Table of contents
If your company is testing AI systems that not only answer questions, but also take actions, access internal and external tools, and make decisions autonomously, then we are no longer talking about generative AI, but rather Agentic Artificial Intelligence — and its data protection risks are very different.
Agentic AI promises efficiency, automation, and cost savings. However, it also introduces a new and far more complex way of processing personal data — one that is difficult to control and, in many cases, incompatible with a “rapid” implementation without prior assessment. Precisely for this reason, the Spanish Data Protection Agency (AEPD) has recently published specific guidance on its use.
In this article, we explain what agentic AI is, how it differs from other AI systems, the specific data protection risks it creates, and what companies should do when deploying these types of solutions.
What exactly is agentic AI?
Agentic AI is a type of artificial intelligence designed to achieve a goal, not merely respond to a specific query. To do so, the system may:
- Break down complex tasks into subtasks
- Access internal databases
- Connect to external services through APIs
- Decide which action to execute at any given moment
- Store information in memory to improve performance
In other words, it does not wait for the user to provide every instruction, but instead acts with a certain degree of autonomy.
A simple example: an AI agent automatically detects a business trip, checks the corporate calendar, searches for flights and hotels, compares options, makes reservations, and sends the documentation to the employee — all without anyone having to intervene in the process.
Generative AI vs. Agentic AI
One of the most common mistakes is to think that agentic AI is simply “a more advanced generative AI.” It is not, for several reasons:
- Traditional generative AI responds to a prompt: it generates text, summarizes, translates, or suggests ideas within a limited scope.
- Agentic AI, on the other hand, chains together actions, interacts with multiple systems, and executes decisions autonomously.
Why does agentic AI multiply data protection risks?
The AEPD makes it clear: when a company introduces agentic AI, it is not simply implementing another tool — it is redesigning the way personal data is processed. The risks change (and increase), and true traceability becomes much more complex.
In many cases, the main issue is not what the company intends the agent to do, but everything the agent may do without anyone directly observing it.
Main risks arising from the use of agentic AI
1. Excessive access to internal information
Agentic AI often accesses emails, documents, calendars, or internal chats in order to “understand the context.” The risk arises when it accesses information that is unnecessary for a specific task.
Example: In managing a business trip, the agent accesses the employee’s email and uses information from unrelated messages or individuals that have nothing to do with the booking.
Security measure: Technically limit which repositories the agent can access, avoiding unrestricted or global access.
2. Disclosure of data to multiple external services
These systems typically connect to external platforms (booking systems, suppliers, cloud services), which may result in excessive sharing of personal data.
Example: The agent sends travel providers more information than necessary, such as personal preferences or internal employee context.
Security measure: Restrict the external tools and services the agent is allowed to interact with.
3. “Shadow leak”: the silent leakage of information
Agentic AI may expose user information indirectly — without any obvious data breach — through patterns, partial responses, or metadata.
Example: A third party does not access direct data, but can infer habits, preferences, or professional relationships based on the agent’s responses.
Security measure: Implement output controls and data loss prevention (DLP) tools.
4. Automation bias and overreliance
The more autonomous the agent becomes, the greater the tendency to accept its decisions without sufficient review.
Example: An employee automatically approves bookings made by the agent without verifying what data was used or whether it is up to date.
Security measure: Design meaningful human oversight, not merely formal supervision.
5. Persistent memory and profiling
Many agents retain memory to improve functionality, which may lead to unwanted profiling or improper reuse of personal data.
Example: The agent remembers frequent destinations or employee spending patterns for future trips, creating profiles that may be difficult to justify legally.
Security measure: Limit, compartmentalize, and regularly purge persistent memory.
6. Uncontrolled agent creation (“BYOAgentic”)
There is a risk that employees create their own agentic AI systems outside corporate governance controls.
Example: An employee connects an unauthorized agent to their corporate email and calendar through an external tool.
Security measure: Prohibit unauthorized agents and establish clear governance rules regarding their use.
Conclusion: efficiency, yes — but not at any cost
Agentic AI offers enormous potential for automating complex processes, but its autonomy significantly increases data protection risks if it is not implemented carefully.
Before deploying these systems, companies should ask not only what the agent does, but also what data it sees, how it makes decisions, who it communicates with, and what information it retains in memory.
In this regard, AI governance will be the most effective measure for leveraging these technologies without compromising compliance with data protection obligations. Alongside specific technical measures, it is essential for companies to adopt an internal AI usage policy establishing which tools are permitted, what access is authorized, and how such tools must be used.
If your company needs advice on data protection and digital law, our lawyers will be delighted to help you.
Data protection and digital law Area
