Understand agentic workflows in practice
d:u Education offers talks and masterclasses from d:u26 on agentic workflows and LLM implementation, free and at your own pace.
What are AI agents?
AI agents are autonomous software systems that independently pursue goals, break down complex tasks into sub-steps and make decisions. The decisive difference to classic AI assistants or simple chatbots: An AI agent does not wait for a single input and provides a single answer. Instead, he plans multi-stage processes, accesses various data, tools and systems, and carries out actions — within clearly defined framework conditions.
Definition
AI agents are AI systems that perceive their environment, plan independently, make decisions and carry out actions to achieve specified goals. Over time, they can learn from experience and adapt their strategies — a skill that fundamentally distinguishes agent-based AI from classic automation tools.
A simple example
A chatbot answers the question about a free meeting room. An AI agent, on the other hand, checks the availability of all rooms, compares the information with the calendars, books the appropriate room and sends out the email invitation — without people having to initiate every step. The agent takes on the entire chain of tasks and carries out the necessary actions independently.
Technological basis
The basis is made up of large language models (LLMs) and other AI models that understand language, recognize context and draw logical conclusions. Modern language models enable AI agents to process natural language, interpret complex tasks and derive appropriate actions. The difference to the simple model: AI agents can act — they use external tools, access databases and trigger actions in other systems.
If you want to dive deeper into how language models work, you can find an overview in Articles about language models and LLMs.
Why AI agents are becoming relevant right now
The idea of autonomous AI systems is not new. What has changed is technological maturity. Three developments are driving the breakthrough forward.
Three factors behind the breakthrough
- Quality of language models: Current large language models understand not only individual sentences, but complex tasks with multiple variables. LLM-based AI agents can evaluate interim results, learn from feedback, and adjust their strategy. They also process unstructured information — such as from emails or reports — and derive actions from it.
- Availability of platforms and frameworks: Salesforce, Microsoft, SAP, and Google offer integrated platforms in which companies can configure AI agents and connect them to existing applications — sometimes with low-code tools. Frameworks such as LangChain, CrewAI or AutoGen speed up development and make it accessible to smaller teams as well.
- Economic momentum: According to one DeepL study 69 percent of the managers surveyed expect AI agents to significantly change their business processes in 2026. The Salesforce Connectivity Benchmark Report 2026 confirmed: German companies are already using an average of ten AI agents, with an expected increase of 80 percent over the next two years.
Which AI agents are there?
Not every AI agent is the same. Depending on the area of application and degree of autonomy, a distinction can be made between different types.
Task-specific AI agents
These AI agents perform clearly defined individual tasks: classifying customer inquiries, extracting information from documents, summarizing content, or compiling reports from various data sources. Task-specific AI agents are the most pragmatic starting point because results can be measured quickly.
Workflow agents
Workflow agents orchestrate multi-stage processes — from data collection to analysis to triggering actions in downstream systems. Example: An AI agent in purchasing recognizes that a material inventory falls below the threshold, compares supplier offers, creates an order recommendation and forwards it to the responsible employees.
Multi-agent systems
The most complex variant: Several AI agents work together. In multi-agent environments, specialized agents hand over subtasks, validate each other's results, and use different tools and data.
Communication between different AI agents requires standardized protocols. The Model Context Protocol (MCP) enables exactly this agent-to-agent interoperability.
Industry-specific AI agents
AI agents are increasingly emerging for specific industries: analysis of patient data in healthcare, fraud detection in the financial sector, dynamic route optimization in logistics. This specialization makes the difference between generic automation and a solution that is tailored to a company's specific environment.
Examples of AI agents: Where they are already having an impact today
Structured, recurring tasks with clear performance measurement are particularly suitable for getting started. Here are examples of AI agents from various areas.
IT service and support
AI agents analyze incoming tickets, assign priorities, start solution processes and only escalate to people when complexity requires intervention. In conjunction with monitoring tools, simple troubleshooting can be fully automated — the IT team can concentrate on more complex tasks.
Customer service
AI agents sort incoming customer inquiries, collect information from various systems, and prepare answers. Compared to classic chatbots, they can also process complex queries that require data from multiple platforms to be brought together.
Finance and accounting
Check invoices, extract payment data from emails, carry out compliance checks: These are exactly the tasks that can be reliably automated with AI agents. Financial agents read email attachments, extract relevant information, transfer data to accounting systems and notify the responsible employees in case of abnormalities.
HR and recruiting
AI agents screen applications, match profiles and orchestrate onboarding tasks: Set up accounts, provide content, send invitations. This significantly relieves HR teams — especially when many new employees start at the same time.
Marketing and sales
AI agents analyze customer behavior, evaluate willingness to buy and prioritize contacts for sales. Marketing agents can create texts and content, evaluate campaign data and prepare answers to customer inquiries — as learning systems that get better with every interaction.
Practitioners report on how companies are bringing AI agents from pilot projects into productive use at d:u27 on April 13 & 14 2027 in Münster. You can find an overview of all speakers here.
What AI agents can't do yet — and where people are still in demand
When it comes to highly complex tasks such as strategic financial planning or legal judgment formation, AI agents do not achieve reliable results. This requires human intervention, experience and the ability to evaluate relationships that cannot be derived from data alone. AI agents don't replace people — they relieve them of operational tasks.
Cumulative error risk
The more complex a multi-agent system, the greater the risk of cumulative errors. Experts recommend starting with clearly defined applications and gradually expanding autonomy — always with the option of human intervention.
Integration gap as a key challenge
The Salesforce Connectivity Benchmark Report also shows: 50 percent of AI agents are still working in isolated silos. The biggest challenge is integration into existing systems and data environments. Without this integration, AI agents remain isolated solutions that do not exploit their potential because they have no access to relevant information.
What this means for SMEs
Many of the tasks in which AI agents deliver the greatest added value take up a disproportionate amount of time in SMEs: document management, answering customer inquiries, internal coordination, email administration. In all of these areas, AI agents can take on tasks, prepare information and prepare decisions.
Three typical hurdles
- Lack of know-how for developing and configuring AI agents
- Governance uncertainties and compliance — particularly when it comes to the question of which decisions AI agents can make autonomously
- Lack of orientation on where the first use case should be
Not every process is suitable for automation by AI agents — and employees in the company must also learn how to use these systems first.
Agentic workflows, tool integration, and orchestration — At d:u27, data teams and developers go into technical depth in over 40 master classes. Find out why the d:u27 is the right event for you and your team.
How companies can now get started with AI agents
Step 1: Process analysis
Where does the company lose the most time with recurring tasks? Where do employees make decisions based on information that has to be laboriously gathered together? Where do people answer the same questions over and over again?
Step 2: Check the data basis
AI agents are only as good as the data they access. The best tools and platforms are of little use if the underlying information isn't accurate. Anyone working with fragmented data sets should start here first.
Step 3: Pilot with a clear goal
For example, reduce the processing time for standard customer inquiries by 30 percent. Such specific applications — in which AI agents work with defined goals and actions — create acceptance and provide the basis for further steps.
Conclusion
AI agents are no longer an experiment
2026 marks the transition from experimentation to productive use of agent-based AI. AI agents are a strategic tool that can give companies a measurable advantage — if the applications are selected correctly and the agents are integrated into existing systems.
The key lies not in technology alone, but in the combination of clear application focus, solid data basis and the courage to start a specific project. AI agents relieve people of routine tasks so teams can focus on strategic decisions and creative solutions.
You can find out how other SMEs are getting started with AI agents on the data:unplugged festival 2027 on April 13 & 14 in Münster. At the SME Stage, companies share their experiences, and in the master classes it becomes concrete: Which examples of AI agents are worthwhile? How do you start a first pilot project?
AI agents affect all areas of the company — from IT to sales to management. For successful implementation, it is important to involve and qualify key people. data:unplugged stands for practical transfer of knowledge — from which the entire team benefits. Get your ticket now!
FAQ
What is the difference between an AI agent and a chatbot?
A chatbot waits for a single input and provides a single answer. An AI agent, by contrast, plans multi-step processes, accesses data, tools and systems and carries out actions itself. A chatbot might name a free meeting room, while an agent checks the calendars, books the room and sends the invitation.
Which type of AI agent is best for getting started?
Task-specific AI agents are the most pragmatic starting point because their results can be measured quickly. They handle clearly defined individual tasks such as classifying customer inquiries, extracting information from documents or compiling reports. Structured, recurring tasks with clear performance measurement are especially suitable.
How do multiple AI agents work together?
In multi-agent systems, specialised agents hand over subtasks, validate each other's results and use different tools and data. Communication between the agents requires standardised protocols such as the Model Context Protocol (MCP).
Why shouldn't AI agents be given full autonomy right away?
The more complex a multi-agent system, the greater the risk of cumulative errors. Experts therefore recommend starting with clearly defined applications and gradually expanding autonomy. The option of human intervention should always remain in place.
Why are many AI agents not yet living up to their potential?
According to the Salesforce Connectivity Benchmark Report, 50 percent of AI agents are still working in isolated silos. The biggest challenge is integration into existing systems and data environments. Without access to relevant information and a solid data foundation, agents remain isolated solutions.
