To keep your data secure when you start using AI, do three things: put everyone on a business plan so your data isn't used for training, decide which data can go in, and use zero data retention for the work that can't be stored. Most businesses need the first two, which take an afternoon. If you need zero data retention in Australia, budget about a day of a developer's time and roughly 10% more per token than the global price.
Zero data retention, or ZDR, means the AI provider doesn't store your prompts and responses once it has answered. It's separate from training. Turning training off stops your chats improving future models, and your chats still get stored.
The defaults are easy to get wrong. Anthropic's API, for example, keeps inputs and outputs for up to 30 days unless you have a ZDR agreement.
What does Anthropic store on each Claude plan?
Anthropic's storage depends on which product you use. Every table in this playbook is as of November 2026. These terms change, so use the links under each table to check the provider's current wording before you rely on it.
| Where you use Claude | Used for training? | What gets stored | Zero data retention? | Can processing stay in Australia? |
|---|---|---|---|---|
| Free, Pro, Max, Ultra | Your choice, in settings | Up to 5 years if training is on. With it off, deleted chats go within 30 days | No | No option listed |
| Team | No, unless you opt in | Chats stay until someone deletes them, then leave Anthropic's systems within 30 days | No | No option listed |
| Enterprise | No, unless you opt in | Kept indefinitely unless an admin sets a period. The shortest is 30 days | Not for chat. Claude Code only, with approval | No option listed |
| Claude API | No, unless you opt in | Inputs and outputs deleted within 30 days | Yes, with Anthropic's approval | No. Requests are processed in the US or globally |
| Claude on Amazon Bedrock, retention set to none | No | Nothing written to durable storage by AWS | Yes, you set it yourself | Yes, between Sydney and Melbourne |
As of November 2026. Check the current terms: API and data retention, business plan storage, consumer plan storage, Enterprise retention controls, data residency, Amazon Bedrock data retention.
Three things apply across Anthropic's own products, even on ZDR or with training off:
- If automated safety systems flag a chat, Anthropic may keep the inputs and outputs for up to 2 years.
- Anthropic keeps what the law requires it to keep.
- Anthropic may keep trust and safety classification scores longer than the chats themselves.
The thumbs up and thumbs down buttons have their own rule. If someone on a Claude business plan presses one, Anthropic keeps the data tied to that feedback for 5 years, on a separate clock from your chats. Tell your team not to press them on anything sensitive.
What do ChatGPT, Microsoft Copilot and Gemini store?
The other big providers follow the same pattern: personal plans are the risky ones, business plans don't train by default, and zero data retention is usually an API feature. The Australian column is where they differ most. Storing data in Australia and processing it in Australia are two separate promises, and several providers only make the first.
What OpenAI stores on each ChatGPT plan
| Where you use it | Used for training? | What gets stored | Zero data retention? | Can it stay in Australia? |
|---|---|---|---|---|
| ChatGPT Free, Plus, Pro | Yes, unless you opt out under Settings, Data controls | Chats stay until you delete them, then go within 30 days | Not listed | No option listed |
| ChatGPT Business | No, by default | Admins set how long. Deleted chats go within 30 days | Not listed | No option listed |
| ChatGPT Enterprise | No, by default | Admins set how long. Deleted chats go within 30 days | Not listed | Storage yes, for new workspaces. Processing no |
| OpenAI API | No, by default | Inputs and outputs for up to 30 days | Yes, for eligible endpoints, with OpenAI's approval | Storage yes, with approval. Processing no |
As of November 2026. Check the current terms: enterprise privacy, how your data is used for training, chat retention, ChatGPT data residency, API data controls.
What Microsoft stores for each Copilot product
| Where you use it | Used for training? | What gets stored | Zero data retention? | Can it stay in Australia? |
|---|---|---|---|---|
| Copilot in Microsoft 365 apps for home | Not used to train foundation models | Activity history, which you manage in your privacy dashboard | Not listed | No option listed |
| Microsoft 365 Copilot for business | Not used to train foundation models | Chats sit in Exchange mailboxes, for as long as your admin's Purview policy says | Not listed | Storage yes, with the paid Advanced Data Residency add-on. Processing expected by the end of 2026 |
| Azure OpenAI in Microsoft Foundry | No, without your permission | Flagged prompts may be stored for human review | Close to it: approved customers can apply to switch that storage off | Regional deployments process in the geography you pick, so check your model is offered in Australia |
As of November 2026. Check the current terms: Microsoft 365 Copilot privacy, Copilot for home privacy, Copilot retention, Advanced Data Residency, in-country processing timeline, Azure OpenAI data privacy, Foundry deployment types.
What Google stores for each Gemini product
| Where you use it | Used for training? | What gets stored | Zero data retention? | Can it stay in Australia? |
|---|---|---|---|---|
| Gemini app, personal account | Yes while Keep Activity is on. You can turn it off | 18 months by default. With Keep Activity off, 72 hours. Chats sent to human reviewers, up to 3 years | Not listed | No option listed |
| Gemini in Google Workspace | No, without your permission | Admins set how long. The Gemini app default is 18 months | Not listed | No. Gemini data regions are the EU and US only |
| Gemini on Google Cloud (Vertex AI) | No | Caching is on by default. Flagged prompts for up to 90 days, unless your Cloud Master Agreement exempts you | Yes, by configuration: turn off caching and logging. Google Search grounding still keeps data for up to 3 days | Yes for some models, including Gemini 3.5 Flash, in australia-southeast1 |
As of November 2026. Check the current terms: Gemini Apps privacy, Workspace privacy hub, Gemini data regions, Google Cloud data governance, Google Cloud data residency, Vertex AI abuse monitoring.
Can an older or cheaper plan get you around AI data storage?
No. Storage follows the product and the contract behind it, not the age or price of your plan. Anthropic lists consumer plans and Team and Enterprise chat as not eligible for ZDR, and I couldn't find a legacy plan that's an exception.
Turning training off on a Pro account doesn't get you there either. It stops training and leaves you on consumer terms, outside ZDR.
The model you pick does matter. Anthropic's Fable and Mythos models require 30-day retention and aren't available under ZDR unless Anthropic expressly authorises it. So if you have ZDR on the API, using one of those models takes you outside it, and using another Claude model keeps you covered, as long as you stick to ZDR-eligible features. Changing your chat plan, whether older, cheaper or more expensive, won't get you ZDR at all.
What you need before you secure your AI setup
- A list of the kinds of data your team handles: client files, staff records, contracts, anything under NDA.
- Your NDAs and client contracts, and someone who can read them.
- An admin who can see which AI accounts your staff use.
- For Step 4 only: an AWS account and a developer for a day.
Step 1: Sort your data into three piles before anyone types it in
Spend an hour and put every kind of data in one of three piles.
- Fine on a business plan. Marketing copy, internal process docs, public research.
- Needs controls. Client documents, financials, anything with personal information.
- Can't be stored by anyone else. Material under strict NDA, anything where a leak ends up in front of a lawyer.
There are two separate confidentiality problems here. One is between people inside your company: who can see what once a tool is connected. The other is your company's data reaching outsiders. A retention setting only helps with the second. The first is about permissions in whatever tool you connect.
The usual failure is skipping this and buying the most expensive option for everything, or the cheapest for everything.
Step 2: Get everyone off personal AI accounts
Move the whole team onto a business plan and make it the only approved way to use AI at work. This one step gets you off consumer terms and turns training off by default.
The Office of the Australian Information Commissioner recommends as best practice that you don't enter personal information, and particularly sensitive information, into publicly available generative AI tools. A staff member's personal account is that kind of tool.
What good looks like: nobody at work is using a personal login for company data. The usual failure is a free account that stays in use on someone's phone.
Step 3: Check what your AI plan actually keeps, and for how long
For each AI tool your team uses, refer to the tables above, or look up the options on the provider's own site if your tool isn't listed. Check what it stores and for how long, and whether that's acceptable for each of your three piles.
If you're on an Enterprise plan, your admin can set how long chats are kept. Claude Enterprise keeps them indefinitely unless someone changes it.
If nothing landed in your third pile, you can stop here.
Step 4: Set up zero data retention in Australia
If your third pile has something in it and it needs to stay onshore, this is the route we could document for current Claude models. As of November 2026, Anthropic's own API offers ZDR through its sales team, but it can only process requests in the US or globally. Claude on Amazon Bedrock has an Australian inference profile that routes only between Sydney and Melbourne, and AWS says it doesn't share your inputs and outputs with model providers.
On Bedrock, AWS is the processor, so you set retention yourself rather than signing an Anthropic ZDR agreement.
First check that your model allows it. Look for none in allowed_modes:
awscurl --service bedrock-mantle --region ap-southeast-4 \
https://bedrock-mantle.ap-southeast-4.api.aws/v1/models/anthropic.claude-opus-5-5
Then set the mode to none in both Regions, following AWS's data retention guide. The setting is per Region and doesn't carry across, and there's no console option for it yet.
for region in ap-southeast-2 ap-southeast-4; do
awscurl --service bedrock --region "$region" -X PUT \
"https://bedrock.$region.amazonaws.com/data-retention" \
-H "Content-Type: application/json" \
-d '{"mode":"none"}'
done
If you call the bedrock-mantle endpoint, it has its own account setting. It runs in Melbourne (ap-southeast-4) only, so set that one too:
awscurl --service bedrock-mantle --region ap-southeast-4 -X PUT \
https://bedrock-mantle.ap-southeast-4.api.aws/v1/data_retention \
-H "Content-Type: application/json" \
-d '{"mode":"none"}'
Call the model through the Australian profile, and confirm the profile only goes to Sydney and Melbourne:
aws bedrock get-inference-profile --region ap-southeast-2 \
--inference-profile-identifier au.anthropic.claude-opus-5-5
Then close the gaps:
- Don't use a
global.model ID. Global profiles can route to any AWS commercial Region. - Pin Claude Code to Australia. Set
ANTHROPIC_DEFAULT_OPUS_MODEL,ANTHROPIC_DEFAULT_SONNET_MODELandANTHROPIC_DEFAULT_HAIKU_MODELtoau.model IDs. Regions starting withap-otherwise prefer the wider Asia-Pacific prefix, and a prefix setting alone is a preference, not a guarantee. - Lock it in. A Service Control Policy can deny any retention mode other than
none. Cover both thebedrock:andbedrock-mantle:actions. - Leave invocation logging off, which is the default, or treat those logs as sensitive data.
- Decide where chat history lives. Bedrock keeps none, so if your app saves conversations, your database is now the thing that needs protecting.
Then read the setting back in each Region. Both should say none:
for region in ap-southeast-2 ap-southeast-4; do
awscurl --service bedrock --region "$region" \
"https://bedrock.$region.amazonaws.com/data-retention"
done
The usual failure is expecting your team to keep logging in to Claude as normal. They can't on this route, because claude.ai and the standard Claude app run on Anthropic's own systems. They need a chat app that connects to Bedrock instead. There are two common options:
- The Claude desktop app in third-party mode. It sends your prompts to Bedrock and keeps chat history on each person's computer. It still contacts Anthropic to download its workspace and CLI, and for updates and analytics unless your admin turns them off. Confirm it's using the
au.profile. - A self-hosted chat app such as LibreChat. It runs in your own AWS account, and the chat history sits in your database.
Claude Code also works on Bedrock. Anthropic's server-side web search and code execution tools aren't available on Bedrock, so you'd add your own search. The desktop app and Claude Code still run code locally.
Step 5: Stop AI agents leaking through their web searches
Zero data retention stops at the model provider. Anthropic's docs say data processed by third-party websites, tools or integrations isn't covered. The moment an agent searches the web, something leaves.
| Who | What they can log | Does ZDR help? |
|---|---|---|
| The model provider | Your prompt and documents | Yes, this is the only layer it covers |
| The search provider | The full search query | Not documented |
| The website | The full address requested, the time, the IP address | No |
Claude's built-in web search sends queries the model writes to outside search providers. As of November 2026, Anthropic lists Brave Search, TurboPuffer and Parallel as subprocessors for it. Brave's own API notice says query logs are kept for up to 90 days unless an enterprise client has ZDR, and Anthropic doesn't publish which terms apply to its traffic. Microsoft is upfront about the same gap: Copilot's web queries go to Bing and its data protection addendum doesn't apply to them, though Microsoft says they aren't used to train its models.
Here's a made-up example. You're quietly planning to close a depot and ask an agent to work out the cost. The agent searches Acme Freight Dandenong depot closure redundancy costs. Your ZDR is working perfectly, and the search provider has just received your plan in one line.
A single search on a public company name is no different to someone Googling it. The exposure is a query, or a run of queries, that joins the name to what you're planning.
A prompt telling the agent to be careful isn't enough on its own. In the MosaicLeaks benchmark, privacy prompting "reduces but does not eliminate leakage". It's a controlled test and doesn't measure real deployments. In 2025, researchers demonstrated that a PDF with hidden instructions could get a Notion agent to put private client data into a web search and send it to an attacker's server.
So separate the jobs:
- The agent holding your confidential material has no internet access.
- A second agent does the web lookups. It only ever receives single public facts, and it has nothing confidential to give away.
- Every outbound request goes through one gateway that logs it, so you can answer "where did this agent go and what did it send?"
Simon Willison's framing is the one to remember: private data, untrusted content and the ability to communicate externally should never sit in the same agent.
How you know your AI data security is working
- Nobody uses a personal AI account for company data.
- Every tool has a written line on what it stores and for how long.
- The retention setting reads
nonein both Sydney and Melbourne. - The Australian profile lists only Sydney and Melbourne.
- You can pull a log of every outbound request an agent made.
Where zero data retention does not protect you
- Flagged content. ZDR has exceptions for safety flags and legal requirements, on Anthropic's own service. AWS also keeps a narrow exception for apparent child abuse material in images, and prompt caches you switch on hold content for 5 minutes or an hour after last use.
- It's a promise. ZDR is a contractual commitment. I found no independent audit that proves deletion.
- Your own systems. Your database, logs and backups are outside every setting in this playbook.
- People. Someone pasting a client file into a personal account on their phone sits outside all of it.
- Your contracts. Whether a particular NDA allows AI processing at all is a question for your lawyer. This playbook explains how the technology works and isn't legal advice.
Before you rely on any of this for something that matters, ask your provider in writing which products are covered and exactly what goes to its search partners.
Related reading

Written by
Anna SekulicAI Specialist
Anna Sekulic is an AI Specialist at Hourglass AI, where she runs the AI audits and team workshops. She has been building AI workflows since she was 16, with three years of hands-on experience designing agent systems for real businesses, and she shipped her own iOS app with AI tools alone.
More from AnnaCommon questions
What is zero data retention?
Zero data retention, or ZDR, means the AI provider doesn't store your prompts and responses after it has answered. It covers the provider only. Safety flags and legal requirements are still exceptions, and web searches an AI agent runs go to other companies, so check what those companies log.
Does turning off AI training stop the provider storing my data?
No. Training and storage are separate. Turning off training stops your chats being used to improve future models. Your chats are still stored so you can see your history, and a business plan doesn't change that.
Can you get zero data retention on a Claude Team or Enterprise plan?
Not for chat. Anthropic lists the Team and Enterprise chat interfaces as not eligible for ZDR. Enterprise lets an admin set a retention period of 30 days or more. Claude Code on Enterprise has its own ZDR offering, subject to approval.
How do you get zero data retention for Claude in Australia?
Use Claude on Amazon Bedrock through the Australian inference profile, which routes only between Sydney and Melbourne, and set the data retention mode to none in both Regions. Anthropic's own API offers ZDR, but it can only process requests in the US or globally.
References
- 1. Anthropic: API and data retention: What ZDR covers, what it excludes and which products are eligible. Read 10 November 2026
- 2. Anthropic: How long do you store my organisation's data?: Retention on business plans and the API
- 3. Anthropic: Data residency: The processing locations the Claude API offers
- 4. AWS: Data retention in Amazon Bedrock: The retention modes and how to set them
- 5. AWS: Claude Opus 5.5 on Amazon Bedrock: The Australian inference profile and its destination Regions
- 6. OAIC: Guidance on privacy and the use of commercially available AI products: The Australian regulator's advice on personal information in public AI tools
- 7. OpenAI: Enterprise privacy: Training and retention on ChatGPT business plans and the API
- 8. Microsoft: Manage public web access in Microsoft 365 Copilot: How Copilot's web queries are handled
- 9. Brave Search API privacy notice: Query log retention at a search provider
- 10. MosaicLeaks: Privacy Risks in Querying-in-the-Open for Deep Research Agents: A benchmark of what research agents leak through their web queries
- 11. CodeIntegrity: The hidden risk in Notion 3.0 AI agents: A demonstrated data leak through an agent's web search tool
- 12. Google: Generative AI in Google Workspace Privacy Hub: How Workspace handles customer content
- 13. Anthropic Trust Center: Subprocessors: The companies Anthropic lists for web search
- 14. Simon Willison: The lethal trifecta for AI agents: Why private data, untrusted content and outbound access shouldn't share an agent










