As AI systems expand in scope and complexity, gaps in data flow, inconsistent access patterns, and brittle pipeline design are becoming critical failure points.
These issues surface most clearly under production workloads, where latency, reliability, and context continuity are non-negotiable.
DBTA held a roundtable webinar, Building AI's Data Core: Architecting Scalable Data Foundations for AI, with experts who explored how technical leaders are addressing these constraints by engineering a scalable foundation built to support modern AI workloads in production environments.
Amar Srivastava, community of practice director at Informatica from Salesforce, revealed the top challenges facing companies when getting AI into production. This includes:
Data quality/retrieval concerns: 50%
Security concerns: 43%
Lack of agentic expertise: 42%
Yet 86% of companies said they are expected to increase their data management investments in the year ahead, Srivastava said.
AI-ready data consists of the following:
- Accessibility
- Discoverability
- Governance
- Observability
- Context
- Quality
- Trusted
- Observability
Srivastava recommended Informatica Headless—a pluggable architecture that works across any cloud, any tool, and any framework and is truly decoupled from the platform UI. It offers trusted data management as a native tool within any agent, app, or developer tool.
According to Vikram Kulkarni, area vice president at Perforce Delphix, software delivery has changed permanently. Execution is no longer slow, linear, or manual. AI-assisted tools augment humans—every action still reviewed and approved by a developer. AI velocity without data readiness creates delivery chaos.
Governance and security strategies that support trusted AI include:
- Maintain data shape: Let the use case drive the solution that best maintain shape and security.
- Mitigate sovereignty threat: Protection, policy, access controls, and audit built in.
- Join lineage and velocity: Automated data operations at AI pace with built in lineage and auditability.
Ebrahim Alareqi, principal machine learning engineer at Incorta, said an AI-ready foundation must reach the operational systems, preserve the details, deliver on the question, and govern once, not per copy.
Adding capacity to a six-stage pipeline gets you a faster six-stage pipeline. The costs that grow with AI workloads grow per copy: the storage, the reconciliation, and the surface you have to secure and prove. Taking a stage out removes all three at once, Alareqi explained.
Integration, masking, and the infrastructure underneath all get easier when there are fewer copies to apply them to.
High-performing AI applications are not distinguished by model weights alone, but by the speed, fidelity, and scale of their underlying data foundations, noted Steve Richardson, senior cloud solutions architect at NetApp Instaclustr.
High-performance AI applications require high-reliability open-source data infrastructure.
For the full webinar, featuring a more in-depth discussion, Q&A, and more, you can view an archived version of the webinar here.