Synthetic Data’s Practical Uses for AI and Agent Systems — KPMG’s Fabiana Clemente Explains

KPMG distinguished engineer Fabiana Clemente discusses practical applications, limitations and best practices for synthetic data in a Generative AI in the Real World podcast hosted by Ben Lorica. Clemente defines synthetic data as non‑real‑world data used across a spectrum of use cases — from test data management and privacy‑preserving data replicas for offshore teams to improving fraud detection and training/evaluating multi‑agent AI systems. She warns against common mistakes: oversimplifying synthetic data generation and failing to match methodology to use case. Key takeaways include the need for governance, evaluation metrics, and established processes for producing usable synthetic datasets. Clemente notes text now dominates synthetic‑data production (driven by LLMs), but stresses that synthetic data also includes simulation approaches. She highlights synthetic data’s role in agent development — creating structured knowledge and scenario simulations for multistep tool‑using agents — while acknowledging concerns such as feedback loops when models train on AI‑generated data. Overall, she positions synthetic data as an important accelerator for AI development when paired with proper planning, governance and tooling.