On March 17, 2026, IBM closed its $11 billion acquisition of Confluent — one of the largest bets ever placed on real-time data infrastructure. IBM's thesis: enterprise AI agents are bottlenecked not by models, but by data. Specifically, the kind of structured, event-driven, continuously updated data streams that let an agent know what's actually happening in a business, right now, before it decides what to do next.
It's a coherent thesis. Whether IBM can execute it is another question.
The framing — that most enterprise AI agents never make it to production — shows up repeatedly in trade coverage and vendor materials. According to TechHQ's reporting on IBM's own announcement materials, 79 percent of enterprises have adopted AI agents in some form, but only 11 percent are running them in production — a gap IBM's internal materials call the largest deployment backlog in enterprise technology history. The 79 percent and 11 percent figures appear in multiple vendor-adjacent sources, but a primary study behind those specific numbers is not immediately identifiable. The claim should be read as IBM's framing, not independently verified industry data.
The reasons for the gap are architectural. Most enterprise data lives in unstructured documents, event logs, and systems designed for human consumption, not agentic consumption. An LLM fine-tuned on last year's quarterly report has no mechanism to know what happened in your CRM this morning.
This is the real-time data problem. And it's exactly what Confluent was built to solve.
Confluent, a San Francisco-based data streaming company built on Apache Kafka, lets companies maintain continuous, organized flows of data across their entire technology stack. Every transaction, event, update, and signal can be captured, routed, and made available in real time. For human analysts, that's useful. For an AI agent that needs to know the current state of a system before acting, it's load-bearing infrastructure.
This is where the dependency chain becomes impossible to ignore. Agentic AI isn't a single inference call. It's a sequence of steps — the model decides, retrieves data, reasons, acts — and at each step, output quality depends on input quality. If an agent queries a data warehouse updated six hours ago, it's reasoning on stale information. If it needs to respond to a supply chain disruption but can only access structured databases, it can't see the unstructured emails, Slack threads, and PDFs where that signal actually lives.
The dependency chain for production agentic AI has several load-bearing links: model capability, instruction quality, data retrieval latency, data freshness, contextual relevance, and action quality. Break any link and the agent degrades. Real-time data infrastructure is the layer that keeps data fresh and retrievable fast enough that the chain doesn't snap. This is also why IBM's competitive position is more precarious than the infrastructure story alone suggests.
Snowflake launched Cortex Agents in general availability on November 4, 2025, giving enterprises a way to run agentic workflows directly against Snowflake's data cloud with built-in guardrails, audit logging, and enterprise authentication. Databricks shipped Agent Bricks on June 11, 2025, its own framework for deploying and managing AI agents against the Lakehouse platform. Both companies had existing data infrastructure relationships with enterprise customers and moved directly into the agent orchestration layer. IBM's equivalent — watsonx.data, which IBM claims delivers 40 percent more accurate AI than conventional RAG in internal testing — exists. But it's behind both on pure software maturity, and IBM's accuracy figure comes from internal testing only, not independent benchmarks. The company has not published peer-reviewed results or third-party evaluations. Enterprise Advantage, the consulting service IBM announced alongside the Confluent news, compounds the confusion. It's a services engagement — IBM consultants help enterprises redesign their data workflows to support agentic AI. Useful, potentially. But it's not a software product. IBM appears to be selling its professional services arm alongside its infrastructure play, which makes the announcement as much a consulting sales motion as a technology statement.
The irony worth sitting with: IBM may be making the right infrastructure bet while running behind on the software layer where that bet gets monetized. The Confluent acquisition gives IBM a real-time data streaming platform with significant enterprise penetration and Kafka expertise. Confluent's customer base and technology are genuinely strong — the company's event streaming infrastructure handles billions of messages daily for enterprises that need exactly the kind of data continuity agentic AI requires. If the question is whether IBM has the plumbing to support agentic AI at scale, the Confluent deal is a credible answer.
But the software frameworks that sit on top of that infrastructure and orchestrate agent behavior, handle multi-step reasoning, manage context windows, and enforce governance policies? That's where Snowflake and Databricks have moved first and fast. IBM's consulting-led approach suggests it knows it's behind on the software side and is trying to win on implementation. The AskIT numbers IBM cites are the most concrete evidence for the infrastructure-plus-services model working in practice: 86 percent query resolution, 74 percent reduction in calls and chats, $18 million in initial cost reduction. IBM's own internal reporting puts those figures forward — they haven't been independently audited — but if they hold at scale, they represent a meaningful proof of concept for enterprise-wide agent deployment.
For enterprises evaluating agentic AI investments, the IBM-Confluent story is a useful forcing function for the architectural question that can't be avoided: where does your data actually live, and how quickly can an agent see it?
The companies winning on the agent software layer — Snowflake, Databricks, Microsoft with Copilot, and the emerging field of agentic middleware — have all made bet-the-company moves on this question. The data infrastructure underneath is becoming a differentiator, not commodity plumbing. IBM's $11 billion says it agrees.
The question IBM hasn't answered yet is whether it can build or buy the software layer fast enough to make the Confluent bet pay off — or whether it bought the pipe but is already losing the house.
