Helping people finish a legal document they actually trust
Rocket Copilot Document Generation helps people create legal documents through a guided conversation instead of a traditional legal questionnaire. As Lead Product Designer and sole design owner, I led a nine-week redesign across six tested iterations, reframing the experience around the user’s real-world situation before partnering with Product, Engineering, Research, and Legal to ship the final flow.
The pivotal decision was to stop asking people to configure a legal document and instead let them identify their situation while Copilot handled the legal reasoning. That strategic shift became the foundation for every design decision that followed, from the interaction model to the recommendation system and trust principles.
MY ROLE
Lead Product Designer and sole design owner:
led product strategy, end-to-end interaction design, research planning, and final design direction
THE BET
People would complete documents more confidently if they could identify their situation and confirm explained recommendations instead of making legal decisions from scratch.
WHO I WORKED WITH
Product:
funnel strategy, scope, and success measures
Engineering:
technical feasibility and recommendation logic
Research:
study design and six rounds of user testing
Legal:
clause accuracy, guardrails, and recommendation boundaries
WHAT TESTING VALIDATED
Production analytics established the problem: only 2.2% of users completed the original experience. Across six rounds of usability testing, the situation-first direction consistently improved perceived ease of completion, while confidence in the resulting document remained the primary design challenge. These findings shaped the final experience that shipped.
Business Context
A high-intent funnel was losing nearly every user before completion.
Document Generation sits at a critical point in the Copilot funnel: it turns high-intent requests, like creating an NDA, into completed legal documents and membership conversions. However, only 21 of 951 users who started the experience completed it end to end (a 2.2% completion rate) with an average completion time of 78 minutes. The post-launch review identified it as one of the three most critical leaks in the Copilot funnel.
I reframed the problem from interface friction to decision-making friction. Users were not abandoning because the questions looked difficult; they were abandoning because the experience required legal judgments before giving them enough context to make those judgments confidently.
The strategic challenge became: How might we help people complete a trustworthy legal document without requiring them to make legal decisions they are unprepared to make?
User Problem
People weren’t struggling to fill a form. They were struggling to feel sure.
Reviewing session recordings alongside funnel data revealed a consistent pattern: users hesitated at the upfront legal questions, reread them repeatedly, attempted to type directly into the document preview, and then abandoned. The issue was not simply the number of fields. Users were being asked to make legal decisions without enough context, while the editable-looking preview created a second and competing place to act.
“Do I even need this clause?”
“What happens if I pick the wrong option?”
“Wait.. am I supposed to type in the chat or in the document?”
Two root causes emerged:
1. Legal decisions were front-loaded. Users had to choose clauses, terms, and configurations before understanding their implications.
2. The experience created competing interaction surfaces. The editable-looking preview drew users away from the guided conversation and made it unclear where they were expected to act.
Strategic Reframing
The reframe: help people recognize their situation instead of configuring legal terms.
I evaluated three directions against completion potential, user trust, and implementation feasibility:
1. Add explanations to the existing questions. This was the lowest-cost option, but it treated uncertainty as a content problem without removing the legal decisions causing it.
2. Progressively disclose the same questionnaire. This reduced visual overwhelm, but users would still need to make unfamiliar legal choices.
3. Reframe the experience around the user’s situation. Copilot would recommend the legal setup while users retained the ability to review and edit every decision.
​
I recommended the third direction because it was the only option that removed the underlying cognitive burden instead of reorganizing it. It required more recommendation logic and legal guardrails, but offered the strongest opportunity to improve completion and trust together. This became the turning point of the project. Every subsequent design decision reinforced this new mental model instead of the original legal questionnaire.
Making Copilot’s reasoning visible
The new flow became: choose a situation → review Copilot’s recommended legal setup → edit any recommendation if needed → provide only the remaining factual details, like names, dates, and addresses.
​
Every recommendation needed to answer three questions:
Why does this fit my situation?
What is still required from me?
How can I change it if I disagree?
​
This balance was critical. Reducing decisions created speed, but explaining and reversing those decisions preserved trust.
Interaction System
One guided system, not a form bolted to a chatbot.
I unified the questionnaire, conversation, and document preview around one interaction rule: Copilot recommends, the user confirms, and the document reflects the decision. Each surface received one clear responsibility, eliminating competing inputs and creating a single, predictable interaction model.
HOW THE FLOW MOVES
THE INTERACTION SHIFT
Why Guided Creation Works
Shifting the task from legal interpretation to situation recognition
Traditional legal questionnaires assume that users understand the terminology and consequences behind every choice. The guided model shifted the task from interpreting legal language to recognizing a familiar situation, reviewing a recommendation, and confirming or changing it.
​
Testing showed that this direction substantially improved perceived ease. However, comprehension declined when users reached dense clause language, revealing the next trust challenge: helping people understand the document they had successfully created.
Traditional Questionnaire
✗ Requires users to interpret unfamiliar legal terminology
✗ Presents one long flow with limited feedback
✗ Allows incorrect choices to fail silently
✗ Provides little sense of progress
Copilot-Guided Creation
✓ Uses plain-language, situation-based prompts
✓ Presents one decision at a time
✓ Provides explained recommendations instead of blank choices
✓ Makes progress and next steps visible
DEEP DIVE
Situation-First Setup
Turning smart defaults into explainable and reversible recommendations
Problem: Situation-based recommendations reduced effort, but they could also feel opaque if users could not understand or revise the legal decisions Copilot made for them.
Design decision: I paired situation-based defaults with visible reasoning, editable recommendations, and an explicit path for users who preferred to make their own choices. This allowed Copilot to reduce effort without removing agency.
Tradeoff: More automation increased speed but risked making the experience feel like a black box. I addressed that risk by showing the factors behind each recommendation, making every choice editable, and preserving a manual path for confident users.
What changed:
-
Users selected a real-world situation before seeing legal configuration choices.
-
Copilot generated a recommended legal setup for that situation.
-
Each recommendation included the factors considered and an explanation of why it fit.
-
Every decision remained editable before document creation.
-
A manual path preserved control for users who preferred to choose independently.
DEEP DIVE
Guided Answers
Reducing effort without hiding the user’s own information
Problem: After Copilot handled the legal setup, users still needed to provide factual details such as names, dates, addresses, and party information. Open text fields increased effort and left users unsure what a complete or appropriate answer looked like.
Design decision: I kept the remaining inputs inside the conversation and used quick choices, contextual examples, and “Help me answer” guidance to reduce typing without silently completing personal details on the user’s behalf.
What changed:
-
Remaining questions focused only on user-specific information, not legal decisions.
Quick choices replaced free text when answers were constrained.
-
“Help me answer” provided examples and plain-language guidance.
-
Progress confirmations made the remaining effort visible.
-
All inputs stayed in the conversation, preventing competition with the document preview.
What I chose not to do:
I avoided silently auto-filling personal information. Users needed to understand and confirm the facts entering their legal document, so automation reduced effort without removing participation.
DEEP DIVE
Live Read-Only Preview
Turning the preview from a trap into a trust signal.
Problem: Blank, editable-looking fields made the preview appear to be a second input surface. Users split their attention between the document and conversation, attempted to type directly into the preview, and lost track of where they were expected to act.
Design decision: I made the preview read-only and updated it live as users answered questions in the conversation. The conversation became the only place to act, while the document became visible confirmation that each answer was being applied correctly.
What changed:
-
The preview became clearly read-only.
-
The document updated in real time as users answered.
-
The conversation became the single place for input and decisions.
-
Microcopy explained that the preview would update automatically.
Trust and Design Principles
A reusable model for reducing effort without removing understanding or control
01
Recognize, don’t configure
Let users identify their situation instead of translating it into legal settings.
02
Recommend, explain, and preserve control
Copilot can reduce decisions only when its reasoning is visible and every recommendation remains editable.
03
Create one place to act
The conversation captures decisions; the document preview reflects them. Each surface has one unambiguous role.
04
Escalate when confidence is insufficient
Make qualified human support reachable when AI guidance alone cannot provide enough assurance.
Extending the Trust Model to Membership
Applying the same recommendation principles at the upgrade moment
Completing the document created a high-intent membership moment, but the recommendation needed to preserve the same principles as the creation flow, visible reasoning, clear value, and user control. I applied the trust model by showing the completed document first, connecting membership benefits to the user’s next actions, and explaining why a specific plan was recommended.
WHAT MEMBERSHIP UNLOCKS AFTER THE DOCUMENT IS COMPLETE
Sign Documents
Send the completed document for signature
Connect with an Attorney
Ask a qualified attorney about the document
Manage Documents
Edit, store, and track documents in one place
Recommendation Logic:
Testing showed that generic language such as “based on your activity” felt opaque and potentially manipulative. I made the recommendation logic explicit by connecting each plan suggestion to a visible signal, such as the document type, a selected next action, or a stated need.
Membership Plans and Pricing
Testing identified three conversion risks:
-
Pricing transparency: Discounts, billing cadence, and checkout totals needed to remain consistent.
-
Recommendation transparency: Users needed to understand why a specific plan fit their activity and document.
-
Control and confirmation: Cancellation terms and document access needed to be explicit near the decision point.
The strongest testing signal was not visual polish but transparency. Users accepted being offered a membership, but ambiguity around price, cancellation, document access, or recommendation logic quickly weakened confidence. Showing the completed document before the upgrade request and explaining the recommendation had more influence on willingness to continue than additional visual refinement.
Testing and Evidence
How six successive studies shaped the shipped direction
Production analytics established the original-flow baseline: a 2.2% end-to-end completion rate and a 78-minute average completion time. Over nine weeks, I led six successive UserTesting studies, the original experience as a benchmark followed by five redesigned iterations. Every round used the same measures for perceived ease of completion and trust in the resulting document, and each study directly informed the next design.
Benchmark (Original front-loaded flow): Familiar, but difficult to trust and complete.
V1 (Lean situation-first flow): Highest overall preference, but limited trust support.
V2 (Recommendations with visible reasoning): Best balance of ease and trust.
V3 (Maximum trust scaffolding): Highest trust, but greater interaction effort.
V4-V5 (Streamlined guided flows): Strongest ease scores, but smaller trust gains.
​
Each version was evaluated on two consistent dimensions, perceived ease of completion and trust in the final document.
FINAL DESIGN DIRECTION:
I used V4’s low-friction interaction model as the foundation and reintroduced the strongest trust elements from V2 and V3: visible recommendation logic, editable defaults, and prominent access to a Legal Pro.
Impact
What shipped and how user behavior changed
59.67%
Interview completion rate
Up from 19.65% in the original flow.
11.78%
Document completion rate
Up from 2.2% in the original experience.
18 minutes
Average completion time
Down from the original 78-minute average, reducing the time required to create a finished document.
67.2%
Selected a situation-based preset
Most users began with a real-world situation rather than configuring unfamiliar legal terms from scratch.
Post-launch data validated the core design hypothesis. Compared with the original experience, the native flow tripled interview completion (19.65% → 59.67%) and increased end-to-end document completion 5.4× (2.2% → 11.78%). In addition, 67.2% of users adopted a situation-based preset, reinforcing the shift from legal configuration to situation recognition.
Production results
Research predicted the direction. Production data measured the outcome.
Six rounds of testing identified the situation-first model as the strongest direction for reducing effort while preserving trust. After launch, production analytics confirmed those improvements translated into measurable gains across the funnel. Starting from the same number of document starts, interview completion increased from 19.65% to 59.67%, document completion increased 5.4× (2.2% → 11.78%), and trial-to-paid conversion increased 4.5× (1.6% → 7.2%).
3.0 X
Interview completion
59.67% vs 19.65%
5.4 X
Document completion
11.78% vs 2.2%
4.5 X
Trial-to-paid conversion
7.2% vs 1.6%
Reflection
What I carried forward
The hard part of AI document creation isn’t generating the document, modern AI models already do that well. It’s deciding how much thinking to take off the user’s plate without making them feel like they’ve lost control of something legal and consequential.
-
Research became a tool for alignment, not only validation.
Product, Engineering, and Legal initially held different views on how much Copilot should decide. Evaluating each version against both ease of completion and trust in the document gave the team a shared framework for resolving tradeoffs and selecting the final direction.
​
-
The problem was behavioral and strategic, not visual.
The strongest improvement came from changing what users were expected to decide, not from shortening the form or refining the interface. That reframe allowed the team to address the root cause and establish a reusable model for future Copilot recommendations.
​
-
Shipping was not the end of validation.
After launch, the next responsibility was to measure whether research gains translated into live completion, faster task time, fewer clarification requests, and greater confidence in the finished document.