Four AI Products, One Agent Platform

Only designer on the suite. I lead design across conversational AI, autonomous agents, and self-service configuration. I owned the Assemble rebuild, Employee AI from scratch, and carrying that chat model into Chat AI. I partner with product managers and a globally distributed engineering team from discovery through implementation. Engineering productionizes what survives review.

Constraint. Assemble is the configuration and knowledge platform that powers Voice AI, Chat AI, and Employee AI. I joined an incomplete system and a legacy UI: components did not scale, layouts were not trustworthy, workflows were brittle.

Alternatives. Document the failures and keep shipping on the old look, or spend the opening weeks on a real system: components, layouts, and the UX of how banks configure knowledge and behavior.

Decision. I chose to establish the system before continuing to layer product work on top of it. I revamped the UI, the library, and the workflows so later products could look and behave like one platform. Short-term velocity was traded for scalable leverage. That platform system became the base for Employee AI. The rebuild is the product design system: Figma components and tokens on Ant Design foundations, in GitHub. Accessibility was included in that system. WCAG 2.2 AA is embedded in the components, tokens, and foundations, and that same bar carried through the refresh across Assemble, Employee AI, and Chat AI.

Constraint. Employee AI is not one user. Contact-center agents, branch tellers, and knowledge-only staff all need the bank's AI. Some banks already have a contact center and will not switch desktops. One agent-queue tool would fail the rest.

Decision. The four depths are how the product balances automation, human oversight, and control. One product model, different levels of depth. Full workspace where Interface owns the employee surface. A thinner knowledge mode. A summary embed when the bank already owns escalations. Same Assemble chrome and chat library — not four products. Designed from scratch, not a reskin of the old tool.

Constraint. Chat AI is the embed on a bank's website: members and prospects asking questions grounded in the same knowledge base that powers everything else. It also had to support applications and transactions — transfers, balance, block card, replace card — without becoming a second product language.

Decision. I deliberately separated platform chrome from conversational behavior so the latter could travel across products without creating a third interaction language. A platform library for app chrome (Assemble, Employee). A chat-behavior library for knowledge and transactions. Employee AI already handled most of those chat interactions, so Chat AI reuses that experience. One UI that is easier to manage, and a consistent conversation for the member whether a human is in the loop or not.

Assemble's configuration workflows are the self-service surface for the agent estate. A copilot is now in progress to help users configure that estate, and it will use the same chat library.

Constraint. Exploration was cheap. Shipping was not. Engineers could not productionize a Figma file they had to reinterpret, and generated screens would drift from the real library the moment they left review.

Decision. I turned the library into markdown chapters — framework, tokens, components, variants — so a design artifact could become an implementation input. The rule is reuse before invent, and document anything new so the guide grows. Human review is the ship gate: generated exploration is disposable until it survives product, technical, and design review.

I prototype on the real system — Claude, Cursor, GitHub, and live APIs — so product and engineering can check function, accessibility, and consistency before a screen is productionized. What we take into that prototype comes from customer feedback, competitive research, and Microsoft Clarity, synthesized with Claude and decided with product managers and engineering.

This is still current work, and still confidential. I set the systems, redesigned the products that sit on them, and kept the interaction model consistent as Assemble grows a copilot. The product decisions are above.