JuriCloud
JuriCloud
Designing an evidence-first legal AI research workspace
Designing an evidence-first legal AI research workspace
Roles
Roles
Product Designer
Product Designer
Scope
Scope
Product strategy, conversational AI UX, information architecture, interaction design, prototyping, design system
Product strategy, conversational AI UX, information architecture, interaction design, prototyping, design system
Focus
Focus
Trust, verification, legal research workflows, reusable AI interaction patterns
Trust, verification, legal research workflows, reusable AI interaction patterns
Company
Company
JuriCloud
JuriCloud
TL;DR
TL;DR
JuriCloud already had strong legal AI infrastructure, including custom retrieval and citation-backed research. The challenge was turning that capability into a product lawyers could actually trust and use. I designed the interaction model, conversational workflows, trust patterns, and reusable system that helped shift the experience from “AI chat” to “legal research workspace.”
JuriCloud already had strong legal AI infrastructure, including custom retrieval and citation-backed research. The challenge was turning that capability into a product lawyers could actually trust and use. I designed the interaction model, conversational workflows, trust patterns, and reusable system that helped shift the experience from “AI chat” to “legal research workspace.”
Evidence-first legal AI workflow
Evidence-first legal AI workflow
|
|
Citation-backed answer review
Citation-backed answer review
|
|
Clarification and uncertainty states
Clarification and uncertainty states
|
|
Reusable conversational design system
Reusable conversational design system
What I designed
What I designed
Show system activity
Show system activity
Show system activity
Communicate observable stages such as retrieving, ranking, and checking sources without exposing internal model reasoning.
Communicate observable stages such as retrieving, ranking, and checking sources without exposing internal model reasoning.
Scope before generation
Scope before generation
Scope before generation
Ask for missing legal context before producing a detailed answer.
Ask for missing legal context before producing a detailed answer.
Cite visibly
Cite visibly
Cite visibly
Keep evidence connected to the claims it supports.
Keep evidence connected to the claims it supports.
Warn honestly
Warn honestly
Warn honestly
Weak, conflicting, or unavailable evidence should change the answer state.
Weak, conflicting, or unavailable evidence should change the answer state.
Legal AI needed trust, not just answers
Legal AI needed trust, not just answers
The Problem
The Problem
The product had strong underlying capability, but the experience still felt too close to a generic AI assistant. In legal research, that creates a trust problem.
The product had strong underlying capability, but the experience still felt too close to a generic AI assistant. In legal research, that creates a trust problem.
Why Generic AI Patterns Failed Here
Why Generic AI Patterns Failed Here
Lawyers do not just need answers. They need to inspect sources, evaluate whether a response is reliable, understand when the system needs more context, and recover safely when certainty is low.
Lawyers do not just need answers. They need to inspect sources, evaluate whether a response is reliable, understand when the system needs more context, and recover safely when certainty is low.
Citations inside the response
Source verification without breaking flow
Source verification without breaking flow
Clear separation between generated text and source material
Citations inside the response
Clear separation between generated text and source material
Most chat-based AI interfaces optimize for fluency.
Legal research requires something else
Context before generation
Context before generation
The challenge was not to create another AI chat interface.
It was to make legal AI feel reviewable, source-grounded, and controlled enough to support research and analysis workflows.
The challenge was not to create another AI chat interface.
It was to make legal AI feel reviewable, source-grounded, and controlled enough to support research and analysis workflows.
Three issues shaped
the redesign
Three issues shaped
the redesign
The interface did not explain what the AI was doing.
Citations existed, but source quality and relevance were difficult to judge.
The product lacked reusable patterns for loading, clarification, evidence, uncertainty, and failure states.
The interface did not explain what the AI was doing.
Citations existed, but source quality and relevance were difficult to judge.
The product lacked reusable patterns for loading, clarification, evidence, uncertainty, and failure states.
This led to one core design question:
This led to one core design question:
How might we help legal professionals use conversational AI for research while preserving visibility into context, evidence, uncertainty, and professional review?
How might we help legal professionals use conversational AI for research while preserving visibility into context, evidence, uncertainty, and professional review?
From Chat Interface to Research Workspace
From Chat Interface to Research Workspace
I first simplified the navigation and existing screens so the platform was easier to enter and understand.
As I worked deeper into the product, the central problem became clear: legal professionals were not only asking for answers. They needed to inspect the basis of those answers.
I first simplified the navigation and existing screens so the platform was easier to enter and understand.
As I worked deeper into the product, the central problem became clear: legal professionals were not only asking for answers. They needed to inspect the basis of those answers.
The redesign shifted the platform from a basic prompt-response model toward a structured research workflow.
The redesign shifted the platform from a basic prompt-response model toward a structured research workflow.
Making AI output inspectable
Making AI output inspectable
Structure the response so users can inspect, extract, and reuse it.
Structure the response so users can inspect, extract, and reuse it.
Design System
The design system supported dynamic AI behavior, dense legal content, and both light and dark themes.
A reusable design system for JuriCloud’s legal AI workflows, covering conversational research, source inspection, matter context, prompt templates, attachments, loading states, citations, and errors.
We started with color.
The palette is warm, restrained, and evidence-forward. Neutrals carry most of the interface, gold is reserved for primary actions and active evidence states, and semantic colors are used only when they clarify risk, success, or status.
Typography for legal reading.
The type system is designed for long-form legal answers, compact controls, and dense source metadata.
Components built around conversational workflows.
Components were not created as isolated UI parts. Each component supports a specific AI moment: asking, scoping, attaching, searching, verifying, citing, saving, and recovering from uncertainty.
Design System
I built the experience around a reusable system for conversational AI in legal workflows. The design system had to support dense reading, layered context, component consistency, and variable AI states across light and dark themes.
A reusable design system for JuriCloud’s legal AI workflows, covering conversational research, source inspection, matter context, prompt templates, attachments, loading states, citations, and errors.
We started with color.
The palette is warm, restrained, and evidence-forward. Neutrals carry most of the interface, gold is reserved for primary actions and active evidence states, and semantic colors are used only when they clarify risk, success, or status.
Typography for legal reading.
The type system is designed for long-form legal answers, compact controls, and dense source metadata.
Components built around conversational workflows.
Components were not created as isolated UI parts. Each component supports a specific AI moment: asking, scoping, attaching, searching, verifying, citing, saving, and recovering from uncertainty.
Extending the system beyond product
Website
Outcome
Outcome
I designed the workflow across four connected layers:
This made the product feel less like a generic chatbot and more like a legal research workspace.
This made the product feel less like a generic chatbot and more like a legal research workspace.
Research entry
Empty and returning-user states for starting or resuming research
Empty and returning-user states for starting or resuming research
Prompt system
Support for short questions, long factual scenarios, templates, and refinement
Support for short questions, long factual scenarios, templates, and refinement
Context building
File uploads and attached cases to ground the model before generation
File uploads and attached cases to ground the model before generation
Answer review
Readable responses with inline citations and verification patterns
Readable responses with inline citations and verification patterns
The redesign gave JuriCloud a clearer product point of view: not a generic legal chatbot, but a structured workspace for evidence-backed AI research.
It also created a reusable foundation for future product surfaces by turning conversational AI behaviors into a system rather than a set of isolated screens.
The redesign gave JuriCloud a clearer product point of view: not a generic legal chatbot, but a structured workspace for evidence-backed AI research.
It also created a reusable foundation for future product surfaces by turning conversational AI behaviors into a system rather than a set of isolated screens.
Reflection
Reflection
The project changed how I think about conversational AI.
In trust-sensitive domains, the interface is not just a wrapper around the model. It is the mechanism that makes the intelligence usable.
The biggest lesson was that citations, uncertainty states, and context controls are not supporting details. They are the product.
The project changed how I think about conversational AI.
In trust-sensitive domains, the interface is not just a wrapper around the model. It is the mechanism that makes the intelligence usable.
The biggest lesson was that citations, uncertainty states, and context controls are not supporting details. They are the product.
