Should This Be an SAP Agent, an Extension, or a Workflow?
A practical decision guide before enterprise AI creates responsibility drift.
Many SAP AI discussions start with the wrong question:
“Should we build an agent?”
The better question is:
“What responsibility are we assigning?”
Before choosing a product surface, force the artifact to answer what it owns: business truth, state, decisions, execution, stable capability, repeatable process, lifecycle versioning, and accountability.
Use an extension when...
- the enterprise needs a stable reusable business capability
- custom UI, backend, authorization, APIs, data model, or tenant lifecycle is required
- extension-owned state must be versioned and migrated
- the capability must be secured, tested, monitored, and evolved
- business logic should not be hidden inside an agent prompt
Use a workflow when...
- the path is known or repeatable
- approval, audit, SLA, escalation, or exception handling is required
- the process must be replayed or explained
- cross-system orchestration needs governance
- agents may help but should not swallow process accountability
Use an agent when...
- input is ambiguous or unstructured
- the system must interpret intent
- the path is not fully predetermined
- dynamic tool selection is required
- contextual judgment, summarization, recommendation, or coordination is needed
- the agent acts under policy through governed tools
- the agent does not bypass extension APIs, workflow accountability, or core-system truth ownership
Use lightweight automation when...
- the task is low-risk
- the process is internal or experimental
- speed matters more than long-term governance
- no core business truth or regulated execution is involved
Keep it in the core system when...
- canonical business truth is involved
- financial, legal, HR, procurement, or compliance truth is at stake
- S/4HANA or another system of record owns the business object
- the enterprise cannot tolerate truth drift
Decision summary
Stable capability: extension.
Repeatable process: workflow.
Contextual decision and tool selection: agent.
Human interaction surface: assistant.
Canonical business truth: core system.
If the action writes to a core business system, it should go through governed workflow or extension APIs, not direct free-form agent execution.
The goal is not to build fewer agents. The goal is to avoid assigning the wrong responsibility to the wrong execution unit.