Share with your CDO
The gap between AI pilots and production isn’t a model problem, it’s an infrastructure and integration problem. Srujana Sree Bathineni, Lead Data and AI Platform Architect at AMS IT Solutions, makes this case through her work building the AMS BackOffice Inventory Module, a retail system that connects invoice processing, expiration tracking, vendor credit workflows, and AI-assisted promotional recommendations into a single operational layer. Her core argument is that AI only earns its place when it plugs into existing processes and produces outputs users can act on without leaving their workflows.
What this means for your business
Most enterprise AI programs stall not because the models fail but because the data plumbing underneath them was never built for production load. If your organization has completed more than two pilots in the last eighteen months without a deployment that business users actually depend on daily, the bottleneck is almost certainly data pipeline reliability and system integration, not model selection. That’s a data leadership problem before it’s an AI problem, and it lands squarely on the CDO’s desk.
Bathineni’s framing of “deterministic processing plus AI where each is most appropriate” is the honest architecture most vendors won’t sell you. Rules handle known formats; AI handles edge cases. This isn’t a philosophical preference, it’s a risk management decision. Systems that route every judgment through a large language model inherit all of its variance, including the embarrassing kind that surfaces in front of a customer or auditor. The retail date-code problem she describes, where manufacturer codes vary wildly across suppliers, is a clean illustration of a category of messy real-world data that structured rules alone can’t fully cover but that AI alone can’t reliably own either. Hybrid routing is the engineering answer, and most data platform teams haven’t codified when to switch between the two.
The interview is a vendor-adjacent piece, published on a platform that sells reach to practitioners building their professional profiles, so Bathineni’s optimism about what integrated AI can accomplish in retail operations should be read against the backdrop of a single internal deployment rather than a multi-customer evidence base. That’s a scope limitation, not a reason to dismiss the architecture argument. The principles she outlines, modular interfaces, pipeline observability, human decision authority over AI recommendations, are well-established. Where I’d push back is on the implicit suggestion that governance “from the beginning” is achievable in most enterprises. In practice, governance gets retrofitted, and the real question is whether your team has built enough observability into existing pipelines to make that retrofit survivable. If you don’t have data lineage tooling in place today, that’s the budget call this story is actually forcing.
Concept deep-dive: Data lineage
Data lineage is the recorded history of where a piece of data came from, how it was transformed, and which downstream systems depend on it, think of it as a version-control log for information as it moves through your organization. It exists because AI models inherit errors from their inputs silently, with no compiler warning. When a model recommendation turns out to be wrong, lineage tooling is what lets you trace the fault back to a bad upstream field rather than blaming the model. Without it, debugging production AI is archaeology.
Based on reporting from Why Enterprise AI Stalls After the Pilot: Srujana Sree Bathineni on Building AI That Delivers, originally published 2026-10-28 13:20:00.
