Case study • 2026 • Rocket Lawyer
The user, the AI, and their attorney, in one loop.
Designing Rocket Copilot, Rocket Lawyer’s AI assistant for small businesses: where the work lives, how the AI behaves, and who makes the call.
This year I led the design of Rocket Copilot. We moved from selling one document at a time to an AI that carries a business’s legal work forward, and brings in a real attorney when a decision needs one. I own the end-to-end experience and how the AI behaves: what it does on its own, when it stops, and when it hands off to a person.
Where the work lives
An AI workspace
Chat for thinking, a place for AI to do the work, and a rail that shows where things stand.
How AI behaves
Five shared AI states
Mandate, Working, Needs you, Resolved, Watching. Every agent workflow plugs into them.
Who makes the call
AI → attorney → user
AI flags a legal issue, the attorney decides, and AI prepares the next step.
Trust
Five AI trust principles
The rules for every prompt change and every screen where AI works with an attorney.
01 • Snapshot
What I owned.
01
I lead the design end to end
02
I own the agent behaviors
03
I own one AI framework
04
Attorney inside the loop
Scope • five areas, one AI experience
From the first question to the attorney. Every area runs on the same AI framework.
02 • Business context
Today, Rocket Lawyer is transactional.
For the business, we sell one document at a time, and few people come back. Small business users create about half of all documents on Rocket Lawyer, but only 7.2% of people who form a business with us create another document in their first 30 days.
For the user, the document gets done, but the legal situation doesn’t. A hire needs a wage notice, a contract renews, a state wants a filing. Nothing in the old experience carried that forward.
The strategic design challenge
How might we take care of a small business’s whole legal situation, instead of selling one document at a time?
Rocket Lawyer today
Home is a to-do list of documents. Each screen is a separate transaction: make a document, upload a document, start again.
Before → After
Before
One document at a time.
Each legal task is a separate transaction, and the user works out what comes next.
Reframe
After
The whole legal situation.
Copilot carries the work forward, and brings in an attorney when a decision needs one.
03 • How I work
I design how AI behaves, starting from what people expect.
I’m a research-led, AI-native designer. Three habits shape everything that follows, and each one is something I actually hand off to engineering, content design, and prompt engineering.
01 • The AI model first
I map the AI model before I draw a screen
The states the agent moves through, and the loop where it brings in an attorney on its own.
02 • Behavior on the screen
I label the AI behavior on every screen
What she expects comes from research. What the AI does is a rule: revealed in the moment, or always.
03 • Frameworks
I prototype frameworks, not screens
One frame with named slots, which every workflow fills.
04 • Decision 1 • Where the work lives
Keep chat for thinking. Give AI a place to do the work.
We started where leadership wanted: put AI in chat. Users could ask a question and create a document. But in research, people told us they could finish a task without feeling their legal situation was moving forward. They trusted the AI. They didn’t feel confident.
Chat
The task gets done, and then the thread scrolls on. To find the contract or its signing status, the user scrolls back up through the conversation.
Tabs
Tabs organized the work, but one task now lived in three places, and users still worked out the next step themselves. When chat was removed, people asked where it went.
Before → After
Before
Chat or tabs.
Chat helped people think. Tabs helped them do. Neither gave AI a way to carry the work forward.
Reframe
After
Think, do, ground.
One workspace: chat to talk it through, a panel where AI does the work, and a rail that shows where things stand.
The workspace
Chat on the left, the AI’s work in the middle, status on the right. The user can ask questions, see progress, and step in when the AI needs a decision.
Problem
Users finished tasks without feeling their legal situation was moving forward, and lost chat when we moved to tabs.
Design decision
Keep chat, and give AI a workspace to track the task, show progress, and ask for a decision when it needs one.
Evidence
In testing, trust held at 4.2 and confidence rose from 3.5 to 4.0. Leadership adopted it as the foundation framework.
Shift
From choosing chat or tabs → to one workspace where people think with AI, and AI moves the work forward.
05 • Decision 2 • How AI behaves
Design the agent’s behavior once, so every workflow can use it.
Four agent workflows were moving at once, each team designing its own flow. Looking across them, I saw the AI asking permission, showing progress, and handing back to the user in different ways. This was more than design consistency. It was agent-behavior consistency.
Same moment, two teams
Follow the steps: permission, progress, asking the user, done, and two weeks later. Business filing and hiring handle each one differently.
So I stopped thinking in screens. With AI engineers and prompt engineers across teams, we pulled the flows apart and asked the same questions of each. What can AI do on its own? When should it stop? When does it need the user?
Underneath, the same five things
The screens look different, but the agent does the same five things in every one.
These aren’t five screens. They’re how the agent behaves. I shared them through a series of workshops that moved teams from prototyping screens to prototyping agent behavior.
Five states in one hiring flow
The steps light up as the agent moves from permission, to working, to a decision only the owner can make, to done, to watching the dates.
Opposite workflows, same foundation
Hiring and ending employment have opposite outcomes and opposite emotions, but they share the same agent behaviors and the same shell. What changes is the content, not the layout.
Problem
Agent behavior was drifting across four workstreams. We’d have reinvented the agent for every new task.
Design decision
Five shared AI states that every agent workflow plugs into, defined with AI and prompt engineers.
Outcome
A framework built for growth. New workflows get permission, progress, decisions, completion, and monitoring, and the screens can still differ.
Shift
From designing screens team by team → to one set of agent behaviors every workflow shares.
06 • Decision 3 • Who makes the call
Connect AI and the attorney, and make the judgment clear.
In legal, as in healthcare and finance, AI can spot an issue, but a person has to make the call. We didn’t want the user to leave Copilot, repeat everything to an attorney, then come back to work out the next step. So I designed one connected journey: AI flags the issue, the attorney reviews it, and AI turns the advice into the next step.
The first connected flow
AI spots that the business may need to register in another state. The attorney reviews it. AI prepares the filing from her guidance.
Then I found a problem in my own design. The handoff was so smooth that the attorney’s advice and the AI’s next step blended into one message. People approved, but they didn’t understand what they approved. Our negotiation research showed the same pattern.
What we learned
They approved the summary, not the deal.
The AI’s sentence said “keeps your limits.” The draft broke two of them. The real terms were one click away, behind “Review draft,” and the design never asked anyone to look.
With data science and prompt engineering, I looked at what the AI actually said when it passed along an attorney’s advice, and checked those answers with simple questions. We kept finding the same gaps, so I turned them into five principles. They became the rules for every prompt change and every screen.
The same flow, with the principles applied
Same journey, same effort. At every handoff, the user knows whose judgment they’re following: what AI isn’t sure about, who owns the decision, the attorney’s own words, exactly what they’re approving, and when AI stops.
Problem
A smooth handoff made people approve what they didn’t understand.
Design decision
One connected AI-and-attorney flow, governed by five trust principles.
Outcome
The rules for every prompt change and every screen where AI works with an attorney.
Shift
From a smooth handoff → to a clear one, where the user always knows whose judgment they’re following.
07 • Impact
Six months in, Copilot turns visitors into paying customers.
Trial → paid conversion
4×Copilot users convert at 11.3%
versus 2.9% for users without Copilot.
of new users interact with Copilot
Copilot satisfaction, 72% promoters
Rocket Lawyer NPS, Excellent band
Workspace testing
Trust held at 4.2 while confidence rose from 3.5 to 4.0.
First half of 2026. Funnel Mar–Jun, Copilot survey May–Jun, NPS Apr–Jul. NPS is the whole Rocket Lawyer brand, not Copilot alone.
08 • What’s next
Where the loop goes from here.
Business home
From a task list to approving AI next steps
The owner approves what Copilot does next, with protection the agent runs on its own.
Legal Health Check
The check joins the loop
What the check finds opens the work, and what the work finishes feeds back into the business record.
Open decision
Attorney advice before approval
Letting the user bring in an attorney before approving an AI recommendation, which isn’t allowed today.
09 • Reflection
What I learned
Keep the work connected, and make the judgment clear.
In legal, AI can carry the work, but a person has to make the call. Every surface I designed had to show what AI is doing, when it stops, and whose judgment the user is following.
The product story
From
“One document at a time.”
To
“The user, the AI, and their attorney, in one loop.”
Case study · 2026 · Li Song • Lead Product Designer, AI • Rocket Lawyer