Metamorphic testing and AI agents in ensuring the reliability of corporate software

Abstract

The article is devoted to studying the importance of AI agents in ensuring the reliability of corporate software. The purpose of the article is to substantiate the role of metamorphic testing and AI agents in ensuring the reliability, semantic consistency and security of corporate software based on artificial intelligence. The research process used methods of analysis and synthesis to generalize scientific approaches to testing AI systems; the comparative method was used to compare traditional QA, Metamorphic Testing and the agent approach. The results of the study showed that the reliability of corporate software based on AI cannot be ensured only by traditional QA approaches, since LLM, RAG systems, chat bots, recommendation services and AI agents operate in a non-deterministic mode and can change the answer depending on the formulation of the query. It has been established that Metamorphic Testing is an appropriate method for testing such systems, since it allows us to evaluate not a single result, but the constancy of the relationship between responses after a controlled transformation of the input data: paraphrasing, changing the format, rating scale, word order, or adding a small amount of noise. It has been proven that if the content of the query does not change, then the main fact, conclusion, class, recommendation, or logic of the response should remain stable, and their significant change indicates semantic inconsistency, the risk of hallucination, or weak robustness of the model. On this basis, the author's concept of AI-Agentic Testing is proposed, in which the AI-agent automates the generation of follow-up inputs, checking metamorphic relations, detecting unstable responses, hallucinations, logical contradictions, and potential security breaches in corporate AI systems. This approach moves QA from static test scenarios to dynamic control of semantic consistency, reliability, and security of AI solutions in a corporate environment. Its practical value lies in the fact that it can be used as a basis for testing LLM, RAG systems, AI agents and enterprise AI workflows before their integration into critical business processes.