ServiceNow's AutoSynthData Automates Training Data Generation for Enterprise AI Agents

ServiceNow CoreAI has unveiled AutoSynthData, a system that automatically converts the capability gaps of enterprise AI agents into high-quality training data. Rather than relying on manual annotation or static datasets, the pipeline observes where a target model fails in a live environment, then uses a stronger teacher model to generate new, feasible tasks that exercise those exact weaknesses.

Curriculum that adapts to the model

Each generated task is defined by a system specification, a user prompt, and a verifier. The specification encodes environment policies, tool constraints, and seeded state; the prompt mirrors realistic user requests; the verifier provides an objective success signal. Tasks must be feasible, realistic, and difficult enough to challenge the current agent — ensuring every new example delivers a meaningful training signal.

As the agent improves, AutoSynthData shifts the curriculum toward remaining failure modes, creating a continuous feedback loop. The approach was demonstrated on EnterpriseOps Gym, a benchmark that simulates enterprise toolchains, databases, and policy layers.

Why it matters for GPU / AI infrastructure: Training enterprise agents at scale demands massive, diverse, and validated datasets. AutoSynthData reduces the human labeling burden while increasing data relevance, which translates into more efficient GPU utilization — fewer wasted epochs on low-signal data and faster convergence on production-ready models.