Workday builds its Sana home screen by letting Jev pick each component

A chat home screen on a company's own tools has to answer the moment the person stops typing, or the person goes back to clicking links in a menu. The New Stack reported on 14 October that Workday is picking the panel the user sees from its own design system each turn, with a decision model, rather than asking a frontier model to compose one. The decision model is Jev, from TypeSafe.

What the setup does

Workday on Tuesday made Sana, which it bought for $1.1 billion last year, the new home screen for its HR, finance and IT software. In the press demo The New Stack describes, a question about a paycheck brought up a card with the amount hidden behind a click, and a request for time off produced a date picker filled from the form. The design system holds the parts; Jev reads what the user asked and picks the right one.

Joel Hellermark, Workday's chief AI officer and Sana's founder, told the briefing that the usual approach, where a model generates the full interface each turn, is too slow. "If you use a classification model instead and you just give it all of the design components to pick from, you can literally do it instantly," he said. The team gave Jev every atomic component in the design system and hard-coded none of the selection logic. In the demo the paycheck card appeared before Hellermark finished typing the question.

Numbers in the story

TypeSafe says Jev answers in 70 to 500 milliseconds, which Hellermark called the reason this way of running a generative UI works at all. Andrew King, who runs solutions architecture at app shop Very Good Ventures, published a test last month: handing every closed question to Jev cut his generative dashboard's median render time to 3.8 seconds from 29 seconds with a frontier model composing it, on synthetic data. On the Banking77 intent classification set, OpenRouter measured Jev at a median 175 milliseconds and 81 percent accuracy against Claude Opus 5's 2.3 seconds and 84.4 percent, with Claude costing about 22 times as much.

Each figure was measured by the party reporting it, not by Workday on the Sana workload. Workday has not said how large its design system is, how often Jev picks the right component on its own catalogue, or what Sana does when no combination of components fits. TypeSafe's own documentation says Jev leans toward whichever option is listed first, struggles with counting and arithmetic, and does not treat injected content as hostile by default. King's rule for his own setup is that anything Jev answers with under 50 percent confidence is sent to a language model.

Where the method breaks

The approach puts the user interface inside the design system, which means a question with no matching combination of components has no answer and the fallback has to be designed in. Workday has added TypeSafe to its list of subprocessors, the outside firms that may process its customers' data, for intent classification and confidence scoring on all of its product lines. TypeSafe processes data in the United States only and offers zero data retention on request.

A product team weighing this pattern has one question to answer first: whether the design system can actually describe every question the home screen is meant to handle. The approach is faster and the surface is more predictable, in exchange for giving up the shapes nobody drew yet.

Source

Workday is using TypeSafe's Jev to build user interfaces on the fly in The New Stack, reporting Joel Hellermark's press briefing. Jev's architecture and limits in TypeSafe's model card. The Banking77 figures in OpenRouter's Jev versus Claude Opus 5 test.