New Delhi, Sept. 7 -- Enterprise AI adoption is racing ahead of the data infrastructure needed to support it, creating a new bottleneck for companies trying to move from pilots to production: trusted enterprise data.
While 88% of organisations globally are now using AI in at least one business function, just 7% say their data is fully AI-ready, according to QuantumBlack AI by McKinsey. The gap is becoming particularly critical as companies move towards agentic AI, where systems are expected not just to generate outputs but make decisions and take actions across business processes.
"The biggest barrier to scaling AI is no longer the model. It is the readiness, governance and trustworthiness of enterprise data," said Dinesh Chawla, executive vice president and chief data and analytics officer at TransUnion, quoting the industry report.
AI's data problem is becoming a scaling problem
For enterprises, the problem is increasingly less about access to sophisticated AI models and more about whether those models can reliably access the right data, understand its context and operate within appropriate controls. Data silos, inconsistent data quality and fragmented governance are emerging as constraints on AI scale. Cloudera estimates that 56% of global organisations cite data silos as a major obstacle to realising AI value.
The challenge is amplified by agentic AI. A model generating an imperfect answer is one problem; an AI agent acting on incomplete or unreliable information is another. "When data is fragmented, trust in those actions quickly erodes," Chawla said.
According to Chawla, the failure of AI projects to progress from experimentation to production is therefore less about model capability and more about the environment in which the model operates. Enterprises can access increasingly capable foundation models, but scaling them requires data integration, quality, governance and accountability.
He identifies three persistent obstacles: fragmented data across enterprise systems, poor data quality and readiness, and governance. Many AI pilots work because teams manually clean and curate datasets. That approach, however, becomes difficult to sustain when AI is deployed across multiple functions and processes. "A world-class model cannot compensate for poor data quality, fragmented systems or a lack of trust in the results," Chawla said.
For large enterprises, building a trusted data foundation also involves more than data quality. It requires lineage, metadata, identity and master-data management, access controls, governance and, increasingly, real-time access to information. The difficult part, Chawla said, is often not the technology but establishing common definitions, ownership and policies across business units, legacy environments and acquired platforms. That makes creating a trusted view of a customer, product or transaction an organisational challenge as much as an architectural one.
The arrival of agentic AI raises the stakes further. Traditional data governance has largely focused on controlling who can access information. Enterprises now need to govern what AI systems can do with that information.
"Governance must evolve from a focus on data access to a focus on data, decisions and actions," Chawla said. That means giving AI agents clearly defined identities and permissions, ensuring human intervention for high-risk decisions, and maintaining end-to-end auditability of the data accessed, decisions made and actions taken by an agent.
The governance gap is already visible. Many organisations planning agentic AI deployments say they have a mature governance model.. For TransUnion, where AI can influence decisions related to credit, fraud, identity and risk, explainability is consequently a core requirement rather than a trade-off against model performance.
"A highly accurate model that cannot be explained, validated or governed has limited value in regulated and high-stakes environments," Chawla said.
The company uses reason codes and model explanations alongside monitoring for bias and model drift, with governance embedded across the AI lifecycle. The objective, he said, is to ensure that business users, regulators and consumers can understand and validate AI-driven outcomes.
India GCC takes on greater AI ownership
The shift towards AI-led decision-making is also changing the role of TransUnion's India GCC. Chawla said the centre has moved beyond a traditional technology and analytics delivery model and is taking greater ownership of global platforms, data and AI capabilities. India-based teams contribute to OneTru, TransUnion's centralised solution-enablement platform supporting credit, fraud, marketing and consumer solutions globally. The GCC is also expanding its role across AI, cloud transformation, data engineering, fraud, security and intelligent automation.
"The focus is no longer just on executing requirements, but on shaping platforms, solving complex business problems, driving innovation and influencing product roadmaps," Chawla said.
TransUnion is following a similar principle in deciding what to build internally and where to rely on external technology. It intends to retain ownership of capabilities that differentiate the company-its proprietary data, identity intelligence, fraud insights, analytics and governance-while using external foundation models and cloud infrastructure where they can accelerate development.
Its collaboration with Google Cloud, including the use of Gemini models through Vertex AI, is an example of this approach. Chawla said the competitive advantage comes from combining external models with TransUnion's proprietary data, domain expertise, governance and semantic context rather than building foundation models itself.
Over the next 12-24 months, the focus will increasingly shift from building AI foundations to scaling them. Chawla expects greater adoption of AI-enabled analytics, automation and agentic capabilities, alongside continued investment in cloud, data engineering and platform capabilities.
For the India GCC, that is likely to mean deeper involvement in global products and platforms rather than simply expanding delivery capacity. The broader enterprise lesson, Chawla said, is that AI readiness will increasingly be determined not by access to the latest model, but by whether organisations can make their data reliable, governed and usable at scale.
Published by HT Digital Content Services with permission from TechCircle.