According to MIT research, 95% of enterprise Generative AI projects have not yet impacted the P&L. That statistic doesn't point to a failure of AI technology. It reveals a failure in how organizations operationalize it. The pilots work. The demos impress. Then the initiative stalls. Models are trained on clean pilot data that bears little resemblance to production. They remain disconnected from the workflows where employees make critical business decisions, and no single business owner is accountable for the outcome or the ROI. The gap between a successful pilot and measurable business impact is rarely a technology problem. It is an enterprise operating model problem.
Praveen SIVA is an enterprise transformation leader and Program Manager at JPMorganChase with over 16 years of experience leading enterprise transformation across global banking, financial services, and fintech. He has led complex initiatives spanning SWIFT ISO 20022 implementations across 58 countries, Generative AI adoption, digital banking modernization, and enterprise automation. Before joining JPMorganChase, he managed multi-million-dollar transformation portfolios at Accenture, Capco, and Altisource, helping financial institutions execute large-scale business and technology transformation.
“GenAI pilots don't fail because of the AI models they use. They fail because organizations never redesign how the business, data, and workflows operate together," says SIVA.
He says the organizations that reach production share a consistent pattern. They start with a business outcome, not a technology question. Companies that succeed don't ask, "Where can we use AI?" They ask, "Which business decision or process will create the greatest measurable value if improved?"
In banking, that looks like reducing account opening from several days to under an hour, or cutting loan approval time from several days to a single day. The business objective comes first. AI is introduced only where it meaningfully improves the decision, process, or customer outcome.
The second pattern is embedding AI directly into core business workflows instead of deploying it as a standalone dashboard or recommendation engine.
He describes an enterprise contact center implementation where customer agents previously searched a knowledge base manually during calls, placing customers on hold for two to three minutes while locating the appropriate information. His team built an NLP engine that transcribed customer conversations and used AI to interpret them, surfacing the most relevant knowledge base response within seconds.
"We didn't build another AI application. We embedded AI into the operational workflow. The recommendation appeared within the same system where employees were already working, eliminating the need to switch between applications or interrupt the customer conversation."
Deployed across more than a thousand customer service agents, the solution delivered significant productivity gains while improving both employee efficiency and customer experience over the course of the year. The success wasn't driven by the AI model alone. It came from integrating AI into the existing business process, where it became part of how work was performed rather than another tool employees had to learn and adopt.
The contact center example illustrates a broader pattern. Drawing on more than 16 years of enterprise transformation experience, Praveen SIVA has consistently observed five characteristics that separate successful production deployments from pilots that never scale. He refers to this as the PACTS Framework. Together, these five principles provide a practical blueprint for moving enterprise GenAI from promising pilots to measurable business outcomes.
PACTS Framework
P - Production-first thinking
Start with a measurable business outcome rather than a technology use case.
A - Accountable business ownership
Assign one leader responsible for adoption, business value, and ROI.
C - Contextual data
Prepare only the data required for the business process instead of attempting to modernize the entire enterprise data estate.
T - Transformation of workflows
Embed AI directly into operational workflows instead of deploying another standalone application.
S - Shared governance
Integrate business, risk, compliance, legal, and technology stakeholders from the beginning to enable responsible innovation at scale.
He challenges the common assumption that organizations need to modernize their entire data estate before deploying AI.
"Organizations don't need perfect enterprise data before they can realize value from AI. They need fit-for-purpose data for the business process they're trying to improve."
For fraud detection, optimize fraud-related data. For sanctions screening, prioritize sanctions and compliance data. Focus only on the data required to improve the targeted business process. The broader enterprise data estate can continue to evolve in parallel and should not become a prerequisite for production deployment.
The disconnect between pilot and production data is one of the most common reasons enterprise AI initiatives fail. Pilot data is carefully curated, standardized, and complete. Production data is dynamic, fragmented, and often inconsistent across systems. As a result, models that perform exceptionally well during pilots frequently underperform once exposed to real operational environments.
"Production environments don't provide the clean, structured data used during pilots. If models aren't designed and validated against real-world operating conditions, production performance will inevitably suffer."
Technology alone doesn't deliver business value. Accountability does. He sees the same pattern repeatedly: data teams manage the pipelines, technology teams build the models, and business teams expect measurable outcomes. Yet no single leader is accountable for translating those capabilities into measurable business value.
"Everyone owns a component of the solution, but no one owns the business outcome. Without clear accountability, AI remains a technology initiative instead of becoming a business capability." Organizations that consistently move GenAI into production establish a single accountable business owner responsible for delivering measurable outcomes, adoption, and ROI, from pilot through production.
He views governance as the defining factor that determines whether enterprise AI scales or stalls. Successful organizations embed risk, compliance, legal, and business stakeholders into the delivery team from day one rather than routing solutions through disconnected review processes after development is complete.
"When these teams work together from the beginning, business leaders understand how the model works, and technology teams understand the business risks that matter most. The conversation shifts from 'Is this safe?' to 'How do we deploy this safely and effectively?'"
He recalls a sanctions alert triage solution that demonstrated how trust is built in practice. The initial response from stakeholders was overwhelmingly skeptical. "Nobody supported the idea initially. It was viewed as too risky because it touched a highly regulated process."
Rather than continuing with presentations and architecture discussions, the team built a working minimum viable product with human-in-the-loop controls and allowed business users to experience the solution in a controlled environment. Seeing the system operate with appropriate safeguards transformed skepticism into confidence and ultimately accelerated adoption.
"Trust isn't built through presentations or dashboards. It's built when people see the system perform reliably in real business scenarios. Start small, put strong guardrails in place, and let users build confidence through experience."
Enterprise AI has reached a turning point. Competitive advantage will no longer come from experimenting with better models. It will come from building operating models that consistently translate AI capability into measurable business value. Organizations that master this transition won't simply deploy AI faster. They'll redefine how work gets done.
