ServiceNow Introduces AutoSynthData for Enterprise Agent Training Data

ServiceNow AI published a new post on the Hugging Face blog titled "AutoSynthData: Generating Training Data for Enterprise Agents," putting a method for automated data generation in front of the wider AI community. The post frames the problem plainly: enterprises that want capable agents need training data that reflects their own workflows, tools, and policies, and that data is hard to come by.

The central claim behind AutoSynthData is that training data for enterprise agents can be generated rather than painstakingly collected. Instead of relying only on hand-labeled examples or scattered logs, the approach points toward synthesizing the examples an agent needs in order to learn how to act inside a business environment.

The timing reflects a broader shift in the agent market. In the past year, the conversation around enterprise AI has moved from chat-style assistants to agents that call tools, complete multi-step tasks, and take real actions. Each of those capabilities raises the bar for training and evaluation data, because an agent has to be correct not just in what it says but in what it does.

Enterprise data is also unusually sensitive. Customer records, internal tickets, and proprietary processes are exactly the material agents need to learn from, yet they are also the material companies are least willing to expose. Synthetic generation is attractive precisely because it can produce realistic training signal without shipping raw internal data to a third party.

Publishing on the Hugging Face blog rather than only through corporate channels is itself a signal. It places the work in the same ecosystem where models, datasets, and agent frameworks circulate, and it invites practitioners to inspect the approach and build on it.

What to watch next is whether AutoSynthData ships as a concrete library or dataset, how it performs on task-specific accuracy, and whether other enterprise software vendors follow with similar synthetic-data pipelines for agents.