Enterprise AI agents are only as useful as the business context they can see. For ERP teams, that puts real-time, governed data movement at the center of production AI readiness, especially when agents need to act on live finance, order, inventory, customer, and supply chain events.
Most enterprise AI projects do not die in front of a customer. They die quietly, upstream, in the plumbing, where the data meant to feed the model turns out to be stale, exposed, or locked behind a wall the security team will not open.
That is the problem Confluent went after when it announced new capabilities across Confluent Intelligence and Confluent Cloud from London. The pitch is blunt. “Most AI projects fail before they reach a single customer because the data layer breaks down,” said Sean Falconer, head of AI at Confluent. Teams have the models and the mandate. Security risk and fragmented data stop them from shipping.
For SAP professionals, this is not a generic streaming story. Confluent, now an IBM company after a roughly $11 billion acquisition completed in March 2026, runs a bidirectional integration with SAP Datasphere that connects ERP data from SAP S/4HANA and ECC into the wider landscape and back again, and is directly integrated with SAP Business Data Cloud (BDC). So when Confluent hardens its AI data layer, it hardens the layer that increasingly sits between an SAP system of record and every agent an organization wants to build on top of it.
Governance inthe Data Stream
The most recent release targets three specific places where AI deployments break down:
- Natural-language operations: Developers can use a fully managed Model Context Protocol (MCP) server as a control plane, letting AI build, manage, and debug streaming operations, with Agent Skills encoding organizational best practices so those operations run consistently.
- Automated data privacy: A built-in machine-learning function detects and redacts personally identifiable information directly in Apache Flink SQL, without custom code or moving data to a warehouse first.
- Secure connectivity: Support for Azure Private Link keeps AI workloads off the public internet, giving Flink jobs private paths to services such as Azure OpenAI.
Underneath sits the Real-Time Context Engine, now generally available, which continuously serves fresh, governed context to any AI agent at low latency. Confluent also moved its Streaming Agents to general availability: event-driven agents that run natively on Flink and Kafka with a four-nines SLA. The framing is sharp. As models become interchangeable, the competitive advantage is whether an organization’s agents can see and act on the live state of the business.
Batch Data Creates ERP AI Bottleneck
SAPinsider’s own 2026 research is unusually direct about the gap. In its benchmark on SAP Business Data Cloud, only 3% of organizations report having a unified, governed data layer, while 38% remain in siloed environments. That is the modernization gap sitting between SAP data and any production AI ambition. A separate January 2026 SAPinsider analysis puts it plainly: real-time SAP data is the missing link in enterprise AI and analytics, and batch-based extraction, overnight loads, and static reporting cycles introduce delays that make AI models underperform and erode user trust.
Read together, those two findings and the Confluent announcement land squarely on the wound. SAP systems capture a continuous stream of financial postings, order updates, and inventory movements, which in theory makes SAP an ideal foundation for real-time AI. In practice, most SAP shops still feed AI from yesterday’s snapshot. Confluent’s argument is that the fix lies in an event-driven data layer that streams SAP and non-SAP data together, redacts sensitive fields in motion, and hands agents governed context the moment a business event happens.
The governance angle is the part regulated SAP shops should not skim past. PII redaction within the stream, before data reaches an external model, and private connectivity that keeps workloads off the public internet are exactly the controls a security team cites when blocking an AI pipeline. Confluent is trying to remove the reason a CISO says no.
What This Means for ERP Insiders
ERP teams need to separate real-time requirements from hype. Not every data element needs streaming access, but agents that support finance, inventory, order management, supply chain, or customer operations cannot rely on stale snapshots when business conditions change by the minute. CIOs and enterprise architects should identify which ERP events require real-time or near-real-time context before building pipelines that are expensive, overextended, or poorly governed.
Data governance has to move into the pipeline. Confluent’s in-stream PII redaction, private connectivity, and Real-Time Context Engine point to a broader shift from governing data after it lands to governing data while it moves. ERP, security, and data leaders should require lineage, redaction, access control, and policy enforcement inside the AI data flow before agents touch regulated or operationally sensitive workflows.
Production agents raise the stakes for ERP integration architecture. When AI agents act across ERP, CRM, supply chain, data platforms, and external models, the weakest pipeline can become the weakest control point. ERP leaders should pressure-test event-driven architectures against real SAP and non-SAP workloads, audit requirements, latency needs, and vendor roadmap dependencies before treating streaming data as the foundation for autonomous operations.
Editor’s note: A version of this article was originally published by SAPinsider on 7/9.
