Five Lessons From "Trusted AI Starts With the Data Stream"

Stale data used to mean a wrong report, and someone would usually catch it in time. AI agents raise the stakes because they act on the business instead of simply reporting on it. When an agent approves a payment or places an order with no person in between, outdated data leads straight to a wrong decision that nobody sees until it has already happened.

Kai Waehner made this point at the start of our recent webinar. As an Advisory Field CTO and one of the best-known voices in data streaming, he argued that data freshness is now a trust requirement rather than a performance metric. He closed the session with five lessons from customer projects around the world.

Pick the pattern for each use case

Event streams, message queues, APIs, and batch processing aren't competing options, and most enterprise architectures need all four. Streaming often sits at the center because data only has to leave the system of record once, which adds no load to the source system and opens no new attack surface. Even so, a message queue or a batch job is sometimes the simplest answer for a particular use case.

Govern events the way you govern APIs

Most companies have managed their APIs for years, but their event streams rarely get the same care. Kai recommends starting with data contracts and then building a catalog that covers events, APIs, and data lakes. A gateway, privacy controls, and end-to-end lineage complete the picture. If one of these pieces is missing, you will end up tracing problems back to their source after the fact.

Let requirements choose your deployment

A streaming platform can be fully managed, self-managed, or run in your own cloud account, and the right choice starts with your requirements. Regulation, latency, and tolerance for data loss will usually rule out some options long before price comes into it. The same thinking applies to sovereignty across the whole AI stack, where every layer needs a clear answer to three questions: where it runs, who operates it, and how you would leave. As Kai pointed out, the model is the easiest part to replace.

Feed agents governed streams, never the core

Agents should only ever receive data that meets your rules for quality and access. An orchestration layer between your systems and your agents makes that possible by handling retries, human checkpoints, and the choice of model for each task. It also gives you a single place to see everything that happens.

Record every agent action as an event

Sooner or later someone will ask why an agent made a particular decision. When every input, decision, and action is recorded as an event, the answer is already there for you to audit or replay.