Role
Product designer
Team
CEO, 4 developers, 1 designer
Timeline
Feb 2025 – Feb 2026
Status
Shipped
Services
RESEARCH
AI WORKFLOWS
DESIGN SYSTEM
STAKEHOLDER MANAGEMENT
Overview
LexFriend is an AI workspace that helps lawyers structure cases, draft documents and prioritise actions. I designed its AI interactions for legal work, and used Claude in my own process to benchmark the dossier wizard.
18
legal processes grouped into 6 domains
3 → 1
separate tools brought into one workspace
9
fixes in the dossier wizard, no extra steps
2
law firms researched, each with 10 to 20 staff
01
Context
Built while it shipped
• Constraints
Design and development ran in parallel
Every change had to fit a running sprint, without launch metrics.
Legal accuracy came first
AI output ends up in client work and billing, so the lawyer confirms every result before it is saved.
Desktop first
Lawyers work at a desk, with documents side by side.
02
Role
My role
As the only designer, I owned research, interface direction and the design system, from first interviews to handover.
I decided
How the design system was built
Which flows to redesign first
Not my call
Product roadmap
AI model choices
Backend architecture
03
Research
Three tools, no shared context
I interviewed lawyers and secretaries at both firms. They described one problem: work scattered across tools, and dossiers only their creator could follow.
• Key findings
Three separate tools
Email, dossiers and documents lived in separate tools, so none showed the whole case.
No shared context
Secretaries had to interrupt a lawyer for case context.
Unbilled hours
Time was logged afterwards, so billable work went missing from invoices.
Legal process blueprint
I mapped the legal process end to end.
18 processes in 6 domains. With stakeholders, I used it to decide which domains to improve first.
04
Foundation
The foundation, sprint by sprint
Sales-driven features came first, while the UI foundation for dossiers, documents and communication fell behind.
• Decision
I built the design system inside feature work, and agreed with the team that every sprint still shipped a visible feature.
• Result
140 colour variables, 36 typography styles, 2 breakpoints.
Variable structure
Without foundation
With foundation
Semantic tokens in use
05
AI in the product
Predictable before flexible
01. Time registration
Lawyers log time by typing a short note to the AI assistant, which turns it into a time entry they confirm before it is saved.
• Decision
One fixed note format, because the model couldn’t yet read free phrasing reliably.
• Trade-off
Less flexible, but lawyers stopped filling in time forms by hand.
• What followed
Once the model handled free phrasing, the same input took voice.
02. Dossier summary
One summary of status, financials and next actions, so lawyers and secretaries work from the same picture.
• Decision
One summary for both roles; two views would have kept the information gap open.
• Trade-off
A summary is only as good as the dossier data behind it.
06
AI in my process
Claude compared, I decided
I redesigned the dossier wizard first, then used Claude to test it against seven competitor flows. Claude compared; I decided.
How the benchmark worked
• My question
Does billing belong in creating a dossier at all?
• Compared
My redesign against Xero, Upwork, Jobber, Midday, Toggl Track, Revolut Business and Mercury, on the same 11 points.
• What Claude did
Through the Figma and Mobbin MCP servers, Claude pulled real competitor screens into Figma, grouped the patterns and turned gaps into recommendations tied to their source screens.
Before my redesign: a bare progress bar, and Cancel as the only way out.
My redesign next to seven competitor flows, on 11 points.
How Xero, Upwork and Midday handle drafts, review and billing.
Claude’s output: strengths, gaps and recommendations.
What I kept, took and turned down
The benchmark confirmed four of my choices: named steps, labelled optional fields, billing as cards and a review that follows the steps. Three decisions show where I came in.
Research + Claude
01. Save as draft
• From research
I watched people start over because a half-finished dossier couldn’t be saved.
• From the benchmark
Claude recommended Save as draft and a Drafts tab, as in Xero and Upwork.
• My decision
I added both: Save as draft on every step, and a Draft tab in the overview.
Before: Cancel only.
After: Save as draft next to Cancel.
Saved drafts get their own tab, with Undo right after saving.
Claude, adapted
02. An ‘Edit’ on every row
• From the benchmark
Claude recommended an Edit on every review row, editing in place.
• My decision
I took the Edit per row, but each one jumps to its field in the step. Editing in place would have duplicated the form logic; jumping back reused the forms we had.
Before: one Edit for everything.
After: an Edit on every row.
Claude, not followed
03. Billing stays in the wizard
• From the benchmark
Claude’s main advice: move billing out of the wizard and drop Skip for now.
• My decision
I kept billing in, but optional. Research had shown unbilled hours, so the wizard asks at creation. But at intake a secretary can’t always choose: sometimes the lawyer decides later, via Edit dossier.
Before: billing hidden in a dropdown.
After: billing types as cards, step marked optional.
After: four steps, nine changes.
07
Impact
What changed in the daily work.
3 tools, none showing the whole case
Time logged afterwards, hours missing from invoices
Case context only in the lawyer’s head
08
Reflection
Map the transitions earlier
I would map critical workflows end to end earlier, including business rules, permissions and exceptions. In modular SaaS, the risk sits in the transitions between parts.
• Key learnings
Keep the lawyer in charge
Confirmation before automation.
Consistency is a speed argument
Framing the foundation as speed got the organisation on board.
Design for today’s model
Scope AI to what works reliably now, then add flexibility.
