Large organisations have reached a familiar crossroads with GenAI. The early rush of experimentation delivered promising pilots, but very few have translated those experiments into dependable, day-to-day systems. The challenge is no longer imagination. It is integration. Enterprise workflows carry years of embedded rules, manual exceptions, policy changes, and operational constraints that AI must interpret with precision rather than optimism.
Few technology leaders have navigated this shift as deeply as Anurag Jindal, a seasoned Technology Client Partner and VP at Vertisystem, and a Forbes Technology Council Member with 18+ years of experience across AI, Cloud, and Data technologies. Through his efforts, he has helped organisations move from fragmented experimentation to governed, measurable adoption, where AI is not a standalone feature but part of how the business operates. His work spans complex, multi-system programmes, including a multi-year modernisation effort for Stanford Medicine that rebuilt the operational groundwork required for future AI participation.
---Advertisement---
In this interview, he reflects on why GenAI initiatives struggle to scale, what separates a pilot from an operational capability, and how enterprises can design AI systems that are accurate, trusted, and ready for long-term use
---Advertisement---
Thanks for being with us today, Anurag. Can you tell us, from your experience across AI and data programmes, where do enterprises lose momentum with GenAI?
They lose momentum where pilots meet real conditions. A pilot succeeds because it avoids the environment. Data is curated. Exceptions are trimmed away. Ownership is clear for a few weeks. That controlled setup makes any model look stable.
The broader industry reflects the same pattern. As of 2025, nearly a quarter of enterprises have moved beyond experimentation: 23% are actively scaling agent-based AI across business functions, and another 39% have begun AI experiments. That jump shows that the challenge is no longer “Does GenAI work?”, but “Can we embed it reliably into complex enterprise workflows?”
Production exposes the truth. Enterprises carry years of conditional logic, undocumented workflows, fragmented data paths, and policy churn that no prompt can neutralise. When input contracts are unclear, when decision boundaries are implicit, and when systems interpret the same rule differently, GenAI does not stabilise the environment. It amplifies the inconsistency.
Most leaders assume the model is failing. It is not. The surrounding system is not deterministic enough for AI to participate safely. Until organisations confront that ambiguity, AI adoption stalls in the same loop: strong pilots, weak production.
What actually separates an impressive pilot from a production-grade AI system?
A pilot demonstrates possibility. A production system demonstrates reliability.
To reach reliability, you treat GenAI as a system participant, not an accessory. That means verified inputs, unambiguous workflows, and audit trails that explain what the model saw, why it acted, and how that action aligns with policy.
The second requirement is resilience to policy change. Most enterprises operate under rules that evolve monthly. If a model’s logic needs manual re-engineering every time a policy updates, the system will break faster than it learns.
The third requirement is measurement. Pilots track outputs. Production tracks correctness, exception paths, human override patterns, and downstream impact. When those metrics exist, AI becomes part of the operating model rather than a high-risk experiment.
This shift from demonstration to discipline is where most enterprises hesitate, but it is the only moment where AI becomes real.
You led a multi-system modernisation programme for Stanford Medicine. How did that work establish the conditions needed for responsible AI?
Before we talk about AI, the context matters. Academic medicine runs on complex, regulated, frequently changing workflows. If GenAI were introduced into that environment without clarity, it would inherit and amplify every inconsistency.
The multi-year programme addressed that head-on. We rebuilt core operations, NIH grants, donor endowments, faculty lifecycle, graduate financial aid, research recruitment, and clinical-trial coordination, on a rules-driven, multi-tenant architecture. Policies were externalised instead of buried in code. Dynamic UIs ensured that rule changes flowed into the system without rewrites. Data sync cycles dropped from more than twenty-four hours to four to six. Manual handoffs were replaced with event-driven integrations.
The result was determinism. Once the organisation had predictable workflows, clear lineage, and enforced auditability, AI could participate without destabilising compliance or decision integrity. The transformation was not about deploying AI. It was about engineering the environment where AI can function responsibly.
The financial impact followed naturally: reduced operational cycle times, fewer manual interventions, and improved accuracy across high-stakes processes. Stability becomes its own economic multiplier.
Governance has become central to enterprise AI. What does responsible governance look like at scale?
Governance is not a checkpoint. It is the architecture.
Responsible AI requires complete traceability: input lineage, model versioning, prompt conditions, exception flows, and structured escalation. If an organisation cannot defend a decision made by an AI-supported workflow, it should not run that workflow.
My experience evaluating impact-driven AI initiatives, including those recognised through the Globee Impact Award, reinforces this. The strongest programmes treat governance as a design requirement. Their reviews tie model behaviour to business standards, and their systems document every action with clarity. The weakest rely on notebook experimentation and hope the surrounding systems absorb the risk.
Enterprises operate under compliance, public accountability, and contractual obligations. AI must operate under the same constraints. When governance is engineered into the system, not stitched on later, innovation becomes stable instead of fragile.
You were recently invited to serve as a reviewer for the EmergIN 2025 Conference. How does that lens shape your view of GenAI’s next phase?
Reviewing research forces a longer horizon. The themes emerging now: sustainability, traceability, and system resilience, are exactly what enterprises will need.
Over the next few years, AI will move from experimentation to participation. Systems will expect AI to validate compliance, document reasoning, analyse architectural drift, and improve decision integrity. The focus will shift from building bigger models to building accountable environments around them.
The enterprises that prepare for that shift will scale AI safely. The ones that ignore it will remain trapped in pilot cycles that never cross into value. The gap will widen quickly.
The principle is consistent across every role I hold. Responsible AI is not a model problem. It is an engineering discipline problem. Organisations that embrace that reality will lead the next phase of adoption.
For leaders under pressure to show AI progress, what principle should guide them through the next two to three years of enterprise adoption?
The pressure to “show progress” often pushes organisations toward rapid deployment instead of responsible deployment. The principle I give every executive is straightforward: build conditions, not shortcuts. When leaders focus on speed alone, they end up scaling prototypes but when they focus on conditions, clarity of rules, quality of data, consistency of workflows and auditability of decisions, AI becomes sustainable instead of cosmetic.
The second part of the principle is equally important: treat every AI decision as an operating-model decision, not an IT project. When AI touches processes, it changes accountability, escalation paths, compliance exposure, and how teams interpret policy. That requires cross-functional ownership, not isolated experimentation.
Leaders who anchor themselves in these two points, conditions before scale and operating model before output, will see AI integrate naturally into their organisations. Everyone else will keep building impressive pilots that collapse under their own weight.