Proof over Promise Insights on Real World AI Adoption from 2025 MINDS Organizations 2026

Enterprise resource planning (ERP) cloud migration copilot BOX 9 AI agents for code migration BOX 10To streamline complex enterprise resource planning (ERP) cloud migrations, KPMG coupled SAP’s AI copilot with 200,000 pages of proprietary documentation and terabytes of curated help content and best practices from previous migration projects (up to 9 terabytes of content/3 million documents). The resulting solution incorporates prior learning while delivering projects faster, aiming for an 18% faster migration rate with 50% less code reworks. In their AI-powered platform migration solution, EXL fine-tuned LLMs on historical migration life cycle data, then orchestrated a multi-agent platform. This platform uses that institutional memory to reduce rework, auto-convert almost 80% of the code (while improving quality) and shorten project timelines. Data gaps are being closed with various augmentation strategies While extensive datasets have long been considered a prerequisite for impactful AI, MINDS organizations are showing that innovation does not have to stall in data-scarce environments (see Spotlight 6). These organizations are mitigating dependence on massive datasets by combining domain expertise, alternative data streams and advanced modelling techniques to unlock new levels of AI-driven impact. Other MINDS organizations are incorporating physics-based models, among others, to augment limited datasets in design and production processes.SPOTLIGHT 6 University of California, San Francisco and SandboxAQ – physics-guided AI under data scarcity University of California, San Francisco’s (UCSF) Institute for Neurodegenerative Diseases partnered with SandboxAQ to speed up drug discovery for neurodegenerative diseases, like Parkinson’s and Alzheimer’s, by combining AI with quantum chemistry and physics-based simulations. Instead of relying only on large, labelled datasets, the team runs simulations that estimate how well a molecule might bind to a target and uses those results as high-quality guidance signals for the AI ranking models. It represents a smart data augmentation; physics adds trustworthy “labels” where real data are limited, so that AI can explore a much larger chemical space with higher precision. Impact: With this approach, UCSF expanded screening from about 250,000 to 5.6 million compounds, narrowing to around 7,000 for lab testing in weeks. This represents a 36-fold reduction in experimental effort and a 30-fold boost in hit rates versus traditional methods. Democratizing battery cell development BOX 11 Physics-informed AI in data scarce environments BOX 12Tsinghua University, working with Electroder, brought this approach to battery cell development. They integrated multiple data sources with mathematical and computational models representing the physical laws underlying a system. They also incorporated a chat assistant to extract design insights from uploaded scientific papers. Ultimately, these efforts sped up the development process by 3.6 times. CATL and AIMS (Hangzhou Augmented Intelligence Manufacturing Solution) used a data augmentation strategy in industrial operations, where a key goal is to reduce failures. Failures are infrequent enough to begin with, however, that there’s limited data to draw on related to past events. They deployed a hybrid solution combining physics-based models, expert rules and ML to learn from limited, imbalanced and noisy data while adapting to changing conditions. Their approach led to a 50% production speed increase and a 50% drop in quality consistency variation. Proof over Promise: Insights on Real-World AI Adoption from 2025 MINDS Organizations 18