MIT researchers develop methods to ground AI models in organizational reality

MIT researchers led by professor Devavrat Shah have developed new approaches to ground large language models and AI forecasting systems in the specific operational data of organizations, addressing a key limitation of current enterprise AI tools. The work, conducted through the MIT-IBM Watson AI Lab and commercialized via Ikigai Labs, targets tabular and time-series data that makes up the bulk of business records but remains poorly handled by general-purpose language models. Standard LLMs excel at natural language but lack the structured, organization-specific context, financial records, supply chain logs, sensor data, needed for accurate forecasting, planning, and decision support. Shah's team created methods that combine LLMs with probabilistic graphical models and large graphical models, enabling systems to learn the statistical structure of an organization's data and generate reliable predictions with uncertainty quantification. The approach has been deployed at companies including a major consumer packaged goods firm, where it reduced forecast error by 20 percent compared to legacy methods, and a global logistics provider that used it to optimize routing across thousands of vehicles. Ikigai Labs, co-founded by Shah and MIT alumnus Vinayak Ramesh, offers the technology as a platform for demand forecasting, scenario planning, and anomaly detection across industries from retail to financial services.