AI-Assistants: Framework for combining LLM-Driven Assistants with corporate features

Article

Abstract

This paper presents a comprehensive review of existing studies and several optimization methods, and suggests a simplified workflow for a unified framework for multimodal systems and AI-driven assistants. It focuses on the integration of AI-based tools into existing IT infrastructures by comparing approaches used in industry use cases of applying AI-based tools in their respective fields. Integrating AI into existing IT infrastructures that were not originally designed to be operated by AI, even at the level of data rendering, causes significant challenges that are worth solving. The main challenges include evaluating the rationality of using AI for certain tasks, managing the computational costs of self-deployed AI tools, managing the quality limitations of self-deployed LLMs to ensure that they can effectively interact with humans in industrial environments, and overcoming the economic limitations of using high-end LLM models such as ChatGPT, DeepSeek, or Grok in terms of token-based pricing. The study demonstrates a significant need for standardization and enhanced operability of existing approaches beyond the highly specialized solutions used in fields such as medicine, software development, and network management. The improvements are intended to enable AI assistants to make critical decisions that still require automation and optimization within industrial systems. The main aim is to systematize modern solutions and design common principles and vectors of development for highly contextual and scalable AI operating systems, enabling both AI technology providers and commercial organizations to adapt even in the short term. An example of the workflow pipeline of the proposed framework is designed to combine the adaptability and intellectual operability of high-end LLMs with the standardization of tasks for highly contextual models. It also addresses the deployment limitations of models built from scratch or derived from local models while preserving optimization for servers of small and medium-sized enterprises at each step by managing small modular tasks.