What you will accomplish
A bounded integration proposal identifies allowed actions, data boundary, owner and validation path.
This guide is for teams evaluating an AI agent, Skill or MCP integration around SAP knowledge or systems. It emphasizes the smallest useful permission and an observable first use case.
Before you start
- A non-production or read-only environment
- A defined user task
- Security and platform stakeholders
Tutorial content
Start with an observable, read-first job
The strongest first use case is not a broad system connection. Pick one valuable question that can be answered through documentation, metadata or other low-risk inspection.
Define what good output looks like and who evaluates it before connecting the agent.
- No production mutation is required
- The data classification is understood
- A named owner can review logs and results
Design the boundary before the tool
Describe the agent host, identity, allowed tools, destination, data boundary and explicit exclusions. A named read-only profile is usually easier to test and revoke than a general-purpose credential.
Only expand capability after the first result, access logs and owner review support a separate decision.
Reading path
Start with documentation, metadata or a low-risk inspection task—not broad write access.
Checkpoint: The task has a clear business benefit without production mutation.Specify host, identity, allowed tools, data classifications and explicit exclusions.
Checkpoint: The team can say what the agent cannot do.Use named profiles, read-only access where possible and a disposable test context.
Checkpoint: Credentials and target environment are deliberately bounded.Review logs and task results with the responsible owner before expanding capability.
Checkpoint: Any next permission increase is a separate decision.