AI agents are computer programs that can retrieve information and carry out tasks across connected systems. An AI assistant that schedules a meeting needs access to a calendar. An agent asked to prepare a funding application might retrieve documents, fill in an online portal and email colleagues. That raises two practical questions for any organisation using them: which information should the system be allowed to use, and what should it be allowed to do with it? 

No single global data protection law answers these questions. Each country sets its own rules. International guidelines, such as those from the OECD and UNESCO, are not legally binding, but they set out widely shared principles that many national laws draw on. 

Why is agentic AI a data protection question? 

Established privacy principles require organisations to collect personal data lawfully and fairly, use it only for the purpose they have stated, keep it secure, and be clear about who is responsible for it. AI agents make these principles harder to apply in practice, because once an agent has been given access to personal data, it can find and use that data without a person checking each step. An organisation therefore needs to know which data the agent can reach, whether the task requires that data, and who is accountable if it is misused. 

More than one country’s rules may also apply. For example, a development organisation may work with communities in one country while using an AI provider and cloud services based in others. In a single task, personal data can pass through several companies and jurisdictions, each with its own data protection law. The organisation that deploys the agent remains responsible for how that data is handled. 

What counts as agentic AI? 

An OECD paper published in February 2026 found that most definitions of AI agents share three features: an objective, outputs that usually take the form of actions, and a degree of autonomy. It treats AI agents and the broader category of agentic AI, in which several agents may be coordinated, as related but distinct. [1] 

In practice, an agent might search a database, draft a reply, complete an online form or update a record. Most of the new data protection questions arise at the point where a tool stops suggesting an action and starts acting autonomously. 

Where can personal data be exposed? 

People often assume an AI tool works only with the information they type into it. An agent can reach much more. Connected to email, shared drives or a case-management system, it can retrieve personal information, pass it to other software tools, and keep copies in its activity records (logs) or memory. Stanford University’s IT guidance for its own staff warns that agents inherit the permissions of the person using them, can take actions at scale, and can act on their own errors, for example by submitting the wrong form. [2] 

For instance, if an organisation asks an agent to compile a donor report, the task needs spending totals and anonymised results. Granting access to the entire project folder may be more convenient, as staff do not need to select the relevant files in advance. However, it also allows the agent to read staff records and files that identify programme participants, none of which the report needs.  

Which international principles apply? 

No international principles have yet been written specifically for AI agents. Two general frameworks apply to them, as they do to any system that handles personal data. 

The OECD Privacy Guidelines, adopted in 1980 and revised in 2013, were the first internationally agreed set of privacy principles. They take the form of an OECD Council Recommendation. That means they are not legally binding, but the countries that adhere to them have made a political commitment to implement them, and they have shaped data protection laws both inside and outside the OECD.[3] They set out eight principles. Two are especially relevant to agents: 

  • Purpose specification: an organisation should state why it is collecting personal data, no later than the time it collects it. 
  • Use limitation: it should then use the data only for that stated purpose, unless the person consents or the law allows otherwise. 

For an agent, this means writing down its task before it goes live, then checking that every data source it can reach is needed for that task. The other six principles cover limits on collecting data, data quality, security, openness about data practices, people’s right to see and correct their data, and accountability.[3] 

UNESCO’s Recommendation on the Ethics of Artificial Intelligence, adopted by its member states in November 2021, also applies. It calls for privacy to be protected at every stage of an AI system’s life, from design and development through use to its eventual retirement. It also says it should always be possible to hold an identifiable person or organisation ethically and legally responsible for what an AI system does.[4] Like the OECD Guidelines, it is not binding law, so organisations still need to check the legal requirements where they operate and where the people whose data they hold are located. 

What should organisations check before deploying an agent? 

  • Start with a test run. Stanford’s guidance recommends trying the agent in a test environment (a “sandbox”) before using it on live systems.[2] 
  • Map the task, the tools and the data. Set out which systems the agent can read, which it can change and which actions need a person’s approval. 
  • Limit access from the outset. Stanford’s guidance recommends giving the agent only the access it needs, read-only permissions unless it must edit, its own account rather than a staff member’s login, a person’s confirmation before it acts outside the organisation, and testing on small batches before wider use.[2] 
  • Follow the data through every provider. Where does information go when the agent calls another tool? What is retained, for how long and in which country? Can the team trace an error and stop a workflow? 
  • Ask the companies supplying the agent, and the services it connects to, to show their controls, not just describe them. For example: how access is restricted, how approval is requested, and how activity records are reviewed. 
  • Tell people clearly how their data is used. An October 2026 OECD policy brief points to short explanations at the moment data is collected (not only in a long privacy policy) and simple ways to opt out.[5] Under the OECD Privacy Guidelines, people should also be able to find out what data is held about them and have it corrected or deleted.[3] Agents make this harder. The UK’s data protection regulator warns that the complex data flows and multi-agent interactions of agentic systems may make it harder for people to exercise these rights.[7] 

Why does this matter for developing countries? 

Ministries, development banks and NGOs are adopting agentic tools that are often supplied by foreign companies and run on cloud infrastructure abroad. Where a national data protection authority is new or under-resourced, there is less outside oversight, so the deploying organisation must itself decide what the agent can access, track where data is sent and handle complaints. 

The World Bank now treats cybersecurity, data protection and AI governance as one function, built into digital systems from the start rather than added later. It plans to make this a standing part of its country digital programmes.[6] For governments and organisations whose budgets are already stretched, that means planning for oversight from the outset, including in donor-funded projects: staff who can review the agent’s activity records, approve important actions and respond to complaints. 

In summary 

Wherever an agent is deployed, three questions apply: can the organisation explain the agent’s task, justify its access and show how a person can step in? 

Notes 

  1. OECD, ‘The Agentic AI Landscape and its Conceptual Foundations’, OECD Artificial Intelligence Papers No. 56, 13 February 2026, https://www.oecd.org/en/publications/the-agentic-ai-landscape-and-its-conceptual-foundations_396cf758-en.html, accessed 12 November 2026. 
  2. Stanford University IT, ‘Responsible Agentic AI’, https://uit.stanford.edu/security/agenticai, accessed 12 November 2026. 
  3. OECD, ‘Recommendation of the Council concerning Guidelines Governing the Protection of Privacy and Transborder Flows of Personal Data’, OECD/LEGAL/0188, adopted 23 September 1980, revised 11 July 2013, https://legalinstruments.oecd.org/en/instruments/OECD-LEGAL-0188, accessed 12 November 2026. 
  4. UNESCO, ‘Recommendation on the Ethics of Artificial Intelligence’, adopted 23 November 2021, https://www.unesco.org/en/artificial-intelligence/recommendation-ethics, accessed 12 November 2026. 
  5. OECD, ‘Enhancing User Control and Privacy Protections in the Context of AI’, policy brief, 15 October 2026, https://www.oecd.org/en/publications/enhancing-user-control-and-privacy-protections-in-the-context-of-ai_e0bd9a46-en.html, accessed 12 November 2026. 
  6. World Bank Group, ‘Trust by Design: Securing Digital Development’, 9 June 2026, https://www.worldbank.org/en/results/2026/06/09/trust-by-design-securing-digital-development, accessed 12 November 2026. 
  7. Information Commissioner’s Office, ‘ICO Tech Futures: Agentic AI – Data Protection and Privacy Risks’, 31 March 2026, https://ico.org.uk/about-the-ico/research-reports-impact-and-evaluation/research-and-reports/technology-and-innovation/tech-horizons-and-ico-tech-futures/ico-tech-futures-agentic-ai/data-protection-and-privacy-risks, accessed 12 November 2026.