Workday lets TypeSafe's Jev choose which design-system components Sana puts on screen

Because Jev only selects, we'd expect each Sana screen to reduce to a series of closed questions, such as which card to show. Every possible answer maps to a component someone already built and reviewed. The list limit matters here. Jev picks from at most 255 options ^[20], so a larger design system would need more than one question to cover it.

Joel Hellermark, Workday's chief AI officer and Sana's founder, said the usual approach, in which a model generates the full interface, is "quite slow" ^[3]. "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 ^[4]. He added that Workday hard-coded none of the selection logic ^[6].

Outside numbers are thin, and TypeSafe's own figures call for caution. It says a third of the Fortune 500 uses Jev, while Andreessen Horowitz's announcement puts the figure at 25 percent ^[24]. Andrew King, who leads solutions architecture at Very Good Ventures, wrote that handing every closed question in his generative dashboard to Jev cut the median render time to 3.8 seconds from 29 seconds with a frontier model ^[10]. That is about 7.6 times faster ^[21]. His tests ran on synthetic data ^[10].

For that ratio to hold at Workday, its screens have to break into closed questions and few requests can fall under the confidence line. In King's setup, anything Jev answers with less than 50 percent confidence goes to a language model, an arrangement he calls "a cage with a door" ^[12]. Those requests get the language model's latency. On the upside, King wrote, "A shape nobody anticipated cannot appear on screen" ^[11].

We think that is the right trade in a closed design system where every component was built and reviewed. The worst output is a wrong card from a reviewed set.

Accuracy gives up a little. In OpenRouter's Banking77 intent-classification test, Jev's median answer came back in 175 milliseconds and was right 81 percent of the time ^[13]. Claude Opus 5 got 84.4 percent right, but it needed 2.3 seconds and cost roughly 22 times more ^[13]. That is 3.4 points of accuracy for about 13 times the latency [d2, d3]. Banking intents are not interface components, so Workday's own numbers will differ.

TypeSafe's documentation says Jev struggles with counting, arithmetic and dates, and leans toward whichever option is listed first ^[14]. The second item is a configuration input. The order of the component list can change the answer.