There’s a well-documented pattern in enterprise AI. The pilots impress, the board sees the demo, the project gets approved and the team starts building. But then – things to fall down when the program goes into full production, where data is messy, timing matters, and the stakes are real.
A retail AI recommends a product the customer already bought. A fraud model flags a transaction after the money has moved. A support agent tells someone their order is on the way when it was cancelled an hour ago.
And yet in most of these cases, the model isn’t the problem. The context is.
What agents need
An AI agent needs two things to reason well: real-time data and historical data. It needs to know what’s happening right now. And it needs to understand what that means, which requires knowing what happened before.
Most enterprise architectures treat these as separate problems, stored in separate systems, accessible through separate pipelines. Either way, the result is flawed: you get agents that are fast but shallow, or agents that are well-grounded but working from yesterday’s picture.
Streambased gets round this and maximises your AI capability by treating these two datasetsas one. As a zero-copy data virtualisation layer built on Apache Kafka and Apache Iceberg it makes live streaming data and historical data visible together at query time – without moving or duplicating anything.
AI agents retrieve structured, current context through MCP, drawing simultaneously on what's happening and what's happened before, in milliseconds.
No ETL pipelines, no accumulation lag, and no separate systems to stitch together.
What unlocks when real-time meets historical data: 5 use cases
When you put the two datasets together, a whole raft of powerful use cases are optimised. For example:
● Customer 360 and Order 360 (FS, retail, media): To deliver smart customer support in any vertical, agents need more than a profile. An FS customer raising a dispute looks very different from a new retail account making the same request – provided you have the full picture. Streambased joins live order events, payment updates and support tickets with the historical record of that customer’s relationship, giving AI the complete context to respond accurately rather than generically.
● Real-time personalisation (retail, media, FS): Real-time signals tell you what someone is doing; history tells you who they are. A media platform pushing content to someone who has just started a true crime binge needs to know both their current intent and their long-running preferences. Streambased serves both simultaneously, so recommendations adapt to the moment without losing the broader picture.
● Anomaly detection (financial services, operations, logistics): A single anomalous event rarely tells you much on its own. Whether it's a suspicious transaction, an operational spike or a network irregularity,understanding what's wrong requires knowing what's normal — and that means access to historical baselines alongside the live signal. Streambased makes both available to investigative AI in real time.
● Patient scheduling and care coordination (healthtech): Clinical scheduling depends on accuracy: clinician availability, patient readiness, test results. But those live signals only make sense against a background of care history, eligibility rules and prior treatment. Streambased serves that combined context to scheduling agents in milliseconds, reducing errors and delays.
● Property and facility management (real estate, facilities, logistics): Live maintenance tickets, rent payment events and building alerts mean little without the lease terms, SLA agreements and repair histories that give them context. AI agents that can only see one dimension make mistakes. Streambased gives them both, enabling proactive rather than reactive management.
The posts in the rest of this series will look in turn at each of these use cases in more detail. But the underlying capability is consistent across them all: AI that knows what’s happening and what it means, because it can see present signals informed by past insights.
.avif)
