How the SAP Change Impact Agent Works
Model judgment inside deterministic guardrails.
An SAP Explore workshop produces narrative decisions. The agent turns them into structured impacts without pretending that schema validation is factual validation.
As-Is + To-Be
Understand
Retrieve
Compare
Validate
Business review
OCM plan
Lifecycle
Eight teaching points
The architecture in plain language
MCP is not the model. A schema is not truth. The agent proposes; the business owns.
Structured output
The schema guarantees shape and types. It does not guarantee that a claim is factually correct.
The model API
Qwen runs server-side through an OpenAI-compatible provider. Secrets never enter the browser.
Tools
Tools calculate or retrieve bounded information. They are capabilities—not hidden reasoning.
MCP
Resources are application-selected context; tools are callable functions exposed through a standard protocol.
Bounded loop
The model may retrieve, draft, validate, and repair, but the run stops after at most six model steps.
Evidence and controls
Every material impact cites evidence. Code recalculates severity and rejects broken references.
Human review
The business owner accepts, edits, dismisses, adds, and ultimately marks the report reviewed.
OCM handoff
Accepted and edited impacts become traceable actions with owners, dates, status, and an explicit lifecycle through hypercare.
Interpret and propose
The model compares narrative states, selects relevant synthetic context, drafts impacts, identifies role changes, and turns missing decisions into questions.
Evidence is inspectable
expands into the supplied workshop statement—not a hidden model thought.
Workshop decision 3
Discounts greater than 10% require sales-manager approval.
Replay the safe agent trace
From lab to enterprise
What would change in production
Enterprise identity, approved process repositories, customer-specific governance, observability, data-retention controls, and formal validation would be added before real workshop content is used.