Ko uporabnik ni več samo človek: od priporočil do avtonomnih agentov v poslovnih aplikacijah

Artificial intelligence (AI) is becoming an indispensable part of business applications, but its integration into business-critical systems requires a fundamentally different approach than simply passing data to large language models. Based on practical experience with the development of a resource management platform in project-oriented service companies, this paper presents a systematic approach to the integration of AI. First, we define the spectrum of AI integration into business processes — from user-confirmed recommendations, through semi-autonomous agents to autonomous agents — and for each level of autonomy we define appropriate governance mechanisms and show where these overlap with the requirements of the EU AI Act. We further address the thesis that information systems will also have to be built for AI agents that will access them programmatically, for example via the MCP protocol, and that safeguards must be built into the tools themselves, not just the agents. This is followed by a practical comparison of seven agent building frameworks and a justification for our choice, followed by the architecture of the agent layer as we built it in the Alloxy platform: two-tiered tool authorization, structured long-term memory, a pattern of human confirmation before actions with consequences, deterministic model routing, observability in our own infrastructure, and verifiable cascading deletion. We conclude with three concrete agents and transferable lessons for developers and architects of business solutions.