\u201cMost AI projects fail before they even reach a single customer because the data layer breaks down,\u201d \u00a0Falconer says. \u201cTeams have the models and mandate, but security risks and fragmented data stop them from shipping.\u201d<\/p>\n
Confluent\u2019s aim is to make real-time data streaming the secure foundation for AI apps and agents, with controls and developer tooling that high-stakes industries need.<\/p>\n
Confluent\u2019s aim is to make real-time data streaming the secure foundation for AI apps and agents, with controls and developer tooling that high-stakes industries need.<\/p>\n
So while teams can build convincing demos, they often can\u2019t go beyond POCs. According to McKinsey<\/a>, eight in ten companies cite data limitations as a roadblock to scaling agentic AI. It\u2019s the gap Confluent wants to close with its new Confluent Intelligence<\/a> and Confluent Cloud<\/a> capabilities, launched in London on May 19. The aim is to make real-time data streaming the secure foundation for AI apps and agents, with controls and developer tooling that high-stakes industries need.<\/p>\n Plenty of firms built security controls across systems and carved their data into separate swimlanes decades ago. It made sense then, but now it gets in the way of AI. Leaders are faced with an ugly trade-off: over-provision access to agents, stall the projects, or centralize the data and hope performance and control hold up. None of those options is ideal.<\/p>\n Why AI needs live context<\/p>\n For Falconer, one of the biggest bottlenecks isn\u2019t the models, it\u2019s access to current data. \u201cI see a demo like a movie set, with a fake cityscape, where I\u2019ve paid the extras to be in the background,\u201d he says. \u201cReal cities are going to be a lot more complicated and unpredictable.\u201d The hard part comes when a polished demo has to survive real users and real data.<\/p>\n Take an airline that uses a customer support agent to help passengers rebook flights. \u201cOld data could mean putting a customer on a flight that no longer exists, or into a seat that\u2019s already been taken \u2014 all of it is bad news,\u201d says Falconer, adding that the whole point is to create a better user experience and a more streamlined process. Real-time accuracy is no longer a nice-to-have for AI systems; it\u2019s the difference between a convincing demo and something that functions in the wild.<\/p>\n \u201cI\u2019ve got 10,000 hours under my belt, so I have the historical reference on how to drive\u2026 But if my camera feed of the road ahead only updates every two hours, how long is it before I crash?\u201d<\/p>\n To deliver on its promises, AI requires two kinds of intelligence: historical pattern recognition and real-time signals. Falconer likens it to driving. \u201cI\u2019ve got 10,000 hours under my belt, so I have the historical reference on how to drive,\u201d he says. \u201cBut if my camera feed of the road ahead only updates every two hours, how long is it before I crash?\u201d\u00a0<\/p>\n The same logic applies to enterprise AI. Companies have spent decades building up historical data and batch systems, but have underinvested in the ability to process what\u2019s happening now. That\u2019s the leap: AI agents need both memory and live context if they\u2019re to make safe operational decisions.<\/p>\n Easier, safer streaming<\/p>\n Confluent wants to help teams move away from hand-crafted plumbing and towards a more guided way of working with streaming systems. The latest additions, including its managed MCP server, agent skills, and dbt adapter, are designed to reduce friction in configuration and setup while democratizing access for developers who aren\u2019t as familiar with, or don\u2019t live in, that environment day-to-day. In practice, that means people can express what they need in natural language, then use familiar workflows to build, manage, and debug streaming operations.\u00a0<\/p>\n The point isn\u2019t just ease. Falconer argues that when teams have to work around cumbersome setup or opaque tooling, they often end up working around the security model too. By contrast, if the system is opinionated and easier to use, more people are likely to stay inside the approved path. As he puts it, the skills are really \u201crecipes\u201d that capture Confluent\u2019s own best practices, so developers don\u2019t have to reconstruct all that domain knowledge themselves.<\/p>\n It comes with a warning label, though. Falconer is clear that models remain non-deterministic, so the question is also how to keep them controlled. Confluent\u2019s managed MCP server is read-only and maps back to existing RBAC, roles, and credentials. At the same time, the skills layer is more flexible and better suited to staging or development environments than direct production changes.<\/p>\n An important distinction for enterprise teams, which reflects the reality of modern AI development: experimentation is useful, but production still needs determinism. \u201cWith production environments, use tools within a staging or deployment environment, and once you know it\u2019s working correctly, use a system like Terraform to push them into production in a repeatable way,\u201d says Falconer.<\/p>\n The same logic applies to data movement. Confluent\u2019s stream processing layer can link to external databases, pull in context for AI, enrich a datastream, or talk to large language models in a customer\u2019s cloud environment \u2014 but those connections shouldn\u2019t need to travel over the public internet. When Confluent Cloud is running alongside a customer\u2019s own Azure-hosted systems, communication between those surfaces stays on Microsoft\u2019s private backbone.<\/p>\n That\u2019s key because the enterprise case for AI is increasingly about securely moving context around, not just moving data quickly. Confluent is also pushing privacy deeper into the stack, with built-in PII detection and redaction in Flink SQL, so sensitive data can be handled in the stream rather than routed out to a warehouse first.<\/p>\n Real-world proof<\/p>\n Financial services already run on streams of events \u2014 card transactions, login attempts, wire transfers, and ATM activity. \u201cBut historically, it\u2019s been difficult to run advanced AI models directly within livestreams in a secure, governed way,\u201d points out Falconer. The usual fallback has been to batch everything into a warehouse and detect fraud later, which means action can come hours or days after the event.<\/p>\n The launch, which includes support for TimesFM, shifts that work closer to the point of data creation, with more adaptive anomaly detection and less model tuning. That way, teams can spot account takeover patterns or abnormal transaction spikes immediately, rather than after costs have already mounted.<\/p>\n Falconer\u2019s broader philosophy is one he learned as head of developer experience at Google: if the tooling is clunky, adoption stalls. He recalls how an API-first business messaging product was powerful, but hard to get started with. A click-through cut onboarding time and lifted the use of advanced features by 3x, while launches doubled. As he says, \u201cmake it easy and fun to use, and developers will be safer and more successful.\u201d<\/p>\n
![](\"https:\/\/www.europesays.com\/ai\/wp-content\/uploads\/2026\/05\/e36dae7b-cropped-6c610b3e-megan-carnegie-square-sizing-600x600.jpg\"\/)
<\/p>\n