AI Agents and Agentic AI: Fundamentals, Applications and Governance

1. What are AI agents?
2. What does ‘Agentic AI’ mean?
3. How do AI agents work?
4. What role do tools play?
5. What distinguishes AI agents from chatbots and generative AI?
6. What types of AI agents are there?
7. What are multi-agent systems?
8. What tasks and applications do AI agents have?
9. What opportunities and risks do AI agents present?
10. How can AI agents be securely integrated into existing IT systems?
11. How do you introduce AI agents into a business?
12. What role do governance, data protection and the EU AI Act play?
13. The future of AI agents
14. Conclusion
15. FAQ

How do AI agents work?

In simple terms, their mode of operation can be described as a recurring cycle:

Perceive → Understand the objective and context → Plan → Act → Check the result → Determine the next step

For example, an agent is tasked with processing an incoming request. First, it analyses the available information. It then determines which data is missing and which systems are required. It retrieves the information, carries out checks and evaluates the results. Depending on the outcome, it continues the process, requests further information or involves a human.

Technically, such a system requires several levels.

The user and application interface

People or other software systems provide objectives and tasks. This can be done via chat and voice, a user interface or directly via an API.

The agent

The agent has a role, a defined scope of responsibilities and decision-making logic. It often uses a large language model to process language, context and complex tasks.

Orchestration

An orchestration instance coordinates work steps, states and, where necessary, multiple agents. It ensures that results are passed on, tasks are distributed and escalations are triggered.

Integration

The agent gains access to existing IT systems via interfaces: databases, document management, CRM, ERP, specialist processes, service management systems or cloud services.

The technical and organisational foundation

This includes models, data platforms, cloud or on-premises infrastructure, identity and access management, monitoring, security and governance.

Such a layered architecture can also be found, for example, in Materna’s MATTER approach to agent-based applications. There, Experience, Agent, Orchestration, Integration and Foundation are regarded as a coherent architecture.

Typical technological building blocks also include Retrieval Augmented Generation (RAG), vector databases, in-memory mechanisms, workflow engines, and planning and reasoning components.

How can AI agents be securely integrated into existing IT systems?

Integration is often a more decisive factor in operational success than the choice of AI model.

An agent requires access to data and applications in order to function. At the same time, this access should be restricted as much as possible.

Several architectural principles are relevant here.

Clear interfaces

Agents should communicate with the existing IT infrastructure via controlled APIs and clearly defined services. Modular microservices are easier to audit and replace than monolithic agents with comprehensive system privileges. This pattern is used, for example, in agent-based administrative applications.

Distinct digital identities

Every active agent should be uniquely identifiable.

This makes it possible to trace which agent carried out which action. At the same time, individual permissions can be granted and, if necessary, specifically revoked.

Least Privilege

An agent is granted only the rights it requires to perform its task.

A research agent does not need write access to master data. A documentation agent does not need to be able to initiate a payment. The scope of action is determined by the task.

Sandboxing

Particularly sensitive or experimental actions can be carried out in isolated environments. This ensures that any errors are more effectively contained from a technical perspective.

This principle is already found in multi-agent architectures, where individual sub-agents run in separate containers.

Human-in-the-loop

Actions with significant consequences can be made subject to human approval.

The extent of this control depends on the level of risk. Reversible standard actions require less supervision than payments, changes to critical systems or legally significant decisions.

Logging and Observability

It should always be possible to trace:

which agent was active, what target they were assigned, what data they accessed, what tools they used, which other agents were involved, and what actions resulted from this.

Controlled operating models

Depending on the protection requirements, public cloud, sovereign cloud, hybrid architectures or on-premises operations may be appropriate.

In this context, sovereignty does not imply a blanket rejection of cloud services. The key factors are data sovereignty, freedom of technological choice, operational control and the ability to enforce governance requirements. The MATTER dossier accordingly distinguishes between technological, operational, data and governance sovereignty.

The Future of AI Agents

Development is likely to proceed in two directions: greater specialisation and less visible AI.

Today, AI is usually accessed via a dedicated interface. Users open a chat, type a prompt and wait for a response.

In future, agents will be able to operate more in the background. They will respond to events, check inputs, coordinate systems, prioritise tasks and initiate processes without a human having to formulate a prompt for every step.

This will also change the role of traditional user interfaces. Employees may find themselves navigating individual applications less frequently. They will formulate a goal, whilst an agent consolidates information from multiple systems.

At the same time, more closely networked agent ecosystems are emerging. Specialised agents from different providers or organisations could communicate with one another and coordinate tasks across system boundaries. Standards for agent-to-agent communication and tool access are therefore becoming increasingly important. Materna’s documents already describe this development as a path towards ‘Agentic Ecosystems’, in which multiple agents handle end-to-end processes and humans are primarily responsible for governance and quality control.

Another question for the future concerns digital sovereignty. As agents prepare and, increasingly, execute decisions themselves, control over data and infrastructure alone is no longer sufficient. Organisations must also be able to understand and manage the machine-based decision-making logic that shapes their processes. In current strategic deliberations, this extension is already being discussed as ‘cognitive sovereignty’.

This is likely to give rise to a paradoxical development: the more powerful and invisible AI agents become, the more visible their governance and responsibilities must become.