Two announcements from October 2 highlight the work required to make AI useful inside an organization: better training examples for models and practical deployment experience for engineers. Both remain relevant as teams decide what to build and how to test it.
ServiceNow CoreAI described AutoSynthData, a pipeline that identifies a target model’s failures, uses a stronger teacher to establish successful behavior and generates new tasks around those gaps. Each example includes an environment, a request and a verifier, with checks that valid solutions succeed and incorrect outcomes fail. In its EnterpriseOps Gym Hybrid experiment, ServiceNow reports a 7.2-percentage-point improvement in mean Pass@1 after supervised fine-tuning on 2,000 generated samples. This is a controlled, author-reported result; the announcement does not establish a generally available training service. Builders can borrow the central idea: turn observed workflow failures into executable test cases before scaling synthetic data generation. Learners get a concrete example of why data quality depends on the quality of the success check.
ServiceNow research, October 2 · Announced 2 October 2026