AI initiatives rarely fail because of model quality. They fail because the underlying data systems were never designed for reliability, context retrieval, or operational consistency.
As organizations move AI from experimentation into real workflows, the data engineering challenges fundamentally change. Pipelines must support continuously changing data, real-time retrieval, lineage tracking, and repeatable outputs. Without these foundations, AI systems drift, hallucinate, and produce results users cannot trust.
DBTA recently held a webinar, Data Engineering for Reliable AI Systems, featuring Kevin Hu, data observability at Datadog and Jerod Johnson, director, technology evangelism at CData Software, who explored the architectural patterns and operational practices required to make AI dependable at scale.
Data consumption has shifted in multiple ways since AI has seen an increase in usage, Hu explained. This is as follows:
- Data for Analytics: Humans consume data through BI tooling.
- Data-as-Product: Teams consume shared, governed assets.
- AI Consumption: Models, agents, and apps consume through context
Data teams are three times more exposed to AI through the data to build AI systems, AI consuming data by default, and AI for data work, Hu said. Data still powers customer experiences and internal decision-making. The need for trust and context remains.
Software and data are complementary goods, according to Hu. As the cost of software decreases, the value of (and therefore demand for) data increases. Systems become interchangeable and the lines between data and software engineering will continue to blur. The value of understanding humans increases.
According to Johnson, the organizations that will win will have one governed data infrastructure that supports the entire autonomy spectrum for AI.
2025 was about experimentation, Johnson said, 2026 is about results. We are at an inflection point that will separate the successful AI companies from the unsuccessful.
Johnson noted that the architectural blueprint for agent-ready infrastructure includes:
- Connectivity: AI can reach every system it needs to act
- Context: AI can understand the data it receives
- Control: AI can be governed reliably
For the full webinar, featuring a more in-depth discussion, Q&A, and more, you can view an archived version of the webinar here.