SANITIZED CASE · SYSTEM ANALYST

Enterprise AI Agent — Requirements and Solution Design

Generalized business context only. No internal names, private APIs, personal data, or confidential employer architecture.

GigaChatLLMRAGMCPREST APIBPMNUML

1. Generalized Business Context

A large organization needed an assistant that could answer operational questions and start controlled actions without exposing every backend system to the model. The analyst work was to turn an unclear request into scenarios, boundaries and a buildable design.

2. User Scenarios

Typical sanitized journeys: an employee asks a policy or status question; the agent retrieves approved knowledge; if an action is required, the agent proposes a tool call that a human or policy layer can accept.

3. Requirements

Functional: grounded answers, tool use, audit trail, role-aware responses. Non-functional: latency budget, fail-closed when retrieval is empty, no leakage of secrets into prompts, human confirmation for write actions.

4. RAG and MCP Components

RAG supplies approved documents. MCP exposes a narrow tool catalog. The model never receives raw credentials; tools execute behind an orchestration layer.

5. REST Integrations

Existing enterprise services stay behind REST contracts. The agent sees tool schemas, not undocumented internal endpoints.

User to AI Agent to Orchestration to RAG and enterprise services User AI Agent Orchestration RAG MCP / tools REST services

6. System Boundaries

The agent is not a general intranet search and not an unsupervised operator. Write paths require policy checks. Knowledge is scoped by role.

7. Risks and Constraints

Hallucination, over-privileged tools, stale indexes, and prompt injection via retrieved documents. Constraints include data residency, audit, and change-control for tool catalogs.

8. Testing Approach

Scenario scripts, golden questions with expected citations, negative tests for empty retrieval, and UAT with business owners — not model-training experiments.

9. Analyst’s Contribution

Stakeholder interviews, use cases, BPMN/UML views, API and data contracts, acceptance criteria, and support through development and acceptance. Role: System Analyst.

10. Lessons Applicable to Future AI Projects

Most of the work is boundaries: what the model may see, what it may call, and how a human verifies the result. That is solutions engineering, not model research.

See related diagrams