Designing an AI legal workflow lawyers can trust
FormPhil drafts legal documents with AI. I designed the workflow that lets lawyers review, verify and stay accountable for what it produces, and took it from early demo to launch in 13 weeks.
Sole product designer · PM, BA and engineers · CORTO, part of LEAP Legal Software · Launched October 2026
Early validation: forms took under 10 minutes, down from 30–60, in pre-launch testing.
The review workflow: AI-filled values highlighted, sources beside the form, one clear next step. Recreated for confidentiality.
At a glance
The challenge: the AI could already draft. Lawyers couldn’t see where its values came from, so they couldn’t stand behind the output.
My reframe: from “make the AI generate faster” to “help lawyers trust, review and take responsibility for AI output”.
What shipped: a review workflow, a source panel, Ask Client and shared trust patterns, launched on time.
Adoption depended on trust, not speed
FormPhil is an AI legal document platform at CORTO, part of LEAP Legal Software. It drafts documents from LEAP matter data, client answers and source documents, for lawyers and legal assistants.
The opportunity: help law firms scale document preparation without removing professional judgement. But lawyers are accountable for every document, so if they can’t verify a value, they won’t use it.
Sole designer, embedded in the squad
• Owned the workflow, interaction design and UI, from early demo to launch.
• Worked daily with the PM, BA and engineers: standups, refinement, handoffs and staging reviews.
• Proposed the AI–lawyer responsibility boundary and agreed it with the PM.
• Designed the component patterns and worked with engineers to implement them.
AI drafts. Lawyers decide.
I designed the product around one responsibility boundary, set by risk rather than by what the AI could technically do.
Automate, where risk is low. Confident values are filled and highlighted, so lawyers see what the AI did.
Judge, where accountability matters. Uncertain values go to the lawyer, with the source beside them.
Ask, where information is missing. The client answers, through Ask Client.
Review is the workflow, not a final checkbox
Lawyers were searching the database for sources every time. A single approve button would have been faster, but it invites blind approval.
So I designed review into the flow:
• AI-filled values highlighted, so lawyers know what to check.
• A source panel showing the document, page and paragraph for each value, inside the form.
• A sticky bar that always shows the one next step.
Ask Client: changing the question
For conflicting values, the PM proposed a quicker modal to pick one. But the documents don’t say which value is right, so lawyers would still call the client. I recommended Ask Client instead, made both costs visible, and the PM made the call.
Proposed: choose a value. Still ends in a call to the client.
Shipped: lawyers preview and edit questions, add follow-ups, and send by email or link.
Designed around the lawyer’s next decision
Before, it was one flag that emailed the client straight away. I designed four states, each tied to an action:
• Pending: prepared, not sent. Review and send.
• Awaiting response: nothing to do yet.
• Client not sure: follow up with the client.
• Response received: review before it’s used.
The PM and I grouped questions by state, and a ribbon counts each one. Clients can answer “I’m not sure” instead of guessing, on desktop or mobile.
“Previewing and editing before sending gave us more control.” Lawyers in testing
Make the system predictable
Testing changed the design twice. Lawyers asked “Is this sending already?” at Send Email, so it became Draft Email in a guided flow. They were also unsure during the automatic portal transfer, so I added an explanation and progress. After both changes, they no longer hesitated.
Patterns that carry trust
With 13 weeks and no design system, I designed patterns as they repeated, on shadcn/ui. Engineers built them once and reused them across form types, so lawyers learned the trust signals once.
Protecting the outcome inside the constraints
Preview vs edit: engineering preferred preview-only to save effort. The question speaks for the lawyer, so we kept editing, and testing confirmed it.
States and edge cases: I worked through every Ask Client state with engineers during refinement and build, including clients who aren’t sure.
Staging sign-off: I reviewed builds in staging before sign-off, so what shipped matched what lawyers needed.
What we left for the next release
I proposed opening source files directly from the source panel. With a fixed date, the PM, engineering and I prioritised the core review workflow instead. The panel still shows document, page and paragraph, and the design is ready for the next release.
Launched on time, with early validation
Delivered: launched October 2026 for lawyers and legal assistants, with the review workflow, Ask Client and shared patterns. Part of version 5 designed.
Validated early: 30–60 → under 10 minutes. Pre-launch testing with two lawyers and a BA on divorce forms. Early validation of the workflow, not a production metric.
Next to measure: adoption, completion time by form type, review time, and where lawyers override AI values, read together to spot blind approval.
What I’d carry into any AI product
• Trust is designed through the workflow, not added through explanation.
• Ask what the user does next. It separated Ask Client from the modal.
• Next time, I’d agree success metrics at kickoff and test the client’s side earlier.
Designing AI into complex workflows?
I’d love to hear what you’re working on.