Enterprises are racing to build physical AI and predictive models, but the algorithm is only half the battle. The real bottleneck? Trusting the data underneath it.
AI captures the headlines, but the underlying requirement hasn't changed in nearly three decades: bad data yields bad results. Treating data as a back-office IT chore instead of a core business asset is actively tanking enterprise AI adoption.
In this episode of Don't Panic! It's Just Data, host Kevin Petrie sits down with 28-year data veteran Poornima Ramaswamy to break down how to build an AI data foundation that actually holds up under real-world scrutiny.
AI Success Starts With Data
AI captures the headlines, but the underlying mechanics haven't changed in nearly three decades.
With a backing of over 28 years working in data, she has seen first hand as the industry moved from decision support systems and management information systems to business intelligence, advanced analytics and machine learning to today's agentic AI.
Across the evolution from management information systems and business intelligence to machine learning and today’s agentic AI, the core dependency remains identical: bad data yields bad outcomes.
Poornima explains that organisations have spent too much time looking at AI through an AI lens. Simultaneously they treat data as a technology problem rather than a source of business value. That disconnect can make it difficult to realise the full potential of AI.
Governing Data for Physical AI
Operating in physical environments where decisions can affect safety, operations and significant financial value is one of the most challenging applications of AI.
Poornima shares a case study involving a global railway operator developing a platform for physical AI and predictive algorithms. The organisation recognised that simply building the technology platform was not enough. It also needed to prove that the data feeding its models was authoritative, trusted and traceable.
The challenge was particularly complex because the organisation had to account for both IT and operational technology data, while also preparing for different regulatory requirements across geographies.
Rather than creating bespoke governance for every customer, PivotX helped the organisation develop a minimum viable governance foundation that could become part of the platform itself.
The Synthetic Data Challenge
The most important scenarios are often the ones organisations have the least real-world data for.
There may be extensive data showing what normal operations look like, but comparatively little data showing rare failures, safety incidents or other undesirable outcomes.
That is where synthetic data becomes valuable.
Poornima explains how synthetic data can help generate the scenarios required to train AI systems, from safety risks in industrial environments to rare diseases and fraud detection.
But synthetic data creates its own governance questions. Should it be governed in exactly the same way as real data? How much rigour is appropriate? And how can organisations avoid creating so much governance overhead that it slows down AI development?
The answer is not to apply a one-size-fits-all approach.
When Vector RAG Stops Working
Why retrieval strategy now hinges on matching Vector, Graph and hybrid RAG to the questions AI must answer across complex enterprise data.
Centralised or Federated AI Governance?
The conversation also explores how governance models need to evolve as AI moves deeper into business operations.
While centralised governance can provide common policies, standards and oversight, Poornima argues that agentic AI increasingly requires a more federated approach.
Business processes rely heavily on unstructured data and domain-specific knowledge. Operational teams often understand that context better than a central IT function.
The emerging model is therefore a balance: establish common governance principles centrally while giving business units and operational experts the ability to apply domain-specific rules.
Why Master and Reference Data Are Back in Vogue
One of the episode's most important takeaways is the renewed importance of master and reference data.
As organisations deploy AI across business processes, consistent definitions and shared context become increasingly important. Master and reference data can provide a common foundation that helps different AI use cases learn from one another rather than becoming isolated pockets of intelligence.
Poornima argues that data professionals need to move further upstream, closer to source systems, and treat enterprise data as something that continuously evolves through the lessons learned from AI deployments.
The result is a broader view of AI readiness: not just better models, but better data foundations.
Inside the Agentic SOC Stack
See how unified telemetry, correlation engines and agentic AI workflows rebuild SOC architecture for autonomous detection and response.
Takeaways
- AI adoption depends on the quality, governance and lineage of the data behind it.
- Physical AI creates particularly demanding data governance requirements because decisions can affect safety and operations.
- Synthetic data can fill critical gaps when real-world examples of rare or harmful scenarios are limited.
- Synthetic data should not automatically be governed with the same level of rigour as real-world data.
- Global AI deployments need common governance foundations with flexibility for local regulations.
- Industry-specific regulations can add another layer of complexity to AI governance.
- Agentic AI is pushing organisations towards more federated governance models.
- Operational SMEs need a role in defining and applying governance frameworks.
- Master and reference data are becoming increasingly important as AI moves into business operations.
- Organisations should think about data governance as an evolving enterprise capability, not a one-off compliance exercise.
Chapters
- 00:00 Introduction to Data and AI
- 01:43 Poornima's Data Journey and the Origins of PivotX
- 03:27 Why Data Remains the Foundation of AI
- 05:58 Governing Data for Physical AI
- 08:47 Why Synthetic Data Matters
- 11:11 Building Governance Across Global Regulations
- 12:43 The Business Impact of Embedded Governance
- 14:23 Managing False Positives and Anomalies
- 15:48 Best Practices for Synthetic Data Governance
- 18:37 Centralised vs Federated AI Governance
- 21:34 Why Master and Reference Data Matter
- 23:32 Turning AI Learnings Into Enterprise Knowledge
Comments ( 0 )