Inside the Data Foundation Problem Behind Enterprise AI Failure: Milan Parikh Takes the Case to Data Summit 2026
For most enterprise technology leaders, artificial intelligence has become the top strategic priority. The results have been mixed.
Milan Parikh thinks he knows why.
“The bottleneck is not the model,” said Parikh, a lead enterprise data architect and a judge for the CES Innovation Awards 2026. “Organizations invest heavily in AI and then wonder why outcomes disappoint. The answer is almost always in the data foundation, not the algorithm.”
Parikh, a Fellow of the British Computer Society, serves as Secretary of both the Enterprise Architecture Specialist Group and the South Wales Branch, and as a Book Reviewer for the society. A Founding Member and Chair of the ACM Houston Chapter and an IEEE Senior Member, he presented his case this month at Data Summit 2026, the 13th annual conference on data management and AI held at the Hyatt Regency Boston. His session, “Revolutionizing Data Value Chains: Medallion Architecture Meets Microsoft Fabric,” drew on peer-reviewed research and 15 years of hands-on implementation experience across life sciences, manufacturing, and financial services to argue that most enterprises are solving the wrong problem.
“Parikh’s session was insightful and addressed both context and results very well,” said Amit Kumar Padhy, speaker and attendee at Data Summit 2026.
The numbers behind his argument are difficult to dismiss. Industry data and his own research indicate that 60 to 70 percent of data teams run duplicated pipelines across departments. Time-to-insight lengthens three to four times when governance is bolted on after the fact rather than built into architecture from day one. Poor data quality costs large enterprises an average of $12.9 million annually.
“Teams build isolated pipelines independently,” Parikh said. “Each handoff creates a new copy of the data. There is no shared lineage and no consistent quality. Governance is added last, if at all. By the time you layer AI on top of that, you have already compounded the problem.”
A Framework for Fixing It
Parikh’s proposed solution centers on what practitioners call Medallion Architecture, a three-layer approach to organizing enterprise data that has gained significant traction among large organizations building AI-ready infrastructure.
The framework divides data into Bronze, Silver, and Gold layers, each representing a progressively higher level of quality and business readiness. Raw data lands in Bronze untouched, preserving source fidelity as an immutable audit baseline. Silver applies transformation, schema enforcement, deduplication, and quality controls, turning raw records into a shared foundation that downstream teams can consume without re-verifying the original source. Gold delivers business-ready analytics models and the structured datasets that serve as the direct input for machine learning and intelligent applications.
“Bronze is your audit baseline,” Parikh explained during the session. “When downstream logic fails, you always have the truth. Silver is the enterprise foundation. Teams consume it without checking the source.”
The distinction he draws between pipelines and value chains sits at the center of his argument. Pipelines move data from point A to point B, optimized for throughput, with no quality contract and governance treated as an afterthought. Value chains refine data through controlled stages, with quality gates at every transition, assets engineered for multi-team reuse, and governance embedded in the architecture itself.
“Gold datasets are your AI foundation,” Parikh said. “Structured data now means faster AI adoption later.”
Microsoft Fabric as the Delivery Vehicle
At the implementation level, Parikh mapped each Medallion layer to specific capabilities within Microsoft Fabric, Microsoft’s unified data and analytics platform, making the case that enterprises no longer need five separate tools to execute the framework.
Data Factory Pipelines handle raw ingestion into OneLake at the Bronze layer. Eventstream supports real-time streaming for operational and IoT workloads. Notebooks running PySpark or SQL manage complex Silver transformations, while Dataflows Gen2 provides low-code cleansing for teams without deep engineering capacity. Power BI Semantic Models serve the Gold analytics layer with DirectLake connectivity, eliminating the data copy that traditional reporting architectures require.
Governance, Parikh argued, must be treated as structural rather than optional. His framework covers four dimensions: data lineage tracked from origin to consumption through automated integration with Microsoft Purview; quality controls enforced at the schema level with reject queues for non-conforming records; access control implemented through row-level security and column-level masking at the Gold layer; and a certified data catalog that makes assets discoverable and trustworthy across teams.
He closed the session with a six-phase implementation blueprint for enterprises starting from scratch: scoping priority domains, building Bronze landing zones with change-data-capture patterns, establishing Silver quality contracts, modeling Gold domain products with Power BI semantic models and access controls, connecting governance and observability tools, and scaling by onboarding additional teams and unlocking AI use cases from the governed Gold layer.
The Researcher Behind the Framework
Parikh’s Data Summit appearance reflects a career that has run on parallel tracks of research and practice. His peer-reviewed work spans 25 papers indexed on IEEE Xplore across 11 international venues, including a journal article in IEEE Transactions on Consumer Electronics, a publication with an impact factor of 10.9. His research has examined federated learning architectures for fraud detection, multi-model database trade-offs for AI workloads, graph neural networks for enterprise trust modeling, and reinforcement learning for adaptive pipeline optimization.
A recent study co-authored by Parikh found that multi-model database systems outperformed both single-model and polyglot alternatives on a Composite Performance Index measuring speed, flexibility, and reliability, a finding reported by Zee News in May 2026 and directly relevant to how enterprises structure data before AI deployment begins.
He holds fellowships from three professional societies. The British Computer Society, one of the world’s oldest and most recognized computing institutions, admitted him as a Fellow, a designation held by roughly the top 6 percent of its 46,000+ members, and he also serves as Secretary of the society. He holds a Fellowship from the Institution of Electronics and Telecommunication Engineers and a Distinguished Fellowship from the Soft Computing Research Society, an invitation-only designation held by approximately 100 of the organization’s 14,000 members globally, representing less than 1 percent of total membership, where he additionally serves as an Assessor. He holds IEEE Senior Member status, a grade requiring demonstrated significant professional performance over at least ten years.
His international conference record includes session chair and speaker roles at IEEE conferences across six countries. He has judged the CES Innovation Awards, the Consumer Electronics Show’s global program recognizing technological advancement, and the IEEE Computer Society Awards across two cycles, evaluating submissions from researchers and engineers worldwide.
Why Boston, Why Now
Data Summit, produced by Database Trends and Applications, has convened data professionals annually since 2014. The 2026 edition drew senior data architects, chief information officers, and enterprise technology strategists for sessions spanning modern data architectures, data engineering, analytics, and generative AI. Co-presenters at the conference included engineers and executives from the US Environmental Protection Agency, Morningstar Investments, IBM, and GoodRx.
The timing of Parikh’s session is deliberate. Enterprise AI spending continues to rise while reported success rates remain stubbornly low. Research cited by Parikh indicates that 34 to 42 percent of enterprise AI initiatives fail to achieve their intended outcomes, not because of model limitations, but because of weak data foundations.
“Enterprises are asking the wrong question,” Parikh said. “They ask which AI model to deploy. The prior question, the one that determines whether any model succeeds, is whether the data architecture can support it.”
Milan Parikh is a lead enterprise data architect, an IEEE Senior Member, and a peer-reviewed researcher with 25 publications indexed on IEEE Xplore. He is a Fellow of the British Computer Society and serves as Secretary of both the Enterprise Architecture Specialist Group and the South Wales Branch, and as a Book Reviewer for the society, a Fellow of the Institution of Electronics and Telecommunication Engineers, and a Distinguished Fellow and Assessor at the Soft Computing Research Society. He is the Founding Member and Chair of the ACM Houston Chapter and a judge for the CES Innovation Awards 2026.
Data Summit 2026 was held May 6–7 at the Hyatt Regency Boston. Full program details are available at https://www.dbta.com/DataSummit/2026/32182-Milan-Parikh.aspx
