Requirements for AI Systems: Why They Differ from Classical Software Systems
At Spirit in Projects, we have spent years helping companies define requirements for complex IT systems and professionally prepare tenders. In recent years, a clear trend has emerged: Artificial intelligence (AI) is being integrated into more and more projects—and brings with it unique challenges.
When it comes to AI systems, defining the right requirements from the outset is crucial. Many companies still rely on the same methods used for classical software projects—and later struggle with poor quality, lack of explainability, or unrealistic expectations.
At Spirit in Projects, we specifically support our customers in formulating AI system requirements in a way that ensures they are robust for implementation, operation, and tendering.
Learning Instead of Programming
Classical software systems follow clearly defined rules: If input X, then output Y. Their behavior is deterministic and predictable.
AI systems work fundamentally differently: They learn from data and develop a model. The “rules” do not come from programming but from machine learning. This means: Data is the key to success. Even in the requirements definition phase, critical questions must be answered, such as:
- What data is available?
- What is its quality and volume?
- How is the data currently maintained and made accessible?
Our experience shows: This aspect is often underestimated in practice—with dire consequences for the project.
Probabilities instead of Guarantees
Another fundamental difference: AI systems deliver probabilistic results. For example, an image recognition system might report, “This object is a cat with 95% probability.” However, it does not guarantee flawless detection of every cat.
Requirements for result quality (accuracy, precision, recall, confidence intervals) must therefore be explicitly defined. In our AI tendering projects, we ensure these quality parameters are clearly specified—to avoid later disappointment.
Dynamic Behavior
Classical software remains largely stable after deployment. AI models, however, often require regular retraining as data and environments change (concept drift).
This leads to specific requirements for:
- Lifecycle management
- Model monitoring and maintenance
- Governance and responsibilities
Spirit in Projects helps customers integrate these aspects into their requirements definitions and tender documents—a critical point often neglected in many projects.
Explainability and Traceability
Especially in regulated sectors (e.g., financial services, public sector, healthcare), explainability is crucial. AI systems must not be black boxes.
Requirements must clearly define:
- To what extent must explainability be ensured?
- For whom must traceability be provided (e.g., internal audits, external regulators)?
In our practice, we ensure these aspects are structured in the requirements catalog—balancing technical feasibility with regulatory demands.
Conclusion
AI system requirements are not just an extension of classical software requirements. They must explicitly account for learning capabilities, probabilistic nature, and data dependency.
Spirit in Projects brings extensive experience from numerous AI projects and tenders. We help companies formulate sustainable and reliable requirements—laying the foundation for successful AI initiatives.
