AWS Introduces Agentic AI Framework for ERP with Deny‑By‑Default Rules and Separate Identity
What Happened — AWS launched the Agentic AI Solutions Framework for SAP‑based ERP automation. The framework runs on Amazon Bedrock AgentCore, uses the open‑source Strands SDK, and enforces deny‑by‑default rules plus a distinct service‑account identity for autonomous actions.
Why It Matters for Compliance & Audit Readiness
- The built‑in identity switching (OAuth 2.0 two‑legged for the agent, three‑legged for humans) creates immutable, auditable trails that map directly to SOX‑type control requirements.
- Layered deterministic controls (structured prompts, SOP retrieval, runtime guardrails, cross‑agent verification) give you continuous evidence that the AI behaved within defined policy bounds – exactly the kind of control‑mapping evidence SOC 2 auditors look for.
- By separating advisory, supervised, and autonomous execution modes, organizations can demonstrate a risk‑based rollout and retain “trust‑as‑you‑go” documentation for each automation phase.
Who Is Affected
- Large enterprises that run SAP or similar ERP systems (manufacturing, retail, professional services, etc.).
Recommended Actions
- Map the framework’s deny‑by‑default and identity‑switch controls to your SOC 2 CC6.1 (Logical Access) and CC7.1 (System Operations) requirements.
- Capture the agent’s audit logs as continuous evidence in your compliance repository.
- Pilot the advisory mode, validate the SOP retrieval accuracy, then expand to supervised execution while documenting confidence thresholds.
Source: Help Net Security
Technical Notes
- Runs on Amazon Bedrock AgentCore with the Strands open‑source SDK.
- Uses OAuth 2.0 two‑legged authentication for the agent’s service account; switches to three‑legged OAuth when a human intervenes.
- Determinism enforced via controlled model parameters, structured prompts, approved SOP retrieval, runtime guardrails, and multi‑agent cross‑verification.
Source: same article