How to Stop Feeding Your Agents Yesterday

An AI agent flags a transaction as fraud and freezes the account. It looks like a clean call: a sharp model, sound logic and high confidence. The only trouble is that the data it reasoned over went stale three hours ago, and nobody will know until the customer is on the phone and the money is gone.

This is the failure that should really worry data leaders, and it has nothing to do with hallucination. It is winning, too. MIT’s NANDA study found that 95% of enterprise GenAI pilots deliver no measurable business impact. Gartner pins 60% of abandoned AI projects on a single root cause: data that isn’t AI-ready.

Sean Falconer has a name for the disease. “Most companies are data delayed,” says Falconer, who runs AI product strategy at Confluent, now an IBM company. “There’s a gap between when something meaningful happens in the business and when that data is collected, processed, and available somewhere you can actually use it.”

You used to be able to live with that gap. In the BI era, a dashboard running a few hours behind still told you something true. Agents are different. “You can’t have an agent helping rebook flights if it doesn’t know which flights are available or which seats are already taken,” Falconer says. “You can’t monitor a patient and surface care recommendations if the last temperature reading is six hours old.”

What should worry a CDO most is that the failure leaves no mark. “A dashboard invites scrutiny while an AI recommendation feels authoritative,” Falconer says. “So when the data behind it is hours old, the failure is silent. Nobody catches it until the damage is done.”

Faster batch jobs won’t fix that. Context has to come from somewhere new: a derived dataset that fuses historical depth with live freshness, updating itself as the business changes and ready the moment an agent asks.

Brain, meet nervous system

That idea now carries a USD11 billion price tag. IBM closed its acquisition of Confluent in March 2026, and skeptics asked the obvious question: Would a Kafka platform remain open once Big Blue owned it? Falconer starts with what doesn’t move. “Confluent is still multi-cloud, still open source at its core, still works with every major cloud provider and available on-prem.”

What changes is the pairing. “For the first time, you have the real-time data platform and a full AI platform under one roof. IBM provides the brain, and Confluent provides the nervous system. The brain is where reasoning and decisions happen. The nervous system is what moves signals through the organization in real time and makes sure they’re trustworthy.”

For AI engineers, the plumbing matters more than the metaphor. Events stream into Kafka, Flink processes them in motion, and the output lands in a Real-Time Context Engine that materializes queryable tables. An agent reaches it through MCP, discovers the schemas, and fires a low-latency point lookup. “The agent doesn’t need to know about Kafka or Flink,” Falconer says. “It asks a question and gets an answer.”

One data flow feeds two query profiles. The context engine handles fast point lookups, the current state of a customer or an order; Tableflow turns those same Kafka topics into Iceberg tables for watsonx.data, where the heavy analytics run. No second pipeline. Freshness lands in single-digit seconds, and queries return in milliseconds, even when the events start life on an IBM Z mainframe.

Governance at stream speed

Nobody starts from a clean slate, and that’s where integration stalls. Mainframes, legacy ERP, a dozen cloud databases, a sprawl of SaaS: each describes the same customer differently. Confluent’s 120-plus connectors and change data capture move data, while Schema Registry and Flink reconcile it into a single canonical shape. The wall teams hit isn’t technical. “Someone has to decide what the authoritative schema for ‘customer’ or ‘order’ looks like, and that requires cross-team agreement.”

Security teams want to know how you govern data at the moment of inference, not just when it lands. Confluent’s answer is architectural. RBAC is enforced at the Kafka broker, “not in the application code, so it can’t be bypassed.” Flink strips PII in the stream before the data ever reaches an agent. And every event is immutable and ordered, a running record of what an agent saw and exactly when it saw it.

Traceability is where Falconer plants his flag against the phrase “trusted AI,” which he says “has become meaningless... used to describe aspirations, not mechanisms.” Trust, according to him, comes down to three things: fresh, governed, traceable. “Traceability is the one most people skip. When an agent makes a bad recommendation, the first question is always ‘why did it do that?’ If you can’t answer that, you don’t have trust.”

The APAC tax

Regionally, the math gets harder, and APAC feels it first. The region is not one market. Singapore and Australia run mature frameworks; Indonesia and Vietnam demand strict residency, and Indonesia’s penalties can reach 2% of annual revenue. “Trying to treat this as one market is where a lot of companies get burned,” Falconer says. The teams getting it right, he adds, “are designing for data sovereignty from day one rather than retrofitting it later.”

The AI layer turns every one of those rules into a sharper edge. “When an AI agent consumes data it shouldn’t have access to and makes a recommendation based on it, that’s a compliance issue that also affects a business decision,” Falconer says. “The blast radius is bigger.”

Ask what it costs to stand still, and Falconer doesn’t reach for the vendor pitch. Batch is fine for quarterly forecasting, he allows. The real danger is competitive drift. “Your competitors who are feeding real-time context into their AI aren’t just getting faster answers. They’re getting better answers.”

The blow, when it comes, is quiet. “You don’t lose a deal because your batch pipeline ran at 2 AM instead of 6 AM,” Falconer says. “You lose because your entire AI layer is operating on a staler version of reality than the other guy’s.” By the time your dashboards show it, the gap has been compounding for weeks.

Ready to learn more about data streaming? Get started with Confluent’s Latest Whitepaper →

Image credit: iStockphoto/mushakesa