Artificial Intelligence
Do you trust your AI agent?

In many production-related scenarios, AI agents now play a crucial role. At the same time, however, it is becoming apparent that expectations for agent-based AI are not always being met at present. Studies indicate that many current approaches do not yet deliver clear economic value, and a significant proportion of projects are terminated prematurely[1]. In addition to data quality and lineage issues, sovereignty and regulatory compliance as well as security and privacy risks are considered major challenges[2]

Given this context, key questions arise for companies: Can AI agents be used responsibly in safety- and business-critical contexts? To what extent can and should one rely on AI agents? And is the associated risk acceptable?

Strategy Instead of Black-and-White Thinking:
The Right Approach to AI Agents

In practice, it is clear that neither the reckless use nor a blanket ban on AI agents is effective. Rather, a structured and risk-aware approach is crucial. Companies face the task of identifying suitable use cases, realistically assessing the expected benefits, and systematically analyzing risks.

Such an approach should include a thorough analysis and weighing of risks and potential, an assessment of economic benefits (return on investment, ROI), and the targeted integration of measures for dependability and traceability.

The goal is to deploy AI agents where they deliver demonstrable added value – while simultaneously making the associated uncertainties manageable.

Structured Development of Reliable AI Agents

The Fraunhofer Institute for Cognitive Systems IKS supports companies in developing AI agents in a sustainable and reliable manner and transferring them into practice. A central component of this is the SCALE framework developed by Fraunhofer IKS. It describes a structured approach and key steps on the path to reliable agent-based AI systems:

• Strategy: Definition of goals, risks, and opportunities
• Concepts: Development of solution approaches and system architectures
• Assurance: Identification, analysis, and mitigation of critical risks
• Learning: Iterative testing and continuous learning using prototypes
• Execution: Transformation into reliable applications

This holistic approach ensures that the central requirement of reliability is taken into account not only at the end, but throughout the entire development process.

Experience the reliability of AI agents

At the all about automation trade fair in Straubing on June 10-11, 2026, researchers from Fraunhofer IKS will demonstrate how this approach can be implemented in practice using various AI agents in automation:

• Troubleshooting Assistant
  An LLM-based assistant helps to analyze production malfunctions. Through transparent and verified outputs, production downtime and manual effort can be significantly reduced.
• Safety Engineering Support
  AI-based tools support for safety engineering tasks, such as the creation of hazard and risk assessments (HARA). Depending on the complexity, this can reduce the time required from weeks to hours.
• Reliable PLC Code Generation
  Generative AI for creating PLC code, combined with iterative verification and validation. This enables a significant acceleration of development while maintaining the necessary quality and traceability.

In a demonstrator for hazard analysis in production environments, interested visitors to the trade show can also test for themselves how well they perform compared to a standard AI agent and a specifically safeguarded AI agent (see Figure 1). This clearly illustrates how various measures contribute to the reliability of AI agents.

Stop by, test your trust in AI agents, and talk to us about how you can safely leverage the potential of AI for your use cases.

Referenzen

[1] Gartner Predicts Over 40% of Agentic AI Projects Will Be Canceled by End of 2027

[2] https://www.fivetran.com/resources/reports/the-2026-agentic-ai-readiness-index