By Leah O’neal 3/21/26
According to the World Quality Report 2025 published by OpenText and Capgemini in November 2025, only 15% of organizations have achieved enterprise-scale deployment despite 89% of organizations actively attempting to introduce generative AI in their workflows. The data from MIT’s State of AI in Business research is even more revealing: only 20% of enterprises that evaluated custom enterprise-grade AI systems made it to the pilot stage, and just 5% made it to full production. The gap between enthusiasm and impact is not a budget problem or a talent problem. The real issue is specificity. Generic AI tools are built to impress in controlled conditions, but business environments are rarely controlled. They may include legacy databases, undocumented formats, fragmented systems, and rules that exist not in writing but only in the institutional knowledge of people working with them. Until AI is built specifically to understand that complexity and operate within it rather than work around it, the solutions will remain in the pilot stage at best.
Raj Bhowmik has been navigating this gap consistently throughout his career. Based in Dublin, California, Raj is currently a Senior Machine Learning Engineer at Cognizant — one of the world’s largest IT and professional services companies — where he specializes in designing and deploying production-ready AI and GenAI systems for complex enterprise environments. His work spans retail, logistics, and multi-database enterprise stacks, where the data can be messy while the business stakes are high.
The core issue, as Raj sees it, comes down to what AI is actually being asked to understand. Most enterprises have plenty of data but suffer from it being structurally inaccessible — spread across incompatible systems or locked behind undocumented formats and legacy business logic. Adding a general-purpose AI layer on top of that infrastructure adds another layer to the problem rather than resolving it.
“Enterprise problems are often defined by systems that aren’t designed to talk to each other and by questions nobody thought to ask, while most AI demos are built to answer the more obvious and expected questions,” comments Raj Bhowmik. “To function on an enterprise level, an AI solution should be able to operate in a real and messy production environment.”
In practice, making enterprise data genuinely accessible means understanding how each system stores information, what format it uses and what business logic surrounds it. This work has to be done before a single line of AI code is written. It results in building interfaces that translate between the language of business users and the language of the underlying infrastructure, ensuring that the outputs are reliable enough to become the basis for real decisions. The AI product demos rarely focus on this work, but it is precisely what separates solutions that scale from those that stall.
“The goal is to build a practical solution — something that a non-technical business user could actually get answers from,” explains Raj Bhowmik. “That requires connecting it to real systems, with all their inconsistencies, while making sure the responses — which can include financial queries and other critical data — don’t contain AI hallucinations.”
The same philosophy applies to one of the more persistent challenges in enterprise operations: establishing reliable data exchange between systems and organizations. nIn large enterprises exchanging documents, such as purchase orders or shipping records, often tied to different platforms, involves layers of format translation, business rule and handling exceptions that generic AI tools are poorly equipped to manage. To create an AI tool that will solve these issues efficiently and reliably, it becomes crucial to map the workflow completely, understand its constraints, and then design a solution that will take them into account rather than expecting to operate in the idealized environment. This approach creates AI systems that streamline document exchange at scale, making it faster and more efficient, while being able to handle the edge cases and avoid errors.
Context-awareness matters just as much on the customer-facing side of enterprise operations. The Product Swap Bot Raj prototyped for a large retail environment demonstrates this clearly. The bot suggests substitutions when a customer requests a product that is currently out of stock, while accounting for multiple factors simultaneously — catalog rules, pricing logic, supplier agreements, and visual similarity — making it more likely that the customer will accept the replacement. The system uses both text and image inputs to evaluate candidate replacements against all of those factors at once, generating suggestions that fit within the operational reality of the business. The result is a tool that behaves less like a search engine and more like an experienced sales associate who knows the catalog, the margin, and the customer preferences.
A similar outcome, AI that enhances human judgement rather than replacing it, lies at the heart of a real-time sentiment analysis system Raj developed for picker-customer chat interactions at a major retailer. In high-volume retail operations, the quality of communication between store pickers and customers has a direct effect on satisfaction and operational efficiency, but monitoring and improving it at scale is not something human supervisors can do in real time. The system Raj built analyzes ongoing chat interactions as they happen, surfactant the right signals and information at the right moment within the existing workflow.
The pattern across all of these projects is consistent: AI earns its place in enterprise environments by operating within specific business rules, not around them. Recognition of this approach is growing beyond Raj’s immediate work at Cognizant. In 2025, his research was presented at two conferences, namely ICNGN 2025 (International Conference on Intelligent Computing) and FMLDS 2025 (International Conference on Future Machine Learning and Data Science) – both serving as peer-reviewed forums where researchers and industry practitioners from across the world present cutting-edge work in AI, machine learning, and next-generation computing. He has also served as a reviewer for an AI book and as a judge for AI and machine learning hackathons across the United States and India..
“Enterprise AI that works is built by someone willing to spend serious time understanding why the existing system operates the way it does — and then building something that fits inside that reality instead of trying to reinvent it,” concludes Raj Bhowmik. “The organizations that figure that out first are the ones that will be genuinely hard for competitors to catch.”
The broader lesson from Raj’s work demonstrates that the competitive advantage in the AI era will go to the companies willing to do the less visible work of making AI actually fit the business, which is not the same as gathering the most data. As the gap between AI enthusiasm and AI impact continues to define the enterprise technology conversation, the practitioners who know how to close that gap while working in real business environments with their constraints, layered rules and changing conditions, are becoming valuable for any enterprise company.
