I spoke with Jason Mills, Global Senior Vice President of Solution Engineering at Cloudera, at the company’s recent EVOLVE 2025 event in New York City. In our conversation, Mills noted that today’s “failed” AI pilots are valuable lessons pushing enterprises toward repeatable, internal wins. He highlighted document automation and fraud detection as sticky AI use cases, underscored the need to “bring AI to the data” across silos, and made the case for true hybrid optionality—on-prem where it stabilizes cost, cloud where it scales.
Mills outlined a pragmatic framework for AI cost control and explained how Cloudera’s open, interoperable platform (bolstered by recent acquisitions like Taikun and Octopai) helps customers govern data estates, ensure sovereignty, and deploy agentic workflows at enterprise scale.
Core Takeaways
- Repeatable value is showing up inside the enterprise. Beyond consumer chatbots, internal LLMs that read documents and automate complex workflows—like fraud detection and document processing—are delivering durable ROI.
- Break the silos, bring AI to the data. Cloudera positions itself as an open, hybrid platform that unifies access to legacy and modern data—on-prem and in the cloud—so governance and interoperability come first.
- Hybrid = choice on cost, performance, and sovereignty. Some workloads are cheaper and safer on-prem (including sovereign and regulated data); others belong in the cloud. The point is workload placement by design, not by habit.
- Cost discipline is a process, not a line item. Plan end-to-end: data prep/labeling, use-case/model selection (including agentic), hidden infra costs (GPUs, storage, egress), and ongoing ops/retraining—often starting on-prem and bursting to cloud.
Key Quotes
“Failure is a feature—pilots teach you what sticks.”
“I actually see the high rate of early AI ‘failures’ as a positive. It means customers are discovering which assumptions don’t hold and where the real value lives. When you move past the splashy B2C chatbots, what sticks are the internal use cases—LLMs that understand your documents and workflows, read at superhuman speed, and help teams act faster.”
“That’s why we’re seeing gravity around things like document automation and fraud detection. Enterprises are lining up their data estates with specific use cases instead of chasing vague promise. The lesson is simple: learn fast, iterate, and focus on the workflows you control.”
“Bring AI to the data—governance and interoperability first.”
“Cloudera’s mission is to unlock data across silos, whether that data lives in legacy systems, a warehouse, or modern lakehouse patterns. Our open architecture is designed to interoperate—so you can govern once and apply AI broadly, instead of copy-pasting data everywhere.”
“We’re not trying to be the ‘end-all, be-all’ platform. We’re trying to be the enterprise-grade fabric that makes data access, governance, and portability smooth—on-prem or in the cloud. That’s where open standards and projects like NiFi matter, and why acquisitions like Taikun and Octopai strengthen deployment and visibility across the entire estate.”
“Hybrid optionality beats one-way cloud economics.”
“Cloud costs tend to scale up naturally. On-prem lets you stabilize costs—run leaner or longer, invest as needed—and choose what truly belongs in the cloud. With Cloudera, customers decide: keep a production workload on-prem to control compute costs, and move other jobs to cloud when it makes sense.”
“It’s not cloud versus on-prem; it’s choosing the right placement for security, sovereignty, and economics. Many governments and regulated industries want sovereign cloud control with a cloud-like experience. That’s exactly the kind of flexibility we aim to provide.”
“Treat AI spend as a lifecycle—plan before you scale.”
“There’s no single formula for AI cost, but there is a discipline: understand the cost of data prep and labeling; decide which use cases to automate and which models or agents to use; account for hidden infra costs; and plan for operations—monitoring, feedback, and retraining.”
“If you open a cloud chatbot to the world without guardrails, usage and cost can snowball. Our guidance is to start in controlled environments—often on-prem—prove value, then burst to cloud as demand and SLAs require. That’s the only way to match speed with control and keep budgets in bounds.”
