Ninety-five percent of enterprise generative AI projects have not yet impacted the P&L. The pilots work. The demos impress. Then the project stalls because the model was tuned on clean pilot data that bears no resemblance to real production inputs, the system sits outside the workflow where employees actually make decisions, and no single person owns the ROI. The gap between strong pilot and business impact is not a technology problem. It is a decision architecture problem.
Praveen Siva, Program Manager at JPMorganChase, has spent over 16 years delivering enterprise transformation programs across banking, financial services, and fintech. His work spans SWIFT ISO 20022 rollouts across 58 countries, GenAI innovation, digital banking modernization, and automation-led change. Before his current role, he led multi-million-dollar transformation portfolios at Accenture, Capco, and Altisource.
"GenAI pilots don't fail because of the model they use. They fail because organizations treat AI models as some kind of tool experiment instead of redesigning how decisions, data, and workflows operate together," says Siva.
Siva says the organizations that reach production share a consistent pattern. They start with a business decision, not a technology question. Companies that succeed don't ask, where can we use AI? They ask, what business decision do we want to improve?" In banking, that looks like reducing account opening from seven days to under an hour, or cutting loan approval time from five days to one. The problem statement comes first. AI enters only where it serves the decision.
The second pattern is embedding AI directly into the workflow rather than layering it on top as a separate dashboard or recommendation tool. Siva describes a banking contact center solution where customer agents historically searched a knowledge base manually during calls, placing customers on hold for two to three minutes each time.
His team built an NLP engine that transcribed calls in real time and fed the transcript into an AI module that surfaced scripted answers from the knowledge base within seconds. "We didn't build a separate AI tool. We built AI into the actual process. The recommendation appears directly inside the system where the employee is already working."
He estimates the time savings across a thousand agents over a full year as substantial, and notes the solution succeeded because it was wired into the operational workflow rather than sitting alongside it.
Siva pushes back on the idea that organizations need to fix all their data before deploying AI. "They don't try to fix all enterprise data. They fix the data tied to the workflow they're improving." For fraud detection, clean up fraud-related data. For sanctions triage, organize sanctions alert data. The scope stays narrow enough to be actionable. The broader enterprise data estate remains a long-term project that should not block near-term production deployments.
The data gap between pilot and production is one of the most common failure points he encounters. Pilot data is curated and clean. Production data is messy, fragmented, and structurally different from what the model was trained on. "When it comes to real-world scenarios, the real-time data is not as clean or consumable as the pilot data. The model won't be tuned for that."
One accountable owner also matters. Siva sees the same pattern repeatedly: data pipeline teams own their piece, tech teams own the model, and business teams expect the outcome. "There is no single owner accountable for translating what you're building in AI into business value. No one owns the ROI." The organizations that break through assign a single leader responsible for the full result.
Siva frames governance as the point where most enterprise AI programs either accelerate or stall. The winning teams embed risk, compliance, and business stakeholders into the delivery team from day one rather than sending work through separate review committees. "If you bring those teams together, business understands how the model works, and the technical team understands what actually matters. The conversation shifts from 'is this safe?' to 'how can we make this work safely and effectively?'"
He illustrates the trust problem with a sanctions alert triage solution his team built. When presented to stakeholders, the pushback was severe. "Nobody was in favor. They saw it as playing with fire." The team abandoned the presentation approach, built a working MVP with human-in-the-loop review, and let users interact with it directly. Adoption followed.
"Trust doesn't come from dashboards or demos. It comes from seeing how the system works in real-world scenarios. Even if you start small, put it in use with guardrails. That's how confidence builds."
