What This Is
On August 28, Alibaba Cloud will host an in-person salon at Shanghai's Xuhui Binjiang district, themed "How to let AI Agents make decisions with real-time data." On the surface it's a technical share-out — but the judgment the organizers are putting forward deserves our attention: model selection stopped being the bottleneck for AI adoption long ago. What's holding Agents back in real production environments is the pipeline that feeds business events into the model.
Specifically, here's the current state of most enterprise AI projects: business data gets aggregated into the warehouse, batch-processed overnight (T+1) into reports, and those static results are then stuffed into the model as context. Agents always end up with facts from the previous batch — decision inputs are stale, retrieved knowledge fragments are expired, and time-sensitive tasks (risk control, recommendations, real-time pricing and the like) routinely miss the mark.
Industry View
The bull case: The compute arms race and large-model iteration cycle are already brutal enough — differentiation at the model layer is narrowing. The next round of AI infrastructure competition will be waged on the line of "whether data can enter the model in real time." Streaming ingestion (Kafka-style), stream processing (Flink-style), and real-time data lakes — components that originally belonged to the big-data stack — are being repackaged as AI-ready capabilities and reappearing on procurement shortlists.
The bear case and risks: Another cohort of practitioners warns that treating the real-time data pipeline as a silver bullet is equally dangerous. Until data governance (definitions, quality, lineage) is done properly, real-time will actually amplify errors faster — batch processing at least has a window to catch problems, but a real-time pipeline failing means a seconds-level incident. Others question whether "real-time" is even a real requirement for most enterprises: T+1 decisions don't affect business outcomes, so there's no need to rebuild a dedicated data foundation just for AI. With investments easily running into the tens of millions of RMB, whether mid-sized and small businesses should follow requires sober accounting.
Impact on Regular People
For enterprise IT: Over the next year, when evaluating AI projects, "whether data can be supplied to the model in real time" will shift from a technical detail to a hard line item on procurement lists. If the internal data warehouse is still on a T+1 architecture, expect most AI projects to stall at the pilot stage.
For individual careers: Colleagues doing data analysis, business analysis, and operations may need to pick up "streaming data" skills going forward. Being able to write SQL for historical reports isn't enough — they need to understand how real-time pipelines are built, how latency is controlled, and how to recover when things go wrong.
For consumer markets: C-end products like AI customer service and AI recommendations will see visible experience improvements — they can finally make decisions based on "what just happened." But the bill enterprises pay for the real-time data foundation will ultimately be partially passed on to consumers through price hikes or subscription fees.