AI-Native Workspace for Bloomberg Law, Tax & Government
Bloomberg's AI Assistant already existed — but users treated it as a chatbot. This project rebuilt it from the ground up as a unified research platform for legal, tax, and government professionals.
Users had AI. They just didn't know what to do with it.
Bloomberg's AI Assistant existed in each product — a narrow panel docked to the right side of the legacy research interface. It was visible, but it felt separate from everything users needed to do. They used it to ask questions. Then they went back to their actual work somewhere else.
The problem wasn't that AI was hidden. It was that users saw a chatbot, not a platform — and the chat lived in a different world from their workflow.
AI Assistant 1.0 appeared as a skinny right-side panel. Users perceived it as a chat tool, not something capable of driving their whole research workflow.
Getting an answer in chat meant returning to the research interface to act on it. The conversation and the work were always disconnected.
Law, Tax, and Government each had different AI entry points and interaction patterns — limiting predictability and cross-product trust.
No way to combine Bloomberg's proprietary content with user-uploaded documents in a shared AI context for deeper analysis.
Core infrastructure, not a BU feature.
This was foundational platform work — shared components that Law, Tax, and Government would each build their domain experiences on top of. Central design defined the core UX; downstream teams extended it. Requirements flowed into the center. The design did not fork per business unit.
The initiative carried hard commercial stakes: core components were due by end of June for team hand-offs, with Law showcasing the integration at industry conferences in July and August.
- Mar 15 Backend support for user-uploaded file AI workflows — prerequisite for Tax showcases and the ML pipeline milestone.
- Jun 30 Core AI + Workspace components delivered with clean UX. Hand-off to Law and Tax teams.
- Jul / Aug Law showcases the integration at industry conferences using the delivered core platform.
AI-assisted. Judgment-led.
The process was structured in two layers — product architecture first, visual exploration second. AI tools accelerated both, but the critical decisions — which panels to build, how agent routing surfaces, where to prioritize transparency — came from design judgment shaped by research and stakeholder alignment.
-
01Architecture before screensDesign · WhiteboardMapped entry points, decision moments, agent logic, and workspace structures before any UI generation began. The hardest part of the project — and the part that made everything downstream coherent.
-
02Structure layerChatGPTProject brief, UX architecture documentation, prompt plan, design-generation prompts, and direction-merge strategies. Used to externalize and stress-test architectural decisions before moving to visual output.
-
03Visual explorationFigma MakeEarly interface directions, full-screen layout exploration, individual panel refinement, and merging multiple generated outputs into a coherent product direction.
-
04Refinement and handoffFigmaFinal screen production, interaction state documentation, design system contribution, and annotated engineering specifications. Clickable prototype with full UX architecture documentation.
The key insight: generating screens was never the hard part. The work was defining a coherent product architecture first — panels, entry points, agent routing, workspace logic — before any UI generation produced meaningful output.
A full-screen AI Mode built around a multi-panel workspace.
AI Mode is opt-in — legacy experiences stay available for users who prefer them. The design had to feel capable enough for complex professional research while giving every user a clear, predictable entry and a coherent path through the full workflow.
The workspace expands progressively as users need more context — starting with a focused single-panel chat, opening to a two-column layout, then a full three-panel workspace. Panels are additive, not overwhelming.
Users can surface and switch agents without disrupting the conversation. The modal surfaces recent agents first, then all available agents — letting users understand the system without needing to understand its architecture.
On mobile, panels collapse into a bottom sheet pattern. The chat stays visible and in focus; sources and context slide up on demand without replacing the screen. The same panel logic applies — just adapted to a single-column viewport.
The hardest call: how much to unify.
Too much integration crowds the workspace. Too little replicates the same fragmentation in a new shell. Every major decision was tested against this tension — entry points, panel logic, agent transparency, and the operating model itself.
-
The problem was perception, not discoverabilityAI was already present in the product. The failure was that users saw a chatbot, not a platform. The design had to reframe what AI could do — not just make it easier to find.
-
AI needs structure to earn trustA blank conversational interface isn't enough for complex legal or tax research. Users need panels, source visibility, workspace context, and clear navigation before they'll act on AI-generated output.
-
Predictability precedes adoptionUsers are more likely to act on AI answers when they understand where answers came from, which agent responded, and what the next step is — not just what the answer says.
-
Core design leads; BUs extendEnforcing a center-out design model was itself a design decision. Without it, three business units would have shipped three incompatible AI experiences on top of the same infrastructure.
What engineering received.
The handoff was structured to give engineering a complete picture — not just screens, but the logic, constraints, and rationale behind every system interaction.
Full workspace and AI assistant designs with annotated panels, interaction states, and workspace structure. Figma Make outputs integrated into final file.
Unified panel structures and AI interaction patterns contributed to the shared component library using Shadcn components.
Clickable prototype with UX architecture docs covering panel logic, AI and workspace interactions, agent routing, entry points, and legacy access patterns.
Success scoped around Amplitude product analytics, structured user testing, and post-launch qualitative research across Law, Tax, and Government.
Pre-launch signals and success criteria.
The product has not launched. Impact will be measured across behavioral, workflow, and qualitative dimensions once Law and Tax deploy their domain extensions on the platform.
Users are overwhelmed not by a lack of tools, but by the number of surfaces they have to switch between to complete a single research task.