The thesis focuses on the design, implementation, and evaluation of an enterprise AI agent for secure access to internal company knowledge and services. The proposed system combines retrieval-augmented generation, authorization-aware access control, and controlled tool invocation to provide grounded answers from internal data while respecting the authenticated user's permissions. The design and implementation are developed with regard to the practical requirements of the company environment, including its technological context, security constraints, and expected future direction. The work also considers practical enterprise requirements such as security, latency, operational cost, and architectural reusability.

• Analyze and design the architecture of an enterprise AI agent for secure natural-language access to internal data.
• Design a security model for AI agents based on RBAC/ABAC principles, including authorization flows, auditability, and data-protection mechanisms.
• Implement a limited functional prototype (MVP) demonstrating selected core concepts, such as a chat interface, retrieval-augmented generation over one internal data source, basic authorization enforcement, and audit logging, with regard to the requirements and technological context of the company.
• Experimentally evaluate selected retrieval, security, and optimization strategies with respect to answer quality, latency, operational cost, and other relevant enterprise criteria.

The software submitted with the thesis will consist of a generalized reference implementation of the proposed enterprise AI agent and will be released under the MIT License. The published part will include the reusable core of the solution, such as domain and API models, REST and WebSocket interfaces, security and authorization logic, audit logging, rate limiting, tool-calling orchestration, generic connector abstractions, and selected non-confidential integration components. Company-specific parts of the implementation, in particular connectors to internal services, generated clients and schemas based on proprietary internal APIs, internal configuration values, deployment settings, credentials, environment-specific identifiers, and any internal datasets or logs, will not be published. These excluded components will be omitted or replaced by sanitized, mocked, or example implementations so that the public artifact remains understandable and reproducible without disclosing confidential company information.