AutoSynthData makes the verifier part of the training example
ServiceNow's pipeline turns capability gaps into executable tasks, checking both successful trajectories and incorrect outcomes.
A failed workflow loses important context when it becomes a request for similar-looking training prompts. ServiceNow CoreAI's AutoSynthData keeps the execution environment attached: a task includes system constraints, a user request and a verifier for the outcome.
The pipeline examines where a target model fails and a stronger teacher succeeds. It converts those gaps into capability descriptions. The generator receives those descriptions rather than the original evaluation prompts, entities or trajectories, then builds new tasks around the capability.
Acceptance is the architectural detail worth examining. A reference solution must execute successfully. The verifier must also reject relevant incorrect outcomes. Repaired candidates undergo checks again, while expansion stays anchored to accepted core tasks rather than repeatedly deriving variants from variants.
I would start with a clearly bounded workflow failure and an observable successful state. Generating additional requests before defining that state leaves the central training question unresolved: what behavior should receive credit?
The operational choice remains who owns the success criteria and which alternative solutions the verifier must accept.
Source: Esakkivel Esakkiraja and colleagues, ServiceNow CoreAI, AutoSynthData: Generating Training Data for Enterprise Agents, published October 2, 2026.