TLDR; Claude Code and privacy law meet when client personal data reaches a prompt. Australia, Japan and the UK all regulate sending personal data abroad, and Anthropic’s and OpenAI’s own services process it outside all three. Use a business or API login that comes with a data processing contract, keep sensitive data out, disclose automated decisions, and in the UK assess the impact first. This is general information, not legal advice.
Why client data in Claude Code is a privacy law question
Say an agent scores the leads in an Australian client’s CRM. Each lead’s details go to a model that Anthropic runs outside Australia. The Australian Privacy Principles (APPs) are the rules in the Privacy Act. For a business they cover, sending those details is in most cases an overseas disclosure: the business hands personal information to someone abroad and gives up control of how they handle it.
From 10 December 2026, the client’s privacy policy may also have to describe the kinds of decisions the agent helps make and the kinds of personal information it uses. Run the same agent over a Japanese or UK client’s records and different rules apply.
All three countries regulate sending personal data abroad, and all three expect you to know what the vendor does with a prompt. Beyond that, the duties differ. This guide sets out the law as five duties.
Claude Code data privacy covers what each vendor keeps and where it processes data. When I checked its table on 21 October 2026, Claude on Amazon Bedrock was the only hosted option that kept processing inside these countries.
A prompt that holds personal data is a transfer abroad
Why: each country treats the prompt as personal data leaving the country, under its own test.
- Australia. The Office of the Australian Information Commissioner (OAIC) is the regulator. For an organisation the APPs cover, it treats sending personal information to an overseas provider as a disclosure in most circumstances. APP 8, the rule on cross-border disclosure, then applies. Sending counts as the organisation’s own use, outside APP 8, only where the organisation keeps effective control of how the provider handles the information.
- Japan. The Personal Information Protection Commission is the regulator. Under the Act on the Protection of Personal Information (APPI), Article 28 requires the person’s consent before personal data goes to a third party in a foreign country. The cases in Article 27(1), such as a disclosure the law requires, are exempt. So are a recipient in a country the Commission recognises as equivalent and one that keeps a system meeting the Commission’s standards. In its Q&A, the Commission treats cloud storage as exempt where the provider does not handle the data.
- The UK. The Information Commissioner’s Office (ICO) is the regulator. Since 5 February 2026, the Data (Use and Access) Act 2025 tests a transfer by whether protection is “not materially lower” than in the UK. A data processing addendum is the contract that sets out how a vendor handles personal data for you. Anthropic’s and OpenAI’s addenda rely on the EU standard contractual clauses: model contract terms for sending data abroad. For the UK they attach the UK Addendum, which the ICO issues to adapt those clauses to transfers from the UK.
In practice: a model that reads the prompt handles the data, so I don’t rely on the Japanese cloud-storage exemption.
Do this: treat every prompt that holds personal data as a transfer abroad. Where a client cannot accept that, use Claude on Amazon Bedrock hosted in the country, or a local model that runs on your own machine. For Bedrock, also check which company you contract with and whether its staff or subprocessors can reach the data from abroad.
A consumer login gives you no contract to point to
Why: Anthropic’s data processing addendum is part of its Commercial Terms, and its consumer logins come with no addendum. The UK’s main data protection law is the UK GDPR. Under it, Anthropic’s and OpenAI’s addenda name each vendor as your processor: a party that handles data on your instructions. They are also the written contract Article 28 requires.
The ICO’s generative AI consultation response doubts that developers of closed models (models the developer keeps private) are processors once a business puts the model to use. A controller decides why and how personal data is used; joint controllers decide that together. The ICO calls joint controllership likely and says a contract does not override the facts.
Japan asks a related question. The Commission’s June 2023 caution on generative AI tells a business to confirm that a prompt holding personal information stays within the purpose the data was collected for, and that the provider will not use it for machine learning. If the person did not consent and the provider handles the prompt for anything beyond producing the response, entering it may breach the Act.
In practice: client personal data does not belong on a consumer Pro or Max login (what each login costs). A careful UK client will ask which role the vendor plays: processor or joint controller.
Do this: run client work on a business login (Claude Team, ChatGPT Business or Enterprise) or an API login, the pay-as-you-go access developers use. Keep the addendum on file, and check the provider’s terms on training with your data before the first prompt.
Sensitive data stays out of the prompt without consent or another legal ground
Why: the OAIC’s guidance on commercially available AI products (October 2024, updated January 2025) is plain: “do not enter personal information, and particularly sensitive information, into publicly available generative AI tools.” It ties AI use to two rules. APP 6 limits use to the purpose the data was collected for, and APP 11 requires reasonable security. The guidance also expects you to keep checking the product.
Japan calls its sensitive categories special care-required information. Article 2(3) lists the categories, among them medical history and criminal record. Collecting them generally needs consent. In the UK, special category data, such as health, needs a lawful basis under Article 6 of the UK GDPR and one of the conditions in Article 9. Explicit consent is one of those conditions.
In practice: one of my CRM tools ships a recommended deny list: contact fields the tool never passes to the model. It covers health and disability notes. They are sensitive information under the APPs, special care-required information in Japan and special category data in the UK.
Do this: strip sensitive fields in the tool that fetches the data, before the model sees them (how to gate the data).
Automated decisions now carry their own duties
Why: in Australia, from 10 December 2026, APP 1.7 to 1.9 require a privacy policy to disclose programs that make, or substantially help make, decisions that significantly affect people.
In the UK, Articles 22A to 22D of the UK GDPR, added by the 2025 Act, now govern decisions with no meaningful human involvement. The person must be told about any significant decision made that way. They must also be able to make representations, get a human to intervene and contest the decision. A significant decision based on special category data, such as health, also needs one of the conditions the law lists.
In practice: an agent that scores leads, like the one in the opening example, may count under both tests.
Do this: list each agent that helps decide something about a person; check it against the UK test now and the Australian test before 10 December 2026.
In the UK, assess the impact before the first prompt
Why: the UK GDPR applies in full. In the ICO’s words, “There is no ‘AI exemption’ to data protection law.” The ICO says that “in the vast majority of cases” using AI triggers the legal requirement for a data protection impact assessment. That is a written review of the risks the work poses to people’s data and how you will reduce them. Where you decide it does not, you document how you reached that view.
In practice: my first test for every client is the same. If the work touches personal information, it runs on a business or API login or a local model, through a tool with a written deny list, and the client knows which. In the UK, that answer goes into the impact assessment.
Do this: write the assessment, or the note on why you need none, before personal data reaches a prompt.
When you don’t need it
- No personal data reaches the model. A deny list in the tool keeps it out, and most of the duties above fall away.
- The model runs on your own machine. A local model sends no prompt to a vendor, so there is no vendor transfer or contract to check. If you work from outside the client’s country, the client sending you the data can still be a transfer abroad. The duties on sensitive data, automated decisions and the UK impact assessment still apply.
- A small Australian business may sit outside the APPs. They bind organisations with annual turnover above A$3 million. They also bind some smaller ones, among them health service providers, businesses that trade in personal information and contractors under a Commonwealth contract. Check whether yours is covered. Separately, since 10 June 2025, a person can sue over a serious invasion of privacy under Australia’s statutory privacy tort.
- Japan and the UK have no such threshold. Article 16 of the APPI binds any business that keeps a personal information database, and the UK GDPR applies in full.
Limits
- This is general information, not legal advice. Ask a lawyer in the country before you rely on it.
- The law is moving. The ICO marks its AI guidance as under review after the 2025 Act, so check the current page first. Japan’s 2026 amendment was enacted on 10 July 2026. It adds surcharges equal to the gain from certain breaches, among them improperly handing personal data to a third party. It also adds a consent exemption for statistics that covers some AI development. Most of it takes effect within two years of its official publication.
- Vendor terms move faster. The privacy guide’s table was checked on 21 October 2026; read the vendors’ pages again before you rely on them.
I build these checks into the agents I set up for Applied AI clients in all three countries.
Related reading
- Claude Code data privacy: what Claude and Codex keep, retention by plan, where the vendors process data, and how to gate it
- Token economics, what each login costs
- Building custom MCP servers for your workflow, credential handling for servers that fetch client data
Sources
- Guidance on privacy and the use of commercially available AI products, OAIC, Australian duties when personal information meets AI tools. Regulator guidance
- New resources on transparency for use of AI and automated decision-making, OAIC, privacy-policy duties from 10 December 2026. Regulator guidance
- APP guidelines chapter 8: cross-border disclosure, OAIC, disclosure versus use and effective control (paragraphs 8.12 to 8.15). Regulator guidance
- Small business, OAIC, the A$3 million turnover threshold and who is covered regardless. Regulator guidance
- Statutory tort for serious invasions of privacy, OAIC, in force from 10 June 2025. Regulator guidance
- What is personal information?, OAIC, sensitive information under the APPs. Regulator guidance
- Act on the Protection of Personal Information, Japanese Law Translation, Articles 2(3), 16 and 28. Legislation
- Caution on the use of generative AI services, Personal Information Protection Commission, 2 June 2023, in Japanese. Regulator guidance
- APPI Q&A, Personal Information Protection Commission, Q&A 7-53 and 12-3 on cloud services, in Japanese. Regulator guidance
- 2026 APPI amendment overview, Personal Information Protection Commission, enacted 10 July 2026, in Japanese. Regulator guidance
- Tackling misconceptions and Allocating controllership across the generative AI supply chain, ICO consultation response on generative AI. Regulator guidance
- Accountability and governance implications of AI and Contracts, ICO. DPIAs for AI and the Article 28 contract. Regulator guidance
- Data (Use and Access) Act 2025, sections 80 and 85, legislation.gov.uk, in force 5 February 2026 (SI 2026/82). Legislation
- UK GDPR Article 9, legislation.gov.uk, special category data. Legislation
- Data processing addendum, Anthropic and Data processing addendum, OpenAI, processor role and UK transfer clauses. Terms
Common questions
Can I put client personal information into Claude Code or ChatGPT in Australia?
Only on a business or API login, and with care. The privacy regulator advises against entering personal information, and particularly sensitive information, into publicly available generative AI tools. For a business the Privacy Act covers, sending personal information to an AI provider overseas is in most cases a disclosure, so the cross-border rules apply. Use a business login (Claude Team, ChatGPT Business or Enterprise) or an API login that comes with a data processing contract, and keep sensitive fields out. From 10 December 2026, if you use programs that make, or substantially help make, decisions that significantly affect people, your privacy policy must describe those decisions and the kinds of personal information the programs use.
Does Japan’s privacy law apply to small businesses that use AI tools?
Yes. Japan’s privacy law has no size threshold: it binds any business that keeps a personal information database. Sending personal data to a third party in another country needs the person’s consent unless an exception applies. The regulator tells businesses to confirm that the AI provider will not use prompts for machine learning.
Do I need a data protection impact assessment to use AI tools on UK personal data?
Usually. The ICO says that in the vast majority of cases using AI triggers the legal requirement for one. If you decide it does not, document how you reached that view.
Is Anthropic or OpenAI a data processor under UK GDPR?
On a business or API login, yes on paper. Each vendor’s data processing addendum (its data protection contract) names the vendor as your processor, and that addendum is the written contract the UK GDPR requires. The ICO doubts the label holds for closed models (models the developer keeps private). It thinks the vendor and the business are likely joint controllers, who decide together how data is used, because a contract does not override the facts. A consumer plan has no addendum at all, so keep UK client personal data off it.