AI Agents in the Company

AI agents go far beyond chatbots: they plan steps, make decisions and act in processes - from email drafts to autonomous execution. We clarify definition & components, classify Agentic Automation vs. Orchestration, show use cases and a 5-step procedure for a quick start.

Doesn't that sound fascinating? A work colleague who works independently, makes decisions and never gets tired. This is precisely the promise of AI agents: a kind of digital person who can act like an employee or colleague - only around the clock.

This topic is particularly topical because the last few years have seen a huge breakthrough in the development of Large Language Models (LLMs). We have seen how rapidly this technology is developing and what the models are capable of. At the same time, it has become clear that it is not so easy to generate real added business value from this pure performance.

To better classify the interaction of AI agents with measurable efficiency gains, it helps to take a look at the efficiency pyramid (Figure 1). It shows the levels at which the use of KI can have an impact - starting with individual employees, through entire teams and departments, to company-wide transformation.

This makes it clear: AI can work on different levels - but the real challenge is to translate this effect into concrete added value.

It is precisely at this interface - between the promise of AI agents and the need for measurable added value - that other terms emerge that are currently frequently discussed: Agentic Orchestration and Agentic Workflows. This blog article aims to clarify the most important concepts and show why they are relevant for companies

Before we delve deeper into the specific fields of application, it is helpful to clarify the key terms. This is because in the discussion about AI agent workflows and orchestration, buzzwords are often used whose meaning is not always clear-cut. With the following definitions, we create a common foundation to better understand the differences and correlations.

Large Language Models (LLMs)

Large language models - known from systems such as ChatGPT or Gemini - are essentially large language models that have been trained on huge amounts of data. Their principle is essentially that of a word prediction engine: they calculate with high precision which word is most likely to appear next. This enables them to generate texts, answer questions or hold conversations - and forms the basis for many current AI applications.

Agentic Workflow

Building on this technology, LLMs can be embedded in predefined workflows. An Agentic Workflow is a workflow or process in which an LLM is used at certain points - for example, to automatically pre-formulate an email or summarize information. The processes are still clearly structured, the model only takes on individual, clearly defined tasks.

Agentic Orchestration

With Agentic Orchestration, we go one step further. Although everything here is also based on defined processes, at certain points the agents are given the opportunity to make independent decisions about the next step. They operate within a limited scope of action in which they are allowed to act independently - for example, by sending an email or searching a mailbox on their own. This means that they not only orchestrate predefined steps, but actively shape the course of the process..

AI Agents

The last and most comprehensive term is the AI agent. This is the idea of a completely autonomous employee. An AI agent can decide autonomously what it does, when it does it, and in what order it proceeds. It only needs a task as input and then takes over planning, execution and prioritization independently - similar to a digital colleague who never tires. As shown in Figure 2, AI agents combine AI models with a memory storage and use this to interact with other tools and systems.

Focus on energy suppliers - customer service agents

A particularly vivid example of the added value of Agentic AI is customer service in the energy sector (see Figure 3). Arvato Systems has developed customer service agents that automatically process and coordinate incoming inquiries and forward them to the right people. This means that customer concerns are not only processed faster, but also more efficiently - while at the same time reducing the workload of service staff.

More use cases relating to AI agents

1. Agentic AI: SAP BANF Agent

A SAP BANF Agent supports the automated creation, checking and release of purchase requisitions (BANFs) in the SAP system. It relieves employees by independently taking over routine tasks such as determining requirements, validating data and forwarding it to the approval workflow.

2. agentic orchestration: e-commerce returns processing

In E-commerce a returns process is orchestrated. An agent decides independently whether a return requires a refund, an exchange or an inspection by customer service. The process is basically predefined, but the agent decides on the next step themselves.

3. Agentic Automation: Pre-structuring insurance claim notifications

An insurance company uses a workflow in which an LLM automatically summarizes damage reports and pre-structures them for processing. The process is fixed, the LLM only takes on a clearly defined task (text analysis & pre-formulation).

At first glance, terms such as agentic workflows or AI agents often seem complex. But getting started is easier than it seems. The most important thing is not to get caught up in individual use cases too early, but to first develop a clear vision and prioritize it.

After an initial proof of concept, scaling should also be considered quickly. This is the only way to unleash the full potential - from initial efficiency gains to the sustainable transformation of processes and organizational structures.

Our tried-and-tested approach at Arvato Systems helps companies step by step into the world of Agentic AI: