Confluent, now an IBM subsidiary after an $11 billion acquisition closed in March 2026, is betting that the AI data layer, not the model, is where enterprise deployments actually break. The company’s latest release adds natural-language streaming operations via a managed MCP server, in-stream PII redaction (personally identifiable information, scrubbed before data ever reaches an external model) inside Apache Flink SQL, and Azure Private Link support for keeping AI workloads off the public internet. For SAP shops specifically, Confluent’s bidirectional integration with SAP Datasphere and direct connection to SAP Business Data Cloud<\/a> puts this infrastructure squarely between S\/4HANA and every agent an organization wants to build on top of it.<\/p>\n Only 3% of organizations report a unified, governed data layer, according to SAPinsider’s 2026 benchmark, while 38% are still running siloed environments. If your SAP data is feeding AI agents through overnight batch loads and static snapshots, you’re not running production AI, you’re running a demo with a delay. The organizations that close this gap first aren’t just faster; they’re building agents that can actually act on live inventory, order, and financial events rather than yesterday’s approximation of them.<\/p>\n The governance angle here is the one that tends to get skimmed and then regretted. The recurring failure mode in enterprise AI pipelines isn’t the model selection or the use case design. It’s the CISO saying no at the infrastructure layer, usually because PII is crossing a compliance boundary or traffic is routing over the public internet. Confluent’s in-stream redaction and private connectivity are architectural arguments aimed directly at that veto. The question for any CDO is whether this capability genuinely satisfies your security team’s objections or just reframes the same exposure in vendor language.<\/p>\n Confluent’s framing, that models are becoming interchangeable and the real competitive advantage is whether your agents can see the live state of the business, is exactly the kind of claim an incumbent data platform vendor would make after being acquired by IBM and needing a growth narrative. But the underlying data holds: batch-based ERP extraction does introduce the kind of latency that erodes agent reliability and user trust. The falsification condition is simple enough. If your agents’ error rates and escalation frequencies don’t improve when you move from batch to streaming context, the data layer wasn’t actually the bottleneck and this investment needs a harder look.<\/p>\n MCP is a standardized interface that lets AI agents request the specific context, tools, and data they need from external systems, rather than having everything baked into a single prompt or hard-coded integration. Think of it as a universal adapter between an agent and the data infrastructure it needs to act. In Confluent’s implementation, a managed MCP server becomes the control plane for streaming operations, letting agents build and manage data pipelines through natural language rather than custom code.<\/p>\nWhat this means for your business<\/h2>\n
Concept deep-dive: Model Context Protocol (MCP)<\/h2>\n