Workday has begun using TypeSafe’s Jev decision model to build user interfaces that change as users type inside its Sana artificial intelligence assistant. Sana has become the new home screen for Workday’s human resources, finance, and information technology software, and system administrators can make it available to users without making it the default interface.
In a demonstration, a question about a pay statement displayed a card that hid the amount behind a click, while a time-off request created a date picker and Sana filled in the remaining fields. The core idea is not for a language model to write an entirely new interface, but for it to select components from a design system that Workday prepared in advance.
Selecting Components Instead of Generating the Interface
Joel Hellermark, Workday’s chief artificial intelligence officer and Sana’s founder, said the usual approach to generative interfaces is relatively slow. According to the explanation provided, Workday supplied Jev with every atomic component in its design system, without writing fixed selection logic for each situation. However, the team also needed to tune data preloading; the pay-statement card appeared in the demonstration before the question had been fully typed.
TypeSafe says Jev responds within a range of 70 to 500 milliseconds, making it suitable for selecting interface elements in real time. Developers send the model application data and a set of typed questions, and it returns a selection from a list of up to 255 options, a score, or a yes-or-no answer, with a probability for each result. Jev does not generate text.
What Changes in Practice?
This design shifts part of the interface task from open-ended generation to controlled classification. As a result, the application can display known, testable elements instead of waiting for a language model to compose a new structure each time. Tests published outside Workday point to a potential difference in speed: Andrew King of Very Good Ventures said that using Jev for questions with defined answers reduced the median rendering time of a generative dashboard to 3.8 seconds, compared with 29 seconds when using an advanced model to compose it, with the tests relying on synthetic data.
In a test on the Banking77 dataset, Jev achieved a median time of 175 milliseconds and 81% accuracy. Claude Opus 5 recorded 84.4% accuracy, but required 2.3 seconds and cost about 22 times more, according to the source.
Limitations and Open Questions
TypeSafe’s documentation says Jev struggles with counting, calculations, and dates, and may favor the option listed first. It also does not treat injected content as adversarial by default. Some experiments use an alternative language model when Jev’s confidence falls below 50%, limiting the scope of its decisions rather than allowing it to generate any possible form.
Workday did not clarify the size of the design system used with Sana, or what happens when a request does not fit any combination of the available components. Workday also added TypeSafe to its list of sub-processors capable of processing customer data. TypeSafe currently provides a SOC 2 Type II report upon request, but the source does not establish additional details about its security architecture or data-retention policies.
The move comes amid broader competition in decision models; OpenAI, Amazon, Cloudflare, and Perplexity have launched similar models, followed by Microsoft the previous week. TypeSafe launched the Jev model on September 15 with a $40 million seed round led by DCVC, then announced an $870 million Series A round led by Andreessen Horowitz at a $7.5 billion valuation.