Why Your Enterprise AI Keeps Failing  And It Has Nothing to Do With the Model

Researcher and architect Milan Parikh has spent 15 years diagnosing the same problem the industry keeps misdiagnosing. His answer: stop blaming the model and start fixing the infrastructure around it.


Enterprise AI has a production problem. Not a model problem  a system problem. And the industry has been looking in the wrong place.

The conventional post-mortems cite familiar culprits: dirty data, talent gaps, organizational inertia. These aren’t wrong. But they’re also not the root cause of why sophisticated AI models, built by talented teams and trained on carefully curated datasets, consistently underperform when deployed in real enterprise environments.

Milan Parikh, Lead Enterprise Data Architect at Cytel and a 15-year veteran of the infrastructure layer where production AI actually lives, has a more precise diagnosis: the pipelines, governance architecture, and data lineage mechanisms surrounding a model determine its real-world ceiling  and most enterprises are ignoring them.

“The model is a component,” Parikh argues across his published research. “The system surrounding it is where deployment outcomes are actually determined.”

FraudSentinel: A Production Problem That Became a Research Finding

Parikh’s 2025 IEEE paper on FraudSentinel didn’t start in a lab. It started in production.

The engineering challenge was deceptively practical: how do you build fraud detection that holds up when the data it depends on is messy, transformed, and arriving through pipelines you don’t fully control? That’s not an abstract research question  it’s the daily reality of every financial institution running AI at scale.

The FraudSentinel architecture arrived at a finding that extends well beyond fraud detection: in production financial systems, model reliability is governed more by conditions upstream of the model than by the model architecture itself. Data quality, event ordering, and pipeline integrity set the actual performance ceiling  regardless of how carefully the model was designed or trained.

The implication is uncomfortable for organizations that have invested heavily in model selection and fine-tuning. You’re optimizing a component. The system around it  pipelines, governance, data lineage  is where the outcome is determined. FraudSentinel’s answer was to build governance into the pipeline layer from the start, not bolt it on afterward. Auditability, lineage tracking, and adversarial resilience became structural properties of the infrastructure, not afterthoughts.

TrustGraph: Making the Invisible Visible

If FraudSentinel identified the problem, TrustGraph  Parikh’s second 2025 IEEE publication  is his attempt at a general solution.

Here’s the enterprise AI reality that rarely makes it into vendor decks: the model you deploy against production data is almost never seeing data in the same condition as the data you trained it on. Between training and inference, data passes through transformation layers, vendor systems, reconciliation processes, and handoffs that introduce variability at every stage. That divergence  training conditions vs. inference conditions  is one of the most persistent root causes of production AI underperformance, and it’s typically invisible at the model layer.

TrustGraph uses graph-based trust propagation across distributed data entities to solve this. As data moves through the pipeline, TrustGraph maintains and updates a trust record at each stage  building an auditable provenance map by the time data reaches inference. When a model’s output degrades unexpectedly, you get an audit trail, not guesswork.

Crucially, TrustGraph was designed for the messy, multi-vendor, cloud-native infrastructure that real enterprise environments actually run on  not for the clean conditions of academic benchmarks.

Event-Driven Architecture as the Foundation  And Why Cytel Bet On It

Both FraudSentinel and TrustGraph share an underlying architectural principle that also defines Parikh’s production work at Cytel: event-driven design as the governance foundation for enterprise AI infrastructure.

The logic is clean: when every state change in a system is recorded as an immutable event, the exact data conditions that produced any model output become reconstructable. Production failures can be replayed. Pipeline failures can be pinpointed. The audit trail survives the complexity of multi-vendor, multi-system environments at enterprise scale.

At Cytel  a global clinical research organization with over 1,500 staff Parikh led the Microsoft Fabric implementation across global operations. The event-driven governance architecture he built delivered documented cost savings of $100,000 to $250,000 and established the data lineage and observability infrastructure now supporting AI deployment across clinical research workflows. In a clinical research environment, data provenance isn’t a nice-to-have. It’s a regulatory requirement.

The Credentials Behind the Research

Parikh holds 14 publications on IEEE Xplore  four published in 2025 alone  covering fraud detection architecture, distributed trust propagation, unified data management, and reinforcement learning for CI/CD pipeline automation. He is a Fellow of the British Computer Society, an IEEE Senior Member, and Secretary of the BCS Enterprise Architecture Group.

In 2026, the CES Innovation Awards appointed him to its judging panel  one of the most prominent technology recognition platforms in the world. That sits alongside IEEE Computer Society recognition, the Conrad Challenge, and 203 peer reviews for AI conferences spanning eight countries including the UK, US, UAE, and India.

A March 2026 Tech Times feature traced the connection between his published research and his production architecture decisions at Cytel, citing FraudSentinel, TrustGraph, and his CI/CD reinforcement learning work by name. It’s the kind of coverage that follows a sustained body of work not a single paper or viral blog post.

The Bottom Line

Enterprise AI’s production problem isn’t going away on its own. And it won’t be solved by better models, bigger training runs, or more sophisticated evaluation frameworks if the infrastructure layer around those models remains an engineering afterthought.

Parikh’s research grounded in fifteen years of production deployments  offers a different frame: treat the infrastructure layer as a first-order engineering concern, design governance in from the start, and measure success at the system level, not just the model level.

It’s a less exciting pitch than the latest model release. It’s also, the evidence increasingly suggests, the one that actually works.

:::tip
This story was distributed as a release by Jon Stojan under HackerNoon’s Business Blogging Program.

:::

Leave a Comment

This site uses Akismet to reduce spam. Learn how your comment data is processed.