps processing onshore if you pick the right deployment
Azure OpenAI now sits inside Microsoft Foundry. It is the route most Microsoft shops use when they build their own AI tool rather than buying Copilot.
Where data is stored. Microsoft's deployment types page states that data stored at rest remains in the designated Azure geography for all deployment types.
Where processing happens. It depends on the deployment type you choose:
- Global types may be processed in any Azure region.
- Data Zone types are processed within a Microsoft-defined zone. The Asia Pacific zone covers several APAC regions, so it is not Australia-only.
- Standard and Regional Provisioned types are processed within the Azure geography you specify.
Only the last group keeps inference in Australia. Microsoft also notes that geography-based types arrive last for new models, so the newest model may not be available onshore yet.
Training. Microsoft's Foundry data privacy page says prompts and completions are not available to OpenAI and are not used to train foundation models without your permission. Abuse monitoring data for human review is stored in the resource's geography. Eligible customers can apply for modified abuse monitoring.
What to switch on. Create the resource in an Australian region and choose a Standard or Regional Provisioned deployment. Then use Azure Policy to block the GlobalStandard and DataZone SKUs, as Microsoft's page shows. That stops a developer choosing a cheaper global deployment by accident.
ChatGPT: business plans store data in Australia, the model may still run offshore
Most of the "half the team" risk sits here, and the plan decides it.
| Free, Plus and Pro | Business, Enterprise and Edu | |
|---|---|---|
| Content used to train models | May be, unless the user switches it off | Not by default |
| Who controls the account | The individual staff member | Your workspace admins |
| Australian data residency | Not offered | Enterprise and Edu: available for new workspaces |
| Model processing in Australia | Not offered | Not offered (inference residency is EU, US and UAE) |
| Feedback buttons and training | A thumbs up can share the whole conversation | Feedback does not permit training |
Personal plans. OpenAI's help centre says that for Free, Plus and Pro "we may use your content to train our models." Users can opt out by turning off Improve the model for everyone under Settings > Data controls. The company cannot enforce that setting on a personal account, which is why the OAIC advice above matters.
Business plans. OpenAI's enterprise privacy page says data from ChatGPT Business, Enterprise, Edu and the API is not used for training by default.
Residency. OpenAI's data residency article lists Australia for storage at rest. Eligibility is "new ChatGPT Enterprise/Edu customers". Inference residency is listed only for Europe, the United States and the United Arab Emirates. So an Australian Enterprise workspace stores conversations, files and custom GPTs onshore, while the model may run elsewhere. Account data, billing and some metadata may also sit outside the region. ChatGPT Business is not named as eligible, so check with OpenAI.
The API. OpenAI's API data controls page lists Australia (au.api.openai.com) with storage "Yes" and processing "No". Using any non-US region also requires approval for abuse monitoring controls and a Modified Retention amendment. The region is set per project when the project is created.
For a feature comparison of the three assistants, see ChatGPT vs Claude vs Gemini for business.
Gemini: Vertex AI in Sydney can keep processing onshore, Workspace cannot yet
Gemini for Google Workspace. Google's data regions documentation lets customers store covered data in "the United States or Europe". Gemini prompts and responses are covered at rest, and during processing on the Enterprise Data Regions versions. Australia is not an option. Google's Workspace generative AI privacy hub states that Workspace does not use customer data to train models without prior permission.
Gemini Enterprise. Google's Gemini Enterprise locations page lists US and EU multi-regions and several in-country regions. Australia is not among them.
Gemini on Vertex AI. This is Google's onshore route. The Vertex AI data residency page lists australia-southeast1 (Sydney) for ML processing on Gemini 3.5 Flash and Gemini 2.5 Flash (128k). Data at rest stays in the location you select. Google is explicit that global endpoints "don't provide regional isolation or data residency guarantees".
Training and logging. Google's Vertex AI data governance page says it won't train on your data without permission. Two defaults are worth knowing. Grounding with Google Search keeps logs for up to three days and cannot be disabled. Implicit caching is on by default and can be turned off per project.
What to switch on. Call the australia-southeast1 regional endpoint, pick a model listed for Australia, and leave Google Search grounding off for sensitive work.
Claude: onshore today through Amazon Bedrock, not through the direct API
Personal plans. Anthropic's consumer privacy article says chats on Free, Pro and Max may be used for training if the model improvement setting is enabled. The article does not state the default, so check the setting on each account.
Team, Enterprise and the API. Anthropic's commercial privacy article states that, by default, it will not use inputs or outputs from commercial products to train models.
The direct Claude API. Anthropic's data residency documentation offers two inference geos, "global" and "us", and one workspace geo, "us". There is no Australian option on the direct API today. Anthropic's regional compliance page lists Australia among its Asia-Pacific regions, through AWS Bedrock and Google Vertex. It does not document Australian residency for claude.ai Team or Enterprise, so check with Anthropic before relying on it.
Claude on Amazon Bedrock. For Claude Sonnet 4.6, the model we checked, AWS's model card lists an AU geographic inference profile. Called from Sydney or Melbourne, it routes only to Sydney and Melbourne. In-region inference in Sydney is not offered for that model. AWS's data protection page says model providers have no access to Bedrock logs or to customer prompts and completions.
Two AWS caveats matter. The global profile routes to any commercial region, so use the au. profile ID. AWS also warns not to rely on the prefix alone: call GetInferenceProfile from each source region and check every destination.
Claude on Google Vertex AI. Google's Vertex data residency page lists no Australian location for Claude or other partner models. Check with Google before choosing that route.
How a prompt stays onshore from question to approved answer
Choosing a compliant tool is the first step. The rest depends on the path a prompt takes and who signs off the result.
- Staff ask a questionIn an approved tool, signed in with a work account, never a personal one
- Policy checks the routeOnly the Australian region and the approved deployment type are allowed
- Model runs onshoreInference in Australia East, Sydney or the AU Bedrock profile
- A person approvesNothing reaches a client or a customer decision without review
- Logged in your tenantPrompts and outputs recorded where your retention rules apply
- Kept under retentionEvidence of the route and the approval kept for audit and APP 8 questions
Which route fits the data you handle
Not every task needs onshore inference. A marketing draft with no personal information carries different risk from a patient referral or a client's financial file.
What data will staff put into the tool?
- Public or marketing content, no personal informationA business plan of Copilot, ChatGPT, Gemini or Claude, with training off
- Internal documents and email, already in Microsoft 365Microsoft 365 Copilot with data location confirmed and the web search policy set
- Client or patient records a contract says stay onshoreYour own build on Azure OpenAI Standard, Vertex AI Sydney or Bedrock AU
- Health, legal or government data with strict termsAn onshore build in your tenant, plus a check of the contract and the vendor terms
If the fourth branch is yours, the infrastructure options are covered in AI data sovereignty: where your data goes. The cloud catalogue comparison is in AWS vs Azure vs Google Cloud for AI in Australia.
An approved-tools list your staff will actually follow
Banning AI rarely works when the free tools are a browser tab away. A short approved list, with a sanctioned option for each job, gets followed. One way to sequence it over four weeks:
- Week 1Find what is in useCheck sign-in logs, expense claims and browser data for free AI accounts. Ask teams what they use and for which tasks.
- Week 2Classify the dataSort tasks into public, internal and client or patient data. Match each class to a route from the decision tree.
- Week 3Configure and lock downConfirm the Australian region, block global deployments and personal accounts, set retention and web search policy.
- Week 4Publish the listOne page per tool: what it is for, what data may go in, who approves outputs. Review it whenever a vendor changes terms.
- Storage: the vendor commits to store prompts and files at rest in AustraliaVendor doc
- Processing: the model runs in Australia, not only the storageDeployment type
- Training: business plan with training off by defaultPlan
- Route: staff can only reach the approved tool and regionPolicy
Where Solve8 fits
If a client contract or a health, legal or government data clause says the data stays onshore, an off-the-shelf plan may not be enough. That is when an onshore build in your own tenant makes sense.
We build in the cloud you already run, whether that is Azure, AWS or Google Cloud, in an Australian region. The work starts with one named bottleneck and a baseline. The tool then runs in shadow mode beside your team until it meets an accuracy threshold you set. After that, a person approves every output. No data goes to a model that trains on it, and the code and credentials stay yours.
The lead on every engagement spent 18 years on operations and data systems for BHP, Rio Tinto and Senex Energy, gained while working with previous consulting employers. These companies are not Solve8 clients. See our private AI infrastructure service for how onshore builds are scoped.
Pick the one AI tool your staff use most, and check its plan and region setting against the table above. If the answer is "offshore" and the data is sensitive, book a scoping conversation. You will leave with a straight answer on which route fits your data and what to switch on first.
