Published thought leadership
Books, articles, and media in industry outlets.
AI-Powered Data Transformation: Modern Solutions for Insurance and Beyond
Co-authored with Pankaj Pant, Anshul Kaushik, and Satya Prakash. A practical guide to data governance, integration, and AI readiness for regulated industries, written from real-world experience on live insurance and financial services programs. The goal: help readers move from project execution to confident, defensible decision-making that creates lasting value.
Large Language Models in Finance: Our Paper at ISBM 2026
Our paper, "Large Language Models in Finance: A Survey of Foundations, Applications, Adoption Strategies, and Open Challenges," was presented at the 5th World Conference on Information Systems for Business Management (ISBM 2026) in Bangkok. It has been selected for publication in the ISBM Conference Proceedings by Springer Nature. It was a privilege to work with my co-authors, Aparna Krishna Bhat and Pankaj Pant. Why we wrote it Financial institutions are no longer asking whether large language models are useful. They are asking how to actually get one. To answer that, a whole family of financial LLMs (FinLLMs) has emerged, and they are built in very different ways. Some are foundation models trained from scratch on proprietary data. Others take an open-source model and fine-tune it at a fraction of the cost. We didn't try to catalog every published system. We organized the field around the decision an institution actually faces: not which model is best in the abstract, but which construction approach fits its own constraints. What the paper covers Foundations: the adaptation techniques behind the differences between models. These range from full fine-tuning, prompt-based methods and instruction tuning to RLHF, parameter-efficient methods like LoRA, and retrieval-augmented generation (RAG). Applications: the finance work FinLLMs are inheriting, not inventing: trading, risk modeling, text mining and advisory. These set the performance baselines any model has to beat. Systems compared: BloombergGPT, FinMA, InvestLM, FinGPT and FinTral. We compare them by how they were built, because that choice drives cost, openness and fit for real deployments. A decision framework: practical guidance for institutions deciding whether to build, fine-tune, or buy. What the evidence says The benchmark picture is mixed. On sentiment analysis and classification, FinLLMs now hold their own against much larger general-purpose models. On multi-step numerical reasoning, they still fall well short, and none of the adaptation strategies we reviewed has closed that gap. The biggest open obstacles are hallucination, knowledge that stops at the training cutoff, and the lack of a shared evaluation framework. So we argue that retrieval grounding, contamination-checked benchmarks, and auditable fine-tuning should come before simply scaling models up. Why it matters to me My day job is helping insurers and financial services firms modernize their core platforms. There, the questions in this paper are concrete: what to build, what to buy, how to govern it, and how to trust the output. This research was another chance to connect academic work with what we see in the industry every day, and to keep learning as AI keeps changing. I'm grateful to my co-authors and to everyone who supported this work. I'll share a link here once the Springer Nature proceedings are published.
AI Driven Predictive Modeling Techniques for Improving Risk Assessment and Policy Management in Modern Insurance Systems
Insurance systems can now be more precise, scalable, and flexible thanks to rapid AI development. This work introduces AI-driven predictive modeling. It improves policy management and risk assessment via gradient boosting machines, deep neural networks, and ensemble learning. Advanced data preprocessing, feature selection, and model optimization were used to process health, vehicle, and property insurance data. Using new data streams to update risk ratings makes policy ideas and real-time premium modifications more detailed. SHAP-based interpretability tools promote open decision-making, which aids rule-following. Experiments used 1.2 million anonymized insurance records. Models were 15.4% more accurate at predictions, 17.8% less likely to give false positives, and 21% faster at policy adjustments than regular statistical models. Customer satisfaction increased 13%, indicating that individuals trust AI-assisted decisions. Co-Author: Pankaj Pant; Anshul Kaushik; Satya Prakash; Tarun Goma
AI's Role in Transforming the TPA Landscape in Life Insurance and Annuities
Authored by Souma Basu. How AI is moving third-party administrators in life and annuities from legacy back-office processing to strategic partnership: intelligent intake, conversational servicing, transparent claims, governance, and the rise of agentic TPAs. Also available as a full article on this site.
Efficient Insurance Operations: How AI Is Helping to Drive Down Costs
Co-authored with Mohammad Ibrahim, Sean Butler, and Matthew Norden as part of Capco's US GenAI series. Examines how AI reduces insurance operating costs across three key use cases: claims processing, fraud detection and prevention, and risk assessment and underwriting, with real-world examples from carriers like Zurich and GEICO.
The Flight of Insurance from California
Authored as Principal in Capco's US Insurance Practice. Examines why major carriers like State Farm and Allstate exited California's homeowners market, covering wildfire risk, Proposition 103 rate regulation, the strain on the FAIR Plan, and parallels in Florida, Louisiana, and Colorado.
Embracing Innovation: The Future of Property and Casualty Insurance
Authored by Souma Basu. With global premiums projected to reach $10 trillion by 2030, this piece maps four forces reshaping P&C insurance: IoT-driven risk prevention, data-driven operational optimization, sustainability and ESG, and embedded insurance at the point of sale.