Why AI Agents Need Better Data With Steffen Hoellinger of Confluent

Confluent bets the agent era is won on data

Confluent's bet is that AI agents break the way enterprises have moved data. Kafka was built to distribute data without touching it, but an agent needs fresh context transformed and filtered as it travels, and batch pipelines are too slow and costly to deliver that. Streamhouse is the architecture Confluent announced for it. The company's harder argument is that the durable advantage sits in data foundations and governance, since models change constantly and agents still need humans accountable for their decisions.

Takeaways

• AI agents need data shaped in flight. Hoellinger says Kafka historically distributed data without transforming it. Agents now need data transformed, joined, enriched and filtered as it travels, so only the right data reaches the agent at the right time in the right shape. Confluent calls the architecture Streamhouse.
• Batch pipelines fail agentic use cases. Moving data in batch takes long and costs a lot, and the result is stale by the time an agent reads it. Agentic use cases need fresh, continuously aggregated data, so pipelines should process information close to the source.
• Governance is harder than building agents. Claude Code and similar tools make agents easy to build. Hoellinger says orchestration and governance are the bottleneck, with the EU AI Act raising the bar, and he cites a Gartner study from April predicting that the average Fortune Global 500 enterprise will run 150,000 agents by 2028.
• Not every AI problem needs an LLM. Traditional machine learning is faster and cheaper, and complex event processing in Flink creates value on its own. In observability, where data volumes are huge, Hoellinger says organisations should filter with CEP or ML before sending anything to a language model.
• Data foundations are the 2027 no-regret bet. Models change constantly as AI labs compete, while data is what a company controls. Hoellinger tells CIOs to build schemas everywhere, metadata, lineage and an ontology, and to stay flexible rather than betting on one platform where all agents will live.
• Real-time AI lands first in fraud and service. Fraud detection in banking, customer service, personal advice and predictive maintenance on the factory floor are the areas Hoellinger names. Organisations that invested in schemas and data lineage adopted generative AI faster than those that did not.

Where Kafka runs when the cloud goes down

• At the America's Cup, two Kafka clusters ran on separate boats, with cluster linking letting the race boat exchange information at low latency with the boat where the team made its decisions.
• Banking and insurance are the most cautious sectors about cloud versus private infrastructure, Hoellinger says.