Naveen Zutshi (Databricks): What separates abandoned AI pilots from real results

Most AI transformations fail before they reach production – not because the technology doesn't work, but because organizations lack the foundation to support it.

Naveen Zutshi has spent his career building that foundation.

As CIO of Databricks, Naveen leads IT transformation for one of the most influential data and AI companies in the world, working with Fortune 500 enterprises that process millions of datasets a day. Before Databricks, he drove technology strategy at Palo Alto Networks and Gap, consistently focusing on how organizations can move from experimentation to production at scale.

In this conversation with Pigment co-CEO Eléonore Crespo, Naveen explains why data maturity matters more than industry when predicting AI success, how CIOs should focus on clarity rather than control, and why fully autonomous AI remains more hype than reality.

Why nearly half of AI projects never make it to production

When Naveen talks about AI transformation, he starts with a number that should concern every enterprise leader: 46% of AI proof-of-concepts are abandoned before they reach production.

The root cause isn't technical sophistication or model quality. It's data. Organizations ship data to models without unified governance, often using datasets that aren't clean enough to support reliable outcomes. The result is failure at scale, repeated across industries and use cases.

The second critical factor is evaluation. Teams need to know what they're trying to achieve with precision, not aspiration. That means defining tests upfront based on the output they need, not the process they'll use to get there.

This approach flips the typical development cycle. Instead of building first and testing later, organizations should define success criteria before writing a single line of code. When evaluation is clear and data is clean, AI projects move from experiments to production outcomes.

Data maturity matters more than industry

When asked which industries are seeing the fastest AI adoption, Naveen's answer surprises most people. It's not about the sector. It's about data maturity.

He points to three companies as examples, all in heavily regulated industries that many assume would move slowly on AI:

• First, Mastercard uses Databricks to process credit card applications and customer onboarding. The work happens generatively now with human oversight and governance, allowing them to handle massive transaction volumes while maintaining compliance.
• Second, Rolls Royce fuses IoT data from aircraft engines with environmental data to predict and prevent maintenance outages. The result isn't just operational efficiency. It's millions of tons of reduced carbon emissions.
• Third, Walgreens optimizes drug distribution across 9,000 pharmacies using both agentic AI and traditional data science. They now process 825 million prescriptions annually, improving productivity by 20% and unlocking $160 million in additional value.

The common thread isn't what these companies do. It's how they manage data. They have proprietary datasets that are well-governed, clear use cases with business ownership, and strong evaluation frameworks. That foundation allows AI to deliver measurable outcomes rather than dissolving into abandoned pilots.

Focus on clarity over control

The traditional CIO playbook emphasizes control, including centralized decision-making, tight oversight, and standardized processes. Naveen believes that approach breaks down when organizations try to scale AI across the enterprise.

"Someone mentioned this to me, and I thought this is a pretty fascinating way of thinking. A CIO should focus on clarity, not control," he says.

Clarity means defining a vision for AI strategy and making sure every employee can participate in that journey. At Databricks, Naveen's team works with exceptionally smart people who constantly generate new ideas. Just minutes before our interview, someone showed him an intelligence engine they'd built for analyzing customer use case conversations. The idea was brilliant, and the team wanted to scale it across the company.

That kind of bottom-up innovation only works when there's a clear framework for how ideas get evaluated, scaled, and integrated into the broader platform. Without clarity, you get shadow AI – unsanctioned tools and processes that create data leakages, security risks, and unintended consequences.

Naveen recommends three core practices for maintaining clarity while enabling innovation:

1. Build an AI security framework. Work with legal, privacy, and security teams to define policies around what tools employees can use, how they can use them, and what's explicitly prohibited. The landscape moves fast, so the framework needs to be adaptable.
2. Create an internal AI strategy document that balances centralization with edge innovation. Some agents and capabilities should be built centrally, but teams also need the ability to build at the edge within a governed environment. Platforms like Databricks' Agent Bricks allow this by providing governance at the data and agent level through unified frameworks like Unity Catalog.
3. Invest in governance that covers bias detection, lineage tracking, auditability, and traceability. When agents make mistakes – and they will – you need to be able to trace what happened, understand why, and correct the system without losing trust.‍

Platform thinking outlasts point solutions

One of Naveen's strongest convictions is that organizations need a platform mindset, not a patchwork of point solutions.

"What is your platform of choice for AI?" he asks. "In early 2000, you had the LAMP stack for software developers. What is the LAMP stack equivalent for AI?"

Most organizations approach AI by adopting individual tools for specific use cases: one vendor for natural language processing, another for computer vision, a third for data preparation. The result is data copied across systems, inconsistent governance, and mounting integration complexity.

A platform approach solves this by keeping models close to data, allowing teams to choose the right model for each use case, and managing governance at the platform level rather than project by project.

At Databricks, this philosophy shows up in the architecture itself. The Lakehouse combines what used to be separate systems (data warehouses for structured data and data lakes for unstructured data) into a unified environment where both types coexist. Teams don't copy data between systems. They work on a single dataset with consistent governance, whether they're running SQL queries or training machine learning models.

The same principle applies to model selection. Organizations often commit to a single model provider, then discover that different use cases require different capabilities. A platform approach gives teams choices. They can test multiple models against the same dataset, evaluate performance, and optimize for both quality and cost without rebuilding infrastructure.

A platform that allows for that kind of dynamic optimization, rather than locking teams into fixed architectures, creates space for both innovation and efficiency.‍

How to empower teams without losing governance

Balancing innovation with governance is one of the hardest challenges CIOs face today. Give teams too much freedom, and you get shadow AI. Lock things down too tightly, and innovation stalls.

Naveen's approach relies on platform-level governance that enables edge innovation within clear boundaries.

For example, Agent Bricks allows teams to build conversational AI, process documents, and classify contracts without requiring expertise in agent architecture. The complexity is abstracted, but governance remains intact through role-based access controls, data lineage, and auditability built into the platform.

The key is making innovation easy while making risk hard. Teams can experiment and build quickly, but they can't bypass governance. Data access is controlled. Model behavior is logged. Actions are traceable.

Naveen is clear about where experimentation stops: agents that take actions on source systems – Salesforce, financial systems, HR platforms – are not allowed at the edge. And, even centrally, they require human oversight.

Humans won’t be leaving the loop anytime soon

The promise of fully autonomous AI agents is seductive, but Naveen is skeptical AI will operate without human intervention at enterprise scale anytime soon.

The reason comes down to reliability and risk. With large language models (LLMs), the same input can produce many different outputs. Small errors compound across multi-step workflows. In enterprise environments, where data is messy and actions need to be verified, the cost of getting things wrong is too high.

That doesn't mean AI can't automate work. At Databricks, two out of three IT support tickets are now deflected before reaching a human. Sales reps receive next-best-action recommendations directly in their workflow. These are real productivity gains, but they're AI-assisted with oversight built in.

What Naveen believes is underrated is governed functional calling over structured context. This approach gives models explicit functions they can call rather than relying on unstructured prompts. It's less flexible than open-ended generation, but far more consistent and auditable – exactly what enterprises need at scale.

When an agent uses a defined function, you can trace what it did, why it did it, and what data it accessed. You can test those functions independently and validate their behavior before production. For CIOs managing risk, the question is whether AI can perform a task reliably, day after day, with outputs you can explain to auditors and regulators.‍

Lessons from Databricks' own AI transformation

Naveen is no stranger to advising other organizations on AI strategy. But he’s also lived it at Databricks, where the company uses its own platform to run operations at scale.

One example is Genie, a text-to-SQL tool by Databricks that translates natural language questions into data queries. Instead of submitting requests to analysts and waiting weeks for dashboards, teams across marketing, finance, and sales can ask questions in plain English and receive structured answers right away. The impact shows up in headcount. What used to require a large data analysis team now runs with two people.

Databricks also uses its own platform for security information and event management (SIEM), processing massive volumes of logs and alerts to detect threats in real time. That same architecture powers customer data platforms and other applications for Databricks' clients.

The philosophy behind these projects is to start with a clear use case, ensure the data is clean and governed, define what success looks like before building, and keep humans in the loop for high-stakes decisions.

Naveen also emphasizes the importance of business engagement. Technical excellence matters, but AI projects that succeed have deep business ownership from the start. The technology team can't drive transformation alone. Business leaders need to lean in, define the outcomes they care about, and stay involved throughout implementation.‍

The data decision Naveen wouldn’t repeat

When asked about a decision he's not proud of, Naveen goes back to 2005, when he was evaluating CRM platforms at Cisco.

The team had to choose between Siebel on Demand and a relatively new company called Salesforce. They analyzed the data, looked at the rationale, and recommended Siebel. It seemed like the safe, logical choice.

"Obviously, in hindsight, it was a really terrible decision," Naveen says.

Fortunately, the business overruled IT and chose Salesforce, which turned out to be one of the best technology decisions the company made.

The lesson isn't that data-driven decisions are wrong. It's that data is only one input. Vision, risk tolerance, and the ability to see around corners matter, too. The best decisions combine rigorous analysis with strategic intuition – and the humility to know you might be wrong.

On the flip side, Naveen points to a decision he's proud of: moving to a Lakehouse-based architecture. That shift unlocked the ability to make decisions faster and build solutions more quickly by eliminating data silos and reducing the complexity of managing multiple storage systems.‍

What keeps Naveen optimistic

When the conversation turns to the future, Naveen's optimism around the possibilities AI brings to life is rooted in specific examples.

He points to AlphaFold, Google DeepMind's breakthrough in protein folding, as an example of what AI can achieve. 

"There is a huge opportunity in using AI to essentially transform and solve some of the biggest problems facing humanity itself," he says.

But he's also realistic about the risks. AI can be hijacked for theft, fraud, or worse. The doomsday scenarios are possibilities that require serious attention from security teams, policymakers, and the companies building these systems.

At a more practical, everyday level, Naveen sees AI reducing the cognitive load that employees face in modern enterprises. People spend too much time navigating systems, stitching together information from disparate tools, and waiting for approvals. AI can compress that work, giving people more time to think, create, and problem-solve.

Advice for leaders: say yes more

Naveen offers a piece of advice that applies beyond AI strategy: say yes more.

It's advice that applies to AI adoption as well. Leaders who default to caution – waiting for perfect clarity, proven ROI, or zero risk – miss the window where experimentation turns into advantage. The companies that succeed are the ones that say yes to pilots, yes to partnerships, and yes to learning by doing.

That doesn't mean reckless experimentation. It means creating frameworks where saying yes is safe: clear governance, strong evaluation, and platforms that make innovation possible without creating chaos.‍

Conclusion

AI transformation succeeds when organizations focus on the fundamentals – data quality, clear evaluation, unified governance, and platform thinking – rather than chasing the latest model or technique.

The companies that move from pilot to production are the ones with clean data, strong business ownership, and the discipline to start from the output they need rather than the tools they want to try.

For CIOs and business leaders navigating AI adoption, that's the actionable insight. Invest in data maturity. Define your platform. Focus on clarity over control. And recognize that autonomous agents, while exciting, remain aspirational for most enterprise use cases.

The opportunity is enormous, but it requires patience, rigor, and a willingness to build the foundation before racing to production.