Important Notice: Beware of Fraudulent Websites Misusing Our Brand Name & Logo. Know More ×
Oracle Partner logo

Agentic AI and Data Protection: How to Map What Your AI Agents Do With Personal Data

Agentic AI and Data Protection

Key takeaways

  • The UK ICO opened a call for evidence on agentic AI and data protection on October 8, 2026. It is not new law, but it signals where regulators want input.
  • Agents differ from chatbots: they read several systems, call tools, pass data to model providers and take actions, so a one-time system review can miss data flows.
  • Keep one short data-flow record per agent: purpose, data in, data out, actions, roles, retention, human involvement and owner.
  • Walk each record against familiar GDPR principles: purpose limitation, minimization, lawful basis, automated decisions, security, processors and impact assessment.
  • Check the ICO page for the current closing date and questions before responding. This is practitioner guidance, not legal advice.

On October 8, 2026 the UK Information Commissioner’s Office opened a call for evidence on how data protection law applies to agentic AI (Squire Patton Boggs summary). Secondary reports put the closing date at November 20, 2026, so check the ICO’s own page for the current date and questions.

A call for evidence is not guidance, and it does not create new rules. It does show where a regulator wants input, and it gives organizations a reason to answer a practical question now: do we know what our agents do with personal data? This article describes a way to find out, using the GDPR principles most teams already work with. It is a practitioner perspective, not legal advice.

Why agentic AI changes the data protection picture

A chatbot answers a question. An agent plans, calls tools, reads systems and takes steps, often across several applications in one task. That raises data-protection questions that a one-time system review can miss:

  • The agent may combine data from systems that were never designed to be read together.
  • It may pass personal data to a model provider, a search tool or another agent.
  • It may take actions that affect people, such as sending a message or updating a record.
  • Its behavior can vary between runs, so what it touched last week may differ from today.

Map How AI Agents Process Personal Data

Identify what personal data AI agents access, retrieve, share, transform, and store across enterprise applications and connected systems.

How to build an AI agent data flow map: one record per agent

Start with a data-flow record for each agent. Keep it short enough that teams will maintain it. If you do not yet have a list of your agents, start with our guide to building an AI agent inventory.

Where personal data moves around an AI agent
Figure 1. Where personal data can move around an agent, and the check that applies to each flow.
Question What to record
Purpose What task is the agent for, and who asked for it?
Data in Which systems and fields can it read? Does that include special category data?
Data out Where do prompts, tool inputs and outputs go: model provider, logs, other tools, other agents?
Actions What can it change or send? Do any actions affect individuals?
Roles Who is the controller and who are processors for each flow, including the model provider?
Retention How long are prompts, outputs and logs kept, and where?
Human involvement Where does a person review or approve before an effect on an individual?
Owner Which named person is accountable?
Eight questions to answer for each AI agent: purpose, data in, data out, actions, roles, retention, human involvement, owner.
Figure 2. A one-page record template. Keep it short enough that teams will maintain it.

Check each agent against GDPR data protection principles

Purpose limitation and minimization. Check each field the agent can read against the task. Agents are often given broad read access because it is convenient. Narrow it to what the purpose needs, and remove access to fields the agent never uses.

Lawful basis and transparency. Confirm the basis for each processing purpose and that privacy notices describe agent-based processing in plain terms. If an agent is new, the notice may predate it.

Automated decisions. Where an agent’s output feeds a decision about a person, check whether the decision-making rules for solely automated decisions could apply, and where meaningful human review sits. A reviewer who approves every output in seconds may not amount to meaningful review. See our design patterns for approval gates in agentic workflows.

Security. Treat agent credentials as privileged (see agent identity and permissions). Limit tool access, log tool calls, and test what happens when untrusted content, such as an inbound email or web page, contains instructions aimed at the agent. Our post on prompt injection in enterprise AI systems covers this in more detail.

Processors and transfers. Review contracts and settings for each model and tool provider the agent uses: whether prompts are retained or used for training, where processing happens, and what sub-processors are involved.

Accuracy and individual rights. Decide how you would find and correct personal data an agent has written into a record, and how you would respond to an access request that covers agent logs.

Impact assessment. For agents that process personal data at scale or affect individuals, a data protection impact assessment is likely to be relevant. The data-flow record gives you most of the inputs, and an enterprise generative AI risk register is a useful place to record the result.

Worked example: mapping personal data flows for a support agent

This is a hypothetical example, not client work. A retail customer-service agent reads order history and drafts refund responses. The map shows it can also read full customer profiles, including fields it never uses, and it sends full conversation text to an external model provider. Two fixes follow: restrict the agent to the order fields it needs, and configure the provider relationship and retention settings to match the purpose. A human approves refunds above a set value. The map made those gaps visible before anyone raised a complaint.

Should your organization respond to the ICO call for evidence on agentic AI?

Organizations that deploy agents are well placed to give practical evidence: where roles are unclear, where controls are hard to apply, and what guidance would help. Check the ICO page for the questions and format, and involve your DPO and engineering leads in the answer.

Strengthen Personal Data Protection Across AI Workflows

Establish visibility into agent permissions, data access paths, processing activities, and third-party integrations to reduce privacy exposure.

Conclusion

Agentic AI does not rewrite data protection law, but it changes what a review has to cover: the agent, the tools it can call and the personal data that moves between them. A data flow map with one record per agent gives privacy, security and engineering teams a shared starting point, and walking that map against familiar principles shows where purpose, minimization, retention and access need attention. Build the map for your highest-risk agents first, then extend it. This article is practitioner guidance, not legal advice, so confirm requirements with your counsel and data protection officer. For help designing agents with these controls in place, see our AI agent development and AI consulting services.

Agentic AI and data protection: frequently asked questions

Does the ICO call change what the law requires?

No. It gathers evidence on how existing data protection law applies to agentic AI.

Does this apply in the EU as well?

The GDPR principles above are shared concepts, but UK and EU rules, regulators and procedures differ. Check the framework that applies to you.

Do we need a separate process for agents?

Most teams can extend existing data-protection processes. What changes is the unit of review: the agent, its tools and its data flows.

Vikas Agarwal is the Founder of GrowExx, a Digital Product Development Company specializing in Product Engineering, Data Engineering, Business Intelligence, Web and Mobile Applications. His expertise lies in Technology Innovation, Product Management, Building & nurturing strong and self-managed high-performing Agile teams.

Build Privacy Controls Into Enterprise Agentic AI

Book an AI Assessment

Fun & Lunch