Share with your CDO
Balaji Raghunathan, Data and AI Engineering leader at Sigmoid, makes a pointed argument: most enterprise AI programs stall not because the models are wrong but because the data underneath them was never ready. Writing for Nasscom’s Leader Talks series, he details how a consumer packaged goods company with $40M-plus in annual retail transactions watched a multimillion-dollar cross-sell AI program freeze entirely because 20-plus retailers across five markets each categorized products differently. The fix required an AI-driven quality accelerator that scored records and routed low-confidence ones to human stewards. The program restarted. The broader lesson is structural, not tactical.
What this means for your business
The recurring failure mode looks like this: an organization approves an AI budget, selects a model vendor, and begins building, only to discover six months in that the training data is too fragmented to produce reliable signals. At that point, the cost of remediation is three to five times what it would have been at project inception. Raghunathan’s argument is that data readiness is the gating condition for AI ROI, and it deserves funding before the model conversation starts, not after the PoC collapses.
The most underappreciated point in the piece concerns governance as a continuous operating discipline rather than a one-time configuration. Static rule sets written against known internal data sources break immediately when AI use cases demand unified feeds from external partners, real-time streams, and unstructured content simultaneously. Organizations still treating data quality as a project with a completion date are essentially betting that their data environment won’t change. It will. Dynamic, automated monitoring with human review reserved for genuine judgment calls is the only architecture that scales alongside agentic AI workloads.
The signal worth watching: agentic AI raises the stakes on data governance from important to existential. When an agent doesn’t just read data but acts on it, modifying records, triggering workflows, sending communications, a single bad data state can propagate across systems before any human sees it. The four metrics Raghunathan recommends tracking (AI adoption rate, time to insight, rate of data incidents, and business outcomes attributable to AI decisions) are the right instrumentation, but most data organizations are currently measuring only the first two. The question worth holding is whether your governance framework was designed for a world where software reads your data or one where software acts on it.
Concept deep-dive: Semantic layer
A semantic layer sits between raw data stores and the AI systems or analysts querying them. It translates technical database structures into consistent business terms, “revenue” means the same thing whether the query comes from a sales agent or a finance model. It exists because enterprises accumulate data across dozens of systems that use different naming conventions for identical concepts. Think of it as a universal translator bolted above your data warehouse. For agentic AI specifically, the semantic layer is what allows an autonomous agent to query live business context without needing a data engineer to interpret the request first.
Based on reporting from The Foundation of Enterprise AI Lies in Trusted Data || Balaji Raghunathan, Data & AI Engineering Business Unit Leader, Sigmoid | Nasscom, originally published 2026-07-23 03:16:00.
