In asset-intensive sectors, the reality of digital transformation is about uptime, safety, compliance, and costs, not just exciting pilots. In a free-wheeling conversation with CIONOW, Sapann Talwar, CDIO, Essar Power, shares what is really working in the field: the one issue resolved only when IT and OT came together as a team, an intelligent roadmap for rolling out AI, analytics and automation, and metrics which show predictive maintenance impacting P&L.
Large conglomerates face a massive balancing act as they must ensure uptime, safety, regulatory compliance and cost pressure. Where can digital technology create the most measurable business value in this mix?
The strategy must address four key areas: uptime, safety, cost and the alignment of physical operations with business decisions. Predictive intelligence is critical—moving from calendar-based maintenance to condition-based maintenance can reduce unplanned outages by more than 20% and extend asset life.
IT and OT must not work in silos; systems like SCADA, DCS and sensors on the OT side need tight integration with IT systems like ERPs to create better metrics. AI can further optimize forecasting and efficiency across both domains. The real value comes from closing the loop between operational events and business outcomes—that’s what the board cares about.
You have emphasized that IT and OT need to break silos. Can you share one operational problem that became easier to solve once data, IT and OT teams began working together, and what made that collaboration stick?
The biggest challenge is that IT and OT see their infrastructure through very different lenses. The breakthrough came when we showed them the common data layer: both have assets, event codes, logs and failure modes that need prediction.
We formed a cross-functional committee, both with collective responsibility and incentives in place. So either we win as a team, or we lose as a team, rather than looking at who to blame. That shift brought the problem-solving time from weeks down to days. The result in bottom-line benefits to management was obvious. They had better problem-solving, more reliable solutions, and regained confidence, because we were able to fill in the blanks, in order to avoid future recurrence.
The common saying is that data is the new oil; yet numerous enterprises find it challenging to extract decisions from this oil. What’s your methodology to create a reliable and accessible data foundation?
Everyone knows data is key, but the focus must be on knitting every data element together into a unified fabric—from OT sensors to IT logs. The challenge is correlating information across different toolsets and validating data quality at the edge, because systems generate a lot of garbage.
We need to validate real data, timestamps and sensor vs gateway levels, then set aside the noise. Tools matter, but humans matter more: engineers must understand the data foundation, correlate insights and enable self-service. It has to be an open-minded way of thinking but with guardrails put in place so the streamed data makes sense to humans not just to dashboards.
With the buzz about AI, how do you figure out when the use case warrants AI, prediction, automation or a better process? How does your team go about doing it?
Begin by establishing the best practice – the majority of the issues result from fragmented processes and manually assigned tasks between functions.
Automation becomes useful for repetitive, rules based functions like a check on a compliance report and triggering alerts and escalations.
Application prediction models use time-series sensor data that must be predictable or optimized, and then in cases of non-structured data (maintenance logs, inspection reports or vendor documents) consider generative and agentic AI where natural language understanding is of great benefit.
If a problem can’t be solved by process, automation, analytics or AI, you may need the next wave of technology but that hierarchy ensures you are not over-investing in AI where simpler fixes work.
It is common to have an Industry 4.0 scenario around predictive maintenance. In addition to providing you with information to prevent equipment breakdowns, how would you indicate how this effort has produced results in terms of increased availability, efficiency or lower cost?
Focus first on model accuracy aligned with systems, not just business outcomes. Key metrics include:
Availability: Reduce your unplanned downtime and increase mean time between failures (MTBF) by quantifying pre vs post-predictive maintenance for asset class.
Efficiency: Link predictive alert to specific performance – heat rate, capacity factor, and reduced startup time to evaluate performance of boilers, turbines etc.
Safety: Identify trends in safety events, audit findings and regulatory issues.
Cost: Optimize inventory of spare parts. Minimize emergency maintenance and avoid oversupply.
We run control pilots on comparable assets—isolating technology impacts from variables like fuel quality or loading profiles—to set benchmarks and scale what works.
As generative AI percolates into enterprises, where should the line be drawn between human-led and machine-led decisions, especially for plant engineers, maintenance teams and business leaders?
I view AI as augmented intelligence, not artificial intelligence. The highest value comes from augmenting human expertise with AI systems. Neither humans nor AI alone can deliver the best performance.
Plant engineers will also use generative AI for asking questions about past production runs in natural language and developing explanations much more rapidly than human experts alone could provide. Nevertheless, the final decision, specifically regarding safety-critical or high-stakes choices, will continue to rest in the hands of humans. AI facilitates more timely and data-driven decisions; humans have accountability and understanding.
What’s one piece of advice you would give to CIOs in asset-heavy industries who are trying to sequence their digital investments for maximum ROI?
Don’t pursue AI because it’s AI. Identify fixes for processes, automate them, then, establish predictive analytics and then only AI where absolutely necessary. Build a unified data layer with IT–OT convergence, validate data quality at the edge, and ensure humans are embedded in the loop.
Measure success by uptime, safety, cost, and not about tech deployment alone. Also, put some joint responsibility between IT/OT, so IT and OT do what they do best in partnership, as a matter of course, not an incident.
