Integrating AI Agents with SAP, Salesforce and Oracle

SAP, Salesforce, and Oracle are where money and contracts live. An enterprise AI agent that “just uses the UI like a user” is RPA with a language model attached — you inherit every screen change and you add a new class of error: fluent, wrong postings.

The factory pattern is a tool layer: versioned APIs and events, idempotency keys, and AgentsAI as the only orchestrator that may choose those tools. The agent never receives a standing superuser role.

SAP

Prefer OData services, BAPIs, or an approved middleware. Idempotency on create. The agent may read business objects in staging freely; production writes require a human and a compensating transaction story. Do not scrape SAP GUI.

Salesforce

Composite API and platform events beat click-paths. Bulk and governor limits are part of the design. An agent that “updates every opportunity” without a bulk pattern will fail in ways that look like a CRM outage.

Oracle

Whether E-Business Suite, Fusion, or a custom Oracle stack, the same rule: published APIs or a governed integration bus. Queue the write. Make the agent’s unit of work a message you can replay, not a session on a form.

type RecordSystem = "sap" | "salesforce" | "oracle";

type ToolCall = {
  system: RecordSystem;
  op: "read" | "write";
  idempotencyKey: string;
  payload: unknown;
};

async function dispatch(call: ToolCall, hitlApproved: boolean) {
  if (call.op === "write" && !hitlApproved) {
    throw new Error("write_requires_hitl");
  }
  // Adapter maps to OData / Composite / REST. No GUI.
  return { ok: true, replayId: call.idempotencyKey };
}