Workday unveiled Sana, its AI assistant, as the default home screen for its HR, finance, and IT applications on Tuesday. The implementation showcases an unconventional approach to generative UI: instead of having a language model compose interfaces from scratch, TypeSafe's Jev decision model selects from a library of components that Workday has already designed and validated.

During a demonstration, Sana responded to a question about paycheck information by displaying a card with the amount concealed behind a user click, while a request for time off triggered a date picker that Sana completed with additional form fields. Workday acquired Sana for $1.1 billion last year and is now rolling it out to users, with administrators able to deploy it without making it mandatory.

Component selection versus generation

Joel Hellermark, Workday's chief AI officer and Sana's founder, explained that conventional generative UI approaches, where models produce complete interfaces, suffer from performance limitations. Jev operates differently: "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," Hellermark said.

Workday provided Jev with every atomic component from its design system without hard-coding selection logic, though the team invested effort in data preloading to ensure responsiveness. TypeSafe reports that Jev delivers responses in 70 to 500 milliseconds, making it suitable for real-time interaction patterns. Hellermark noted that AI systems operating through actual interface elements frequently outperform pure conversational approaches.

How Jev operates

According to TypeSafe, developers supply Jev with application data and typed questions, and the model returns a selection from up to 255 options, a confidence score, or a binary answer, each with an attached probability. Unlike generative models, Jev does not produce text.

TypeSafe launched Jev on September 15 with a $40 million seed round led by DCVC. On Friday, the company announced a $870 million Series A funding round led by Andreessen Horowitz, valuing the company at $7.5 billion. TypeSafe states that one-third of the Fortune 500 uses Jev, though the company has not disclosed which ones. Andreessen Horowitz's announcement cites 25 percent adoption, and TypeSafe's sole published case study involves Jack & Jill, an AI recruiting marketplace.

The decision model space has grown crowded. OpenAI, Amazon, Cloudflare, Perplexity, and Microsoft have all released competing decision models.

Emerging patterns in developer adoption

Using Jev and similar models for UI construction is gaining traction among developers. CopilotKit, an open source framework for integrating AI copilots into applications, published a recipe in which Jev selects which prebuilt panel to display, with an OpenAI model serving as a fallback. Vercel Labs is testing a Jev mode in json-render, its framework for rendering AI-generated UIs from developer-specified components.

Andrew King, solutions architecture lead at Very Good Ventures, reported last month that routing all closed-ended questions to Jev reduced his generative dashboard's median render time to 3.8 seconds, down from 29 seconds when using a frontier model to compose the interface. His evaluation relied on synthetic data.

"A shape nobody anticipated cannot appear on screen," King wrote. He implemented a hybrid approach where answers with confidence below 50 percent route to a language model instead, which he described as "a cage with a door."

Testing on OpenRouter's Banking77 intent-classification dataset showed Jev responding in a median 175 milliseconds with 81 percent accuracy. Claude Opus 5 achieved 84.4 percent accuracy but required 2.3 seconds and cost approximately 22 times more.

TypeSafe's documentation acknowledges limitations: Jev struggles with counting, arithmetic, and date handling, exhibits bias toward the first listed option, and does not treat injected content as potentially malicious by default. Workday has not disclosed the size of its design system or how Sana handles scenarios where no component combination fits the user's request. King noted in September that he was avoiding Jev for client data because the vendor was "weeks old with no published security certification yet." TypeSafe now provides a SOC 2 Type II report upon request.

Data handling and integration

Workday recently added TypeSafe to its subprocessor list—the external vendors authorized to process customer data. This designation applies across all Workday product lines for natural language processing, intent classification, and confidence scoring, not limited to Sana.

TypeSafe states that Jev operates exclusively on its hosted API, processes data within the United States, and provides zero data retention for enterprise customers.

Hellermark demonstrated Sana functioning within Anthropic's Claude, an AI-to-AI integration. Gerrit Kazmaier, Workday's president of product and technology, noted that this setup results in users losing the Jev-powered interface and some contextual information. Workday's May partnership with Google Cloud encompasses A2UI, Google's specification for agents to describe interfaces that host applications then render. MCP Apps, a Model Context Protocol extension, enables interactive UI within Claude and ChatGPT. Neither appears in Workday's announcements this week, suggesting that employees accessing Sana through Claude currently receive conversational interaction without the date picker and other component-based interface elements.