Clarion VP Palash Bharadwaj: Why Enterprise AI Stalls and Why Security Dies by a Thousand Time Zones

Clarion Technologies VP of Technology Palash Bharadwaj and Nathan Kirk on NR Labs Podcast Episode 3: Security Dies by a Thousand Time Zones

NR Labs Podcast, Episode 3 · Host: Nathan Kirk · Watch on YouTube

On Episode 3 of the NR Labs podcast, our cybersecurity podcast for technology and security leaders, host Nathan Kirk talks with Palash Bharadwaj, Vice President of Technology at Clarion Technologies. They cover why enterprise AI pilots stall, how to keep AI costs under control, when on-prem models make sense, and why, in Palash's words, "security dies by a thousand time zones, not by one big breach."

Watch and follow

Who is Palash Bharadwaj?

Palash Bharadwaj is Vice President of Technology at Clarion Technologies, based in Silicon Valley, where he leads technology strategy and AI transformation and advises enterprise clients on AI and cloud. He previously founded a medical IoT platform for heart health.

Why do enterprise AI projects stall?

Enterprise AI projects rarely fail because of the model. They stall in what Palash calls "the seam between the model and the organization." "It almost never stalls on a model," he said.

He sees three common causes:

  • Trying instead of committing. Many pilots are funded to spend this year's budget, with no plan for production. "Try versus really wanting to see a result are two different things."
  • No owner. If responsibilities aren't defined before the project starts, it fails no matter how good the technology is.
  • Resistance to change. Comfortable teams sometimes want the project to fail so they can return to the old way of working.

How do you turn an AI pilot into production?

Start with an assessment, not a new pilot. The real issue is often the data, not the AI. "Finding a solution is the easiest part, or the last part," Palash said.

In one example, a client had a proof of concept that worked but had no adoption, because nobody could say what it saved or who owned it. Clarion added a utilization, impact and cost scorecard and turned "an orphan pilot into a measured and owned initiative." "It has to be a joint effort from everybody."

How do you control enterprise AI costs?

Route each task to the right model instead of sending everything to the most expensive one. Token costs are now one of the biggest challenges for enterprises, and AI can end up costing more than doing the work with people.

Palash expects "AI factories" to pick the best model for each request automatically. "You should not even need to choose what to use. It should be seamless," he said.

Should enterprises use open source or on-prem AI models?

Often, yes, especially when data can't leave the organization. Palash's team chooses models by task and by data sensitivity. Some clients, including large customers in Saudi Arabia, want everything on-prem so their data is never exposed, and many now want a hybrid model where some data stays local and some goes to the cloud.

What does healthcare AI teach enterprise AI teams about risk?

"When the bug in your system is a missed heart failure signal, you learn a different relationship with risk than most enterprise AI teams ever have to," Palash said. During his heart-health platform's pilot, it flagged a patient who had just come home from surgery and was having a heart attack while driving. The team called 911. That's the standard of signal quality he brings to enterprise AI.

Why does security die by a thousand time zones?

Most security failures in global teams come from small handoffs, not one big breach. "It's a hundred small handoffs where nobody was quite sure whose job it is," Palash said.

His example: a ticket touching personal data is handed to an offshore team at the end of the day, which finishes the feature and ships it on a green build. Nothing flagged the field as sensitive. "Nobody skipped a step, but the step simply didn't exist."

The fix was moving data classification tags into code review so sensitive fields are flagged wherever the work travels. Building classification into everyday workflows is core to our Advanced GRC services, and stress-testing incident response is part of our CISO Advisory services.

Key takeaways

  • AI projects stall at the seam between the model and the organization, not in the model.
  • Define ownership and get stakeholder buy-in before the project starts.
  • Measure utilization, impact and cost to move a pilot into production.
  • Route tasks to different models by cost and by data sensitivity.
  • Put data classification tags into code review so handoffs across time zones don't leak sensitive data.

Watch the full episode

Hear the full conversation between Nathan Kirk and Palash Bharadwaj: Watch Episode 3 on YouTube. For more AI and cybersecurity podcast conversations, subscribe to the NR Labs YouTube channel and follow NR Labs on LinkedIn.

FAQ

Why do enterprise AI projects fail?
According to Palash Bharadwaj, they usually stall at the seam between the model and the organization: no clear owner, no measured value, data that isn't ready, and resistance to change.

How can companies control AI costs?
Route each task to the most cost-effective model instead of using the largest model for everything. Palash Bharadwaj expects AI factory tools to automate this routing.

What does "security dies by a thousand time zones" mean?
It means security failures in global teams usually come from many small handoffs where nobody owns a step, not from one big breach. Putting data classification tags into code review helps close those gaps.