Claude and the Privacy Act 2020: what a New Zealand business must do

Start from the information, not the tool

The Office of the Privacy Commissioner's starting point is plain: the Privacy Act 2020 applies to everyone using AI tools in New Zealand². Its guidance on artificial intelligence and the information privacy principles, published on 21 September 2023, walks through how each principle applies across the life of an AI system, including generative tools such as Claude³.

For an owner or manager, that means the useful question is not whether Claude is allowed. It is which personal information will go in, why, who will see what comes out, and where it will be stored afterwards. A team that summarises supplier contracts has a very different privacy profile from one that drafts responses to customer complaints, even if both use the same Claude Project.

The obligations stay with you. When your business uses Claude under the Commercial Terms, Anthropic acts as your processor under its Data Processing Addendum, and you remain the controller^10,11. Choosing a reputable vendor is part of meeting the Act. It does not hand the duty to someone else.

The Commissioner's expectations, applied to Claude

The Commissioner has set out what organisations should do before and while they use generative AI³. Each expectation maps onto a concrete step in a Claude rollout.

• Get senior leadership approval, with the privacy officer in the room. In Claude terms, the person who becomes Primary Owner in Organization settings should not be the only person who decided to buy it³.
• Ask whether the use is necessary and proportionate. Drafting a policy from your own templates needs no personal information at all; screening complaints might³.
• Do a privacy impact assessment before rollout, and ask the provider how privacy is built in. Anthropic's Data Processing Addendum and data usage pages answer most of the vendor questions^3,10,14.
• Be open with customers and staff about how Claude is used, especially where it drafts something they will receive³.
• Seek the views of people affected, including Māori, when the use touches decisions about them³.
• Keep a person reviewing accuracy before anything Claude produced is acted on³.
• Make sure the provider does not keep or disclose personal information beyond what you agreed. That is a plan and settings question, covered below^3,13.

The principles you will meet in everyday Claude use

Principle 1 says you collect personal information only for a lawful purpose connected with your business, and principle 10 limits using it for something else¹. If a client gave you their file to prepare a tax return, pasting it into Claude to prepare that return is a related purpose. Pasting it in to build a marketing persona is not.

Principle 3A is new. Since 1 May 2026, an agency that collects personal information from someone other than the person concerned must take reasonable steps to make that person aware of matters such as the fact of collection, the purpose and the intended recipients, unless an exception applies⁶. This matters for Claude because web search and Research can gather information about people from public sources. If staff use Claude to research a job applicant, a debtor or a complainant, treat it as indirect collection and check whether a notice is needed⁶. One exception covers information that is already publicly available, such as in a newspaper, a public register or on a website⁶.

Principle 5 requires reasonable security safeguards¹. In Claude this means a commercial plan, single sign-on, sensible connector permissions and a clear rule on what never goes into a prompt. Principle 8 requires reasonable steps to check information is accurate before you use it¹. Claude can be confidently wrong, so any reply that describes a person, a debt or an incident needs checking against the source before it lands in a letter or a record.

Principle 12 deals with disclosure outside New Zealand⁵. The Commissioner's guidance explains that, under section 11 of the Privacy Act 2020, an overseas provider that only stores or processes information on your behalf, and does not use it for its own purposes, is treated as your agent, so it is not a disclosure, but you remain responsible for what that provider does with it⁵. Anthropic contracts with New Zealand customers as Anthropic PBC under California law, and its privacy policy describes transfers to the United States^11,12. Record that in your privacy impact assessment rather than leaving it implied.

Plan choice is a privacy decision

The most important control is the one made at purchase. On Claude's consumer plans, Free, Pro and Max, each person decides whether their chats are used to improve Anthropic's models, and data shared for that purpose can be kept for up to five years¹³. On Team, Enterprise and the API, Anthropic may not train models on customer content¹¹.

A business that lets staff put client or employee information into personal consumer accounts will struggle to show it took reasonable security steps under principle 5 of the Privacy Act 2020^1,13. Move work onto a Team or Enterprise organisation first, then write the rules.

Retention also needs a decision. On Enterprise, owners set a custom retention period under Organization settings, then Data and Privacy, with a 30-day minimum; without one, chats and projects are kept indefinitely¹⁵. Incognito chats are not saved to history or memory and are not used for training, yet they are still kept for 30 days and appear in organisation exports¹⁶. Tell staff that incognito is not a way around the rules.

Privacy officer, breaches and your vendor

Any business that handles personal information must have a privacy officer, whose job includes encouraging compliance, handling access and correction requests, and working with the Commissioner during an investigation⁷. Give that person a seat in the Claude decision and a copy of the settings you chose.

If a privacy breach has caused or is likely to cause serious harm, you must notify the Commissioner and the affected people as soon as practicable⁴. The Commissioner's online NotifyUs tool walks you through whether a breach is notifiable⁴. Typical Claude-related incidents are an export of chats sent to the wrong person, a connector given wider access than intended, or client data pasted into a personal account that is later compromised.

Your contract helps here. Anthropic's Data Processing Addendum commits it to notify you of a security breach without undue delay and in any event within 48 hours¹⁰. Put that clause, and the name of whoever receives Anthropic's notices, into your incident plan.

Step by step: a privacy-ready Claude rollout

Work through these steps in order. Allow an hour or two for most steps; the assessment takes longer if Claude will touch health, financial or employment information.

• List who already uses Claude and on which plan, including personal accounts paid for on company cards¹³.
• Choose Team or Enterprise for any work involving personal information, so the Commercial Terms apply¹¹.
• Write a short privacy impact assessment: the purposes, the data classes, the overseas processing, and who reviews output^3,5.
• Sort information into three groups: fine to use, use only with care in a named Project, and never paste in.
• Set Project instructions that remind Claude, and the user, of the review step for anything about a person.
• On Enterprise, set retention to match your records policy¹⁵.
• Check whether Claude-assisted research about people triggers a principle 3A notice under the Privacy Act 2020⁶.
• Update your privacy statement to say you use AI tools to help process information, and name the purposes³.
• Train staff on the rules with their own examples, then review after the first month.

Worked example: a recruitment agency

Consider a generic Christchurch recruitment agency with around thirty staff that wants Claude to summarise CVs and draft shortlist notes for clients. The purpose is clear and connected to the business, so principle 1 is met if candidates were told their CVs would be assessed for roles¹. The agency moves everyone onto Team, sets up a Project per client role, and writes a Project instruction that bans guesses about age, ethnicity, health or family status.

Two risks remain. First, accuracy: a shortlist note that misstates a candidate's experience could cost them a job, so a consultant checks every note against the CV before it goes to the client^1,3. Second, indirect collection: if consultants ask Claude to search the web about a candidate, that is information gathered from a source other than the candidate, and principle 3A may require the agency to tell them⁶. The agency decides web search stays off in that Project and records why in its privacy impact assessment.

Staff, consultation and the public sector

Bringing Claude into a team changes how people work. Where a change could affect jobs, employers need genuine business reasons, must consult the employees affected and must follow a fair process; improved technology is one of the reasons Employment New Zealand lists⁹. Even when no roles are at risk, telling staff what admins can see, such as Enterprise audit logs and exports, is part of acting in good faith.

If you supply government agencies, expect them to ask how you use generative AI. The Government Chief Digital Officer's guidance for the public service covers governance, security, procurement, privacy and transparency for those tools⁸. MBIE's guidance for businesses points the private sector at the same building blocks: clear accountability, a privacy officer and risk-based governance⁷.