A new class of small “decision models” is being tested inside application code for choices, scores and probabilities that do not need generated prose. At the AI Engineer Code Summit, teams described early uses of Jev, a non-generative model released by TypeSafe AI less than two months ago.
Glean said Jev powers a significant number of internal flows alongside language models and conventional classifiers. Its attraction is zero-shot flexibility: the set of choices can change without fine-tuning a new classifier. Glean did not provide independent accuracy or production-volume figures, so the claim mainly establishes adoption rather than comparative performance.
UiPath demonstrated a tournament-style retrieval method in which Jev repeatedly selected relevant document chunks from small groups. The company said thousands of calls completed in under a second. Workday showed a prototype that begins filling a time-off form while a person is still typing, using a fast model to predict an intended action rather than only the next words.
These examples suggest a narrower role for models in software: frequent bounded decisions can go to a low-latency system, while a larger reasoning model handles open-ended work. The summit presentations were demonstrations and company reports, not a common benchmark. Reliability, calibration and the cost of incorrect automated actions still need to be measured in each workflow.