SANITIZED CASE · SYSTEM ANALYST
Generalized business context only. No internal names, private APIs, personal data, or confidential employer architecture.
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.
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.
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.
RAG supplies approved documents. MCP exposes a narrow tool catalog. The model never receives raw credentials; tools execute behind an orchestration layer.
Existing enterprise services stay behind REST contracts. The agent sees tool schemas, not undocumented internal endpoints.
The agent is not a general intranet search and not an unsupervised operator. Write paths require policy checks. Knowledge is scoped by role.
Hallucination, over-privileged tools, stale indexes, and prompt injection via retrieved documents. Constraints include data residency, audit, and change-control for tool catalogs.
Scenario scripts, golden questions with expected citations, negative tests for empty retrieval, and UAT with business owners — not model-training experiments.
Stakeholder interviews, use cases, BPMN/UML views, API and data contracts, acceptance criteria, and support through development and acceptance. Role: System Analyst.
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.