What the regulators agreed in Brussels
Privacy and data protection authorities met in Brussels from 16 to 19 November for the 48th Global Privacy Assembly annual conference. The closed members' session ran on 16 and 17 November. The GPA's own release says more than 200 members took part in that session. The list of adopted resolutions has five entries. The Resolution on Privacy-protective Agentic AI sits alongside resolutions on AI in education, privacy-enhancing technologies, data sharing for anti-money-laundering work and a new GPA Practice Hub. Canada's Privacy Commissioner Philippe Dufresne chairs the Assembly and wrote the agentic AI resolution, his office said on 19 November.
The resolution covers AI agents and agentic systems that can process personal information and make automated decisions, and the published summaries put human oversight at its centre. In his speech to the conference on 18 November, Dufresne said the resolution asks organisations to weigh privacy risks when they design or use agentic systems, with particular attention to risks that come from autonomous actions. MLex reported that regulators from more than 80 countries adopted the text as a non-binding standard. According to MLex, the resolution limits how agents handle personal data and calls for ongoing testing and real-time monitoring. Members also approved a strategic plan for 2027 to 2030 that makes the privacy impact of technology a shared priority and promises more year-round cooperation between authorities.
Why a non-binding resolution deserves attention
GPA resolutions carry no legal force in New Zealand. Regulators use them to signal where their attention is heading, and the positions agreed at the Assembly tend to show up later in national guidance and in the questions investigators ask after a complaint. New Zealand's Office of the Privacy Commissioner lists the Assembly among its international privacy networks. At the 2025 meeting in Seoul, the NZ office was one of 19 authorities that signed a joint statement on trustworthy data governance for AI. At the time, Commissioner Michael Webster said AI done well depends on privacy done well. An agreed international position on agents gives the NZ office a ready reference when it next revises its own AI guidance.
The OPC's guidance is from September 2023. The OPC's AI page expects a privacy impact assessment before an AI tool goes live, with regular updates after that. It also expects accuracy and fairness testing under IPP8, protection of prompts and training data under IPP5, and a check on whether reusing personal information fits the purpose it was collected for. The page ends with blunt advice: "If in doubt, we recommend you do not use AI tools to handle personal information."
The OPC wrote those expectations with chat assistants and generative tools in mind. An agent that opens a customer file, checks payment history, approves a refund and emails the customer engages several privacy principles in one run, and usually nobody sees the steps in between. The Brussels resolution is the first time privacy regulators have collectively treated that kind of autonomy as a privacy risk of its own. We expect the NZ office's questions about an agent to move beyond what data went into the tool and towards what the agent was allowed to do with that data.
The Australian deadline arrives first
NZ firms with an Australian presence face a firmer date. From 10 December 2026, APP entities must say in their privacy policy when a computer program uses personal information to make a decision that could significantly affect someone's rights or interests. The rule also applies when the program does something substantially and directly related to that decision. The policy has to describe the kinds of personal information used. It also has to separate decisions made solely by software from decisions where software does part of the work. The OAIC released its fact sheets and updated APP 1 guidelines on 30 September, after receiving 90 written submissions. Allens notes that the obligation can reach well beyond AI and catch decision-support tools and software built into other products. An agent that approves credit or screens job applicants clearly falls within the rule. A New Zealand company with Australian customers or staff should confirm whether the Australian Privacy Act covers it before 10 December.
What NZ businesses should do now
Most mid-sized organisations already run AI agents somewhere, often as features inside Microsoft 365, a CRM, a help desk platform or a finance package. Few have written down what personal information those agents can reach or what they're permitted to do with it. Because the resolution focuses on autonomous action, an inventory of agents and their permissions is the sensible place to start. Practical steps:
- List every agent. Include vendor features that were switched on by default and agents staff have built themselves in tools such as Copilot Studio or Power Automate.
- Record each agent's reach. Note the personal information it reads, the systems it can write to, and the actions it can take without a person approving them.
- Run or update a privacy impact assessment. Do one for every agent that touches customer or staff data, use the OPC's 2023 guidance as the baseline, and add autonomous actions as a separate risk line.
- Set approval thresholds. Require sign-off for actions that are hard to reverse, such as refunds, account changes, deletions and messages sent to customers.
- Keep logs. Capture what each agent saw and did, and keep the logs long enough to answer a privacy complaint or an IPP6 access request.
- Ask vendors in writing. Find out how agent features can be limited, monitored and switched off, and where agent activity logs are stored.
- Update the Australian privacy policy. If you operate in Australia, map agent-driven decisions against APP 1.7 and update the policy before 10 December.
The resolution's full operative text is published as a PDF on the GPA site. The NZ office hasn't said whether it'll update its 2023 guidance in response. Firms that complete the inventory and assessments now won't need to rush when it does. Our AI consultancy page sets out how we approach AI governance. If you're building agents in-house, our AI agents page covers the design questions around permissions and oversight.
Sources
- 48th GPA Annual Conference, Global Privacy Assembly
- International authorities gather at the 48th Global Privacy Assembly annual conference, Global Privacy Assembly
- Adopted resolutions, Global Privacy Assembly
- News release: Privacy Commissioner of Canada at the 48th Global Privacy Assembly, OPC Canada
- Remarks by the Privacy Commissioner of Canada to the GPA's 48th Annual Conference, OPC Canada
- Global Privacy Assembly issues guidance for agentic AI & PETs, MLex
- Global Privacy Assembly adopts AI agent resolution, Digital Watch
- International privacy networks, Office of the Privacy Commissioner NZ
- Joint statement on trustworthy data governance for AI, Office of the Privacy Commissioner NZ
- AI, Office of the Privacy Commissioner NZ
- New resources on transparency for use of AI and automated decision-making, OAIC
- Mandatory automated decision-making disclosure requirements are coming, Allens
Want to know what this means for your organisation? Our team in Auckland can show you your top three opportunities in a free 30 minute assessment.
Get my free assessment How we help