For many organisations, the first wave of enterprise AI focused on proving what the technology could do.
But as AI moves from experimentation into core business processes, a more difficult question is emerging: when does an AI initiative become a genuine operational transformation rather than simply another productivity tool?
For Prabu Sangar, director of Technology Operations and Integration, Asia,at FCM Travel, the answer lies not in how impressive an AI system is, but in whether it fundamentally changes how work gets done.
Short-term productivity vs long-term impact
If you remove the AI and the business immediately goes back to the old way of working, you probably deployed a tool. You did not transform the operation. Prabu Sangar
How will organisations differentiate between AI initiatives that create a short-term productivity boost from those that deliver lasting operational improvement?
For Sangar, the key is whether AI transforms how the work gets done.
“The test, in my view, is simple. If you remove the AI and the business immediately goes back to the old way of working, you probably deployed a tool. You did not transform the operation,” he said.
For him, AI that drives long-term operational improvement goes beyond productivity or time-saving; it identifies issues, understands context, and acts within guardrails with human judgement when needed.
For him, this is where agentic AI comes in. It helps “move towards systems that do not just generate an answer, but can understand context, make decisions and take action across workflows.”
He added: “In an enterprise environment, AI needs to understand what is happening around the task, not just the prompt in front of it.”
Measuring success
For Sangar, transformation therefore cannot be measured simply by counting AI deployments or calculating how much time employees save. The more meaningful measure is whether the underlying operation has become faster, more accurate, more responsive and better able to learn.
“I think one of the problems today is that we are trying too hard to reduce AI success to one ROI number. In my opinion, the better question is: what actually changed in the operation?” he said.
Common success indicators include cycle time, error rates, rework, customer effort, decision speed, service quality and cost-to-serve.
He said it is also vital to watch exception rates, which he considers one of the strongest indicators of whether an AI-enabled process is becoming mature.
“How often does the AI get stuck? How often does a person need to step in? Why did they step in? And is the system learning from those interventions?”
Human capacity is also a factor. Saving 500 hours is less important than what happened during those hours. Time savings should lead to more time spent with customers, solving complex problems, and improving decision quality and trust.
“Speed without trust is not an improvement,” he emphasised. “Accuracy, compliance, explainability and human oversight all need to be part of the measurement. Ultimately, I want to know whether AI is making the organisation better at deciding, adapting and learning.”
Scaling AI initiatives in APAC
Yet proving value in one workflow is only the beginning. The challenge becomes considerably more complex when organisations attempt to extend successful AI initiatives across multiple markets.
“I think this is especially important in APAC because there is a tendency to talk about the region as if it is one market. It is not,” Sangar said.
With different regulations, languages, levels of digital maturity, customer expectations, infrastructure and ways of working, scaling AI initiatives in APAC for Sangar does not mean expecting every market to work the same way.
“I prefer the idea of a common core with local intelligence. Architecture, security, governance and core standards can be consistent, but markets need room to adapt how AI is applied locally,” he emphasised.
He believes localisation should not be an afterthought, but carefully designed earlier in the process, not after solutions are already built.
“If one market has mature digital workflows and another still relies heavily on manual processes, deploying the same AI in both places will not give you the same outcome. It will simply give you different problems,” Sangar explained.
He added that future operating models will use a more federated approach with strong guardrails, but with the flexibility to adapt to market needs and trends.
“The job of leadership is not to remove complexity – it is to design for complexity without allowing it to become chaos,” he said.
Correcting misconceptions
The complexity of scaling also exposes a misconception about enterprise AI: that greater adoption automatically translates into greater value.
For Sangar, the most common wrong belief in scaling AI is that it means buying more AI.
“More licences, more copilots, more models and more use cases do not automatically create more value. The difficult part is scaling everything around the technology, such as data quality, trust, governance, decision rights, integration and organisational readiness,” he said.
He also notices how many enterprises focus on autonomy without proper insight.
He emphasised: “The question should not be, ‘How quickly can I remove the human?'” It should be, “Where does the human genuinely add value?”
He believes that organisations must intently assess which processes need to be autonomous. Some require empathy, judgement, negotiation, or accountability; others do not.
“We are moving quickly from copilots to agents, and eventually towards multiple agents working together across enterprise systems. But that makes governance even more important. My view is very simple: Do not scale autonomy faster than you can scale accountability,” he said.
Gartner offers the same perspective, emphasising the importance of continuous oversight and robust technical guardrails to keep humans in the loop so autonomy does not come before governance.
Deploying lasting operational improvements from their AI initiatives
Automating a bad process just gives you a faster bad process. Prabu Sangar
When asked for advice for CIOs wanting to deploy long-term operational improvements from AI initiatives, he simply said: “Do not start with the model; start with the problem.”
He said organisations should keep a keen eye on what causes friction, bottlenecks, rework, repeated decisions, or unnecessary effort for customers.
After that, they can assess where AI can transform processes and workflows.
“I also think CIOs need to resist the temptation to layer AI onto an existing process and call that transformation. Automating a bad process just gives you a faster bad process,” he said.
He also emphasised the importance of governance and preparing for a much more agentic environment by considering identity, access, permissions, auditability and accountability for AI agents. He anticipates AI increasingly observing, anticipating, deciding and acting, rather than just waiting for prompts.
“Finally, I think the organisations that succeed will not necessarily be the ones with the best model. They will be the ones with the best learning loop. Measure what happened. Understand where people intervened. Learn from the exceptions. Improve the system. That is ultimately how I think about AI: automating the ordinary, while humanising the extraordinary,” he concluded.
The organisations that succeed will not necessarily be the ones with the best model. They will be the ones with the best learning loop. Prabu Sangar
