This podcast is drawn from a conversation between Dinkar Charak, who heads solution engineering at The Modern Data Company, and Niranjan Nilekani, VP - Head BIU & Analytics at the Mahindra Group.
About the Guest
Niranjan Nilekani is Vice President and Head of BIU & Analytics at Mahindra Group, with over 15 years of experience across business intelligence, data science, and analytics.
He has previously held analytics leadership roles at ICICI Lombard, Accenture, TCS, and Deloitte, working across banking, risk, and data science. His expertise spans data-driven decision-making, AI, and analytics-led business transformation.
This discussion covers where agentic AI is working in regulated industries, why adoption remains limited, and how enterprises should think about measuring ROI.
TOC
* Is agentic AI solving a real business need, or is it mostly hype?
* How did the industry arrive at agentic AI?
* Where is agentic AI already delivering measurable ROI?
* Why hasn’t agentic AI taken over core functions like underwriting and fraud detection?
* What role does regulation play in slowing adoption?
* How expensive is agentic AI in practice, and how are enterprises weighing that cost?
* How should enterprises measure ROI on agentic AI projects?
* What rollout strategy do the speakers recommend?
* Where should enterprises start when piloting agentic AI?
* How big a barrier are hallucinations to enterprise adoption?
* What comes after agentic AI matures?
* What’s the overall takeaway for BFSI leaders considering agentic AI?
Is agentic AI solving a real business need, or is it mostly hype?
It is a mix of both, with genuine potential still outpacing actual maturity. Niranjan points out that the underlying concepts behind agentic AI have existed for roughly a decade in other forms, and what is new is mainly the branding and the current wave of hype. He estimates that only 10 to 15% of organisations attempting agentic AI implementations are seeing it succeed in production, with most still struggling to justify the return on investment against the cost of running these systems.
How did the industry arrive at agentic AI?
Agentic AI is the latest step in a longer analytics evolution rather than a clean break from what came before. Niranjan traces the path from Excel-based analysis in the early 2010s, to dashboarding tools like Tableau and Power BI, to statistical and machine learning tools such as SAS and Python, and eventually to robotic process automation. In his view, agentic AI is functionally close to RPA with sub-agents and a supervising “master agent” layered on top, rather than a fundamentally new category.
Where is agentic AI already delivering measurable ROI?
The clearest wins are in redundant, customer-facing workflows rather than core decision-making processes. Niranjan highlights WhatsApp-based customer journeys and contact centre automation as areas where agentic AI is already creating smooth, successful outcomes for BFSI organisations. These use cases succeed because the workflows are repetitive and sit outside the industry’s most tightly regulated processes.
Why hasn’t agentic AI taken over core functions like underwriting and fraud detection?
Adoption stays low in core BFSI processes because regulation, cost and reliability concerns all compound at once. Niranjan cites his own experience that only around 30% of underwriting decisions are currently handled by AI, with the remaining 70% still requiring human judgment. Fraud detection follows a similar pattern: agentic systems can flag high-risk transactions, but pinpointing fraud with confidence still relies on human review layered on top of rule engines.
What role does regulation play in slowing adoption?
Regulatory restrictions directly limit which AI tools BFSI institutions are allowed to use, which in turn limits performance. Niranjan notes that RBI guidelines have already restricted the use of certain large language models over concerns about data processing happening outside India, pushing institutions toward lower-quality LLMs with correspondingly weaker output. He also points to RBI’s stated position against black-box models and its discussion of a “kill switch” requirement, both of which conflict with how opaque agentic systems currently behave.
How expensive is agentic AI in practice, and how are enterprises weighing that cost?
The cost of running agentic AI at scale is proving high enough that leadership is now comparing it directly against the cost of human analysts. Niranjan describes CEOs asking employees to report token consumption for individual proof-of-concept projects, and notes that while a single token costs roughly $0.01, a simple POC can burn through millions of tokens. In many cases, his organisation has concluded that paying a data scientist is currently more cost-effective than paying for the infrastructure and token usage an agentic solution requires.
How should enterprises measure ROI on agentic AI projects?
ROI should be judged by the same top-line and bottom-line standards used for any prior analytics technology, not by a new or looser bar. Niranjan argues that agentic AI is an enhancement on the same continuum as logistic regression, deep neural networks and LSTMs, so the benchmark for success has not fundamentally changed: does the technology grow revenue, or does it reduce cost through mechanisms like fraud control or lower customer acquisition cost. Dinkar adds that ROI conversations increasingly include FinOps-style cost transparency, since running a POC and running a system in continuous production carry very different cost profiles.
What rollout strategy do the speakers recommend?
Large agentic AI initiatives should be broken into small, independently measurable phases rather than attempted as one large rollout. Niranjan uses the example of automating an entire BI workflow, covering BRD generation, dashboard automation and insight generation, and explains that trying to implement all of it at once makes ROI nearly impossible to isolate. His recommended approach is to split a large initiative into four or five smaller projects, measure the top-line or bottom-line impact of each individually, and aggregate the results before presenting them to leadership.
Where should enterprises start when piloting agentic AI?
Non-regulated, peripheral workflows are the safest and fastest path to an early win. Both speakers agree that starting with the edges of a process, rather than its regulated core, produces faster success and builds the internal case for expanding agentic AI further. Niranjan adds that teams should also avoid use cases with heavy token consumption early on, since high inference costs can undermine a pilot’s business case before it has a chance to prove value.
How big a barrier are hallucinations to enterprise adoption?
Hallucinations remain the first objection raised by leadership and a major reason full production deployment stays limited in customer-facing BFSI use cases. Niranjan explains that after significant investment in a solution, a single hallucinated output can be enough to block productionization, since reputational risk is high in an industry that interacts directly with customers. As a result, many institutions continue to rely on traditional rule engines for critical applications, using them to surface risk flags even when they cannot precisely pinpoint fraud.
What comes after agentic AI matures?
The two speakers expect agentic AI to eventually become self-governing and as unremarkable as GPS or the broader Internet of Things is today. Dinkar predicts that agents will increasingly come with built-in governance shaped by their specific purpose, behaving differently in a regulated context than in a low-stakes one, similar to a well-trained employee who has completed compliance and domain training. He also anticipates a shift toward “agents for things,” where AI is embedded in physical, human-facing systems in the way IoT sensors are embedded in everyday devices today, largely invisible once adopted at scale.
What’s the overall takeaway for BFSI leaders considering agentic AI?
The core message is that old business discipline still applies to a new technology, and rushing implementation without validation is the biggest risk. Niranjan closes by stressing that agentic AI remains genuinely new, still functions as something of a black box, and should be validated thoroughly rather than adopted for its own sake. His expectation is that meaningful, widespread adoption is still three to four years away, arriving once regulation, governance and cost structures mature enough to support it at scale.
📝 Note from EditorThe above insights are summarised versions of the actual dialogue. Feel free to refer to the transcript or play the audio/video to capture the true essence and details of his as-is insights. There’s also a lot more information and hidden bytes in the interview, listen in!
Thanks for reading Modern Data 101! Subscribe for free to receive new posts and support our work.
Guest Connect 🤝🏻
Connect with the guest on LinkedIn 💬
Host Connect 🎙️
Connect with the host on LinkedIn 🎙️
This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit moderndata101.substack.com
