The art of operationalising AI: perspectives from the field

AI experimentation is accelerating at the enterprise level, but the real challenge begins after the pilot phase. While organisations are actively exploring agentic AI and automation, many are still thinking about how to operationalise these technologies securely and reliably across their environments. Increasingly, the focus is shifting from isolated AI tools toward orchestrating AI agents, automation, systems, and people within governed, end-to-end workflows. 

TechTrade Asia polled Sharath Keshavamurthy (SK), Director of Engineering, South Asia at UiPath, on why pilots stall, where enterprises encounter friction, and what it takes to operationalise AI successfully within complex business environments. Keshavamurthy works closely with enterprises across the region on proofs-of-concept (POCs), automation programmes, and agentic AI initiatives. He was formerly an enterprise automation practitioner, and is now as part of UiPath’s engineering organisation. 

TechTrade Asia: Singapore has one of the highest rates of AI experimentation in the region, yet it seems that many pilots never make it to production. From your experience working with enterprises, what's really behind that gap, and is it fair to call it a technology problem? 

Source: UiPath. Keshavamurthy.

SK: Singapore doesn't have an AI capability problem – it has a path-to-production problem. Across the pilots we have been involved in, the barrier is almost always structural rather than technical. 

Singapore enterprises are genuinely enthusiastic early adopters. They invest early, explore broadly, and are willing to back ideas that have not yet been proven at scale. The challenge arises when exploration becomes a destination in itself. 

A pilot should exist to answer a specific question: can this technology solve this problem, for this organisation, at an acceptable cost? Without that clarity upfront, pilots often demonstrate capability without creating a clear path to production. 

In a market like India, pilots are expected to demonstrate ROI from day one, and that commercial pressure becomes a forcing mechanism that shortens the path from concept to deployment. Singapore's stronger innovation funding removes that pressure. It encourages experimentation, but it also removes the urgency to scale. 

Q: You have seen PoCs up close - the early excitement, the bottlenecks, the moments where things start to unravel. What does a POC journey typically look like, and when does it tend to get complicated? 

SK: Every PoC starts in roughly the same place: a clear customer challenge, a specific use case selected to address it, and a scoped set of success criteria. That is the foundation. 

Where things tend to slow down is the environment setup phase. Customers understandably want to see the technology working on their own applications and data, not on a generic demo. That means gaining access to their live systems: ERPs, CRMs, and other internal tools. 

But most enterprises do not maintain separate test environments for these systems, which means teams spend weeks navigating access approvals, compliance reviews, and infrastructure provisioning before any meaningful testing can begin. What should be a two-to-three-week pilot can stretch into two months before a single line of code runs against real data. By the time the environment is ready, the early momentum has faded. 

The second point of complexity comes once the pilot is running, and stakeholders begin to see results. Requests naturally expand into adjacent processes, edge cases, and new scenarios. While each addition may seem reasonable individually, together they can quickly turn a well-scoped pilot into an open-ended project. Scope creep is one of the most underappreciated risks in the POC journey. 

The pilots that progress most smoothly share one trait: the teams align early on the problem being solved, how success will be measured, and what is explicitly out of scope. That discipline makes the difference. 

Q: There's a lot of buzz in the market right now around agentic AI. When you sit down with a customer, how do you cut through the noise and get to what actually matters for their business? 

SK: Customers today are better informed about agentic AI than ever – and at the same time, more confused. Different vendors use the same terminology to mean different things. 

Two misconceptions come up repeatedly: that agentic AI replaces existing automation investments, and that it operates without human oversight. Neither is true. 

Agentic capabilities extend the automation foundations organisations have already built, and the best deployments are the ones where human judgment is wired into the workflow by design. Enterprises are now recognising is that isolated AI tools on their own rarely create lasting operational value. 

The bigger shift is toward orchestrating AI agents, automation, systems, and people together within governed workflows that can operate reliably at enterprise scale. But the most productive approach is to spend less time on the technology itself and more time understanding the business situation. 

The questions that matter are simple: 

• What problem are you trying to solve? 
• What does the current process actually look like? 
• Where are the bottlenecks? 

Starting from those questions tends to cut through the noise more effectively than any explanation of the technology. It is also important to establish early that agentic automation builds on existing automation investments. Organisations are not starting from scratch. Rather, these capabilities extend and strengthen the foundations already in place. 

Once the discussion is grounded in specific outcomes rather than broad concepts, the pace of the conversation changes considerably. The focus moves to implementation, timelines, and priorities, which is where meaningful progress begins. 

Q: Many organisations still think about automation use case-by-use case. What changes when you start thinking about it at the process level - end-to-end - instead? 

SK: The shift tends to become clear very quickly when customers see it in practice, and it often happens in the first discovery conversation. 

Take procurement as an example. In the earlier phases of automation adoption, organisations would typically deploy separate bots for separate tasks: one for purchase requisitions, another for invoice matching, another for payment posting. Each delivers value individually, but they often operate as disconnected islands with limited visibility across the broader process. Process-level thinking changes that picture entirely. 

The full purchase-to-pay journey — from requisition through approvals, goods receipt, invoice matching, and payment — can be orchestrated within a single governed workflow, with complete visibility at every stage and clear routing for exceptions that require human judgment. 

This is also where purpose-built solutions become increasingly important. Rather than building everything from scratch, organisations are looking for preconfigured, governed workflows that accelerate time-to-value while still integrating into their existing ERP, CRM, and operational environments. 

Q: For regulated industries like financial services, healthcare, and the public sector, agentic AI can feel out of reach because of compliance and data residency concerns. How do you approach those conversations, and is anything changing? 

SK: Regulated industries enter these conversations from a fundamentally different starting point. Most organisations want to understand the value proposition first: what this technology can do for their business. Regulated enterprises typically need to address a prior set of questions: 

• Does this meet our compliance requirements? 
• Where is our data being processed? 
• Who has access to it? 

Those questions have to be answered before any discussion of value can meaningfully proceed. However, this has historically meant that adoption in these sectors moves more gradually than in others. 

But deployment options are catching up to those requirements. The ability to run agentic capabilities on-premises, which gives enterprises direct control over how and where their data is processed, addresses one of the most significant structural barriers these industries have faced.

Once a regulated organisation has cleared its compliance requirements and established its governance framework, the scale of adoption can be substantial – these are highly process-intensive environments, which makes them well suited for agentic automation once the right safeguards are in place. 

*ERP refers to enterprise resource planning solutions, and CRM for customer relationship management tools.

Editor's note: UiPath launched agentic solutions for retail and manufacturing in March 2026, and released agentic AI capabilities on UiPath Automation Suite for the public sector in May. The same month, the company launched UiPath Automation Cloud on the Microsoft Azure cloud platform in South Korea, giving customers in Korea access to domestic data residency without sacrificing enterprise-grade automation capabilities.