Everlaw

Improving clustering data visualization

🏢

B2B SaaS

🏢

B2B SaaS

🖥️

Desktop

🖥️

Desktop

Sifting through millions of documents is daunting. In a world where people are generating more data than ever, identifying evidence is getting harder for legal teams. To reduce the time and effort it takes to understand documents, we built key improvements to Everlaw's clustering AI data visualization.

Everlaw is a cloud ediscovery platform serving several Fortune 100 corporations and all 50 state attorney generals.

I was the sole product designer on this feature, collaborating with a product manager and 2 engineers. Design on this project lasted February–March 2023.

Everlaw is a cloud ediscovery platform serving several Fortune 100 corporations and all 50 state attorney generals.

I was the sole product designer on this feature, collaborating with a product manager and 2 engineers. Design on this project lasted February–March 2023.

Context

How clustering helps legal teams

Context

How clustering helps legal teams

Legal cases, especially government and corporate cases, involve huge volumes of data. These millions of documents might hold critical evidence, and it's essential to legal teams to know where to start looking.

Everlaw's clustering tool uses unsupervised machine learning to enable early case assessment (ECA) by grouping conceptually similar documents based on key terms and identifying high-level trends for legal teams to review.

Legal cases, especially government and corporate cases, involve huge volumes of data. These millions of documents might hold critical evidence, and it's essential to legal teams to know where to start looking.

Everlaw's clustering tool uses unsupervised machine learning to enable early case assessment (ECA) by grouping conceptually similar documents based on key terms and identifying high-level trends for legal teams to review.

Legal cases, especially government and corporate cases, involve huge volumes of data. These millions of documents might hold critical evidence, and it's essential to legal teams to know where to start looking.

Everlaw's clustering tool uses unsupervised machine learning to enable early case assessment (ECA) by grouping conceptually similar documents based on key terms and identifying high-level trends for legal teams to review.

Legal cases, especially government and corporate cases, involve huge volumes of data. These millions of documents might hold critical evidence, and it's essential to legal teams to know where to start looking.

Everlaw's clustering tool uses unsupervised machine learning to enable early case assessment (ECA) by grouping conceptually similar documents based on key terms and identifying high-level trends for legal teams to review.

💵

ECA can save firms from tens of thousands to millions of dollars.

Early case assessment is the process of quickly identifying facts to estimate legal liability and calculate risks and costs. Depending on how well it's done, this can be a huge savings or cost center for firms.

💵

ECA can save firms from tens of thousands to millions of dollars.

Early case assessment is the process of quickly identifying facts to estimate legal liability and calculate risks and costs. Depending on how well it's done, this can be a huge savings or cost center for firms.

💵

ECA can save firms from tens of thousands to millions of dollars.

Early case assessment is the process of quickly identifying facts to estimate legal liability and calculate risks and costs. Depending on how well it's done, this can be a huge savings or cost center for firms.

Problem

Scaling to thousands of terms

Problem

Scaling to thousands of terms

Problem

Scaling to thousands of terms

The current clustering interface focuses on large shapes of data, making it difficult to dig down and investigate individual terms identified by machine learning, and tedious to take action with terms of interest. This was especially critical for a strategic client who raised this issue, wanting to use the clustering tool on a 3 million document case.

Users need to better understand trends in their data by investigating cluster terms at scale, with a unified experience to make those insights actionable.

Current state
Current state
Current state
Current state

Outcome

The dedicated selection mode

Outcome

The dedicated selection mode

Outcome

The dedicated selection mode

To solve that problem, I made two key changes.

1. I introduced a new cluster term selection mode that:

  • Increases the space available to explore thousands of terms

  • Gives deeper context behind terms identified by the algorithm

  • Affords deeper interaction including search, sorting, and export

1. I introduced a new cluster term selection mode that:

  • Increases the space available to explore thousands of terms

  • Gives deeper context behind terms identified by the algorithm

  • Affords deeper interaction including search, sorting, and export

1. I introduced a new cluster term selection mode that:

  • Increases the space available to explore thousands of terms

  • Gives deeper context behind terms identified by the algorithm

  • Affords deeper interaction including search, sorting, and export

2. I improved the selection state of clustering to:

  • Emphasize the most relevant information

  • Communicate multi-modal selections

  • Integrate clustering with other parts of the platform, including Search

2. I improved the selection state of clustering to:

  • Emphasize the most relevant information

  • Communicate multi-modal selections

  • Integrate clustering with other parts of the platform, including Search

2. I improved the selection state of clustering to:

  • Emphasize the most relevant information

  • Communicate multi-modal selections

  • Integrate clustering with other parts of the platform, including Search

End solution

Process

Quick research

Process

Quick research

Process

Quick research

Process

Quick research

At the beginning of the project, I knew we were faced with immediate user need, such as a client needing to support a 3 million document database. Despite it being a fast-paced project, I was able to complete several forms of research:

  • Usability tests on the current state (and later, on wireframes)

  • Secondary research via existing user interviews

The Product Lead and I identified some recurring sentiment among legal teams:

"We're unsure about the veracity of cluster terms."

Because of the blackbox nature of machine learning, users are uncertain why certain terms look more relevant than others.

"We're unsure about the veracity of cluster terms."

Because of the blackbox nature of machine learning, users are uncertain why certain terms look more relevant than others.

"We're unsure about the veracity of cluster terms."

Because of the blackbox nature of machine learning, users are uncertain why certain terms look more relevant than others.

"We're unsure about the veracity of cluster terms."

Because of the blackbox nature of machine learning, users are uncertain why certain terms look more relevant than others.

"It's hard to figure out what I'm looking at…if you open it up it's a mess of terms."

Without a structured way to parse cluster terms, users feel overwhelmed.

"It's hard to figure out what I'm looking at…if you open it up it's a mess of terms."

Without a structured way to parse cluster terms, users feel overwhelmed.

"It's hard to figure out what I'm looking at…if you open it up it's a mess of terms."

Without a structured way to parse cluster terms, users feel overwhelmed.

"It's hard to figure out what I'm looking at…if you open it up it's a mess of terms."

Without a structured way to parse cluster terms, users feel overwhelmed.

"I'm using it to figure out what kind of term to do more targeted searches on."

After identifying terms of interest, users use other Everlaw features such as Search to get more in depth.

"I'm using it to figure out what kind of term to do more targeted searches on."

After identifying terms of interest, users use other Everlaw features such as Search to get more in depth.

"I'm using it to figure out what kind of term to do more targeted searches on."

After identifying terms of interest, users use other Everlaw features such as Search to get more in depth.

"I'm using it to figure out what kind of term to do more targeted searches on."

After identifying terms of interest, users use other Everlaw features such as Search to get more in depth.

Process

From flows to wireframes

Process

From flows to wireframes

Process

From flows to wireframes

Process

From flows to wireframes

Research helped me chart a user flow to visualize our assumptions around exploring cluster terms. Here we identified additional value-adds, such as exporting cluster terms. This was a user need not captured in the original project definition, but our interviews had identified a minority of users with ~1000+ terms and preference for off-platform tools.

As a designer, understanding how to balance UX with business and technical constraints is crucial. We knew the level of tooling desired by this user segment would be infeasible to develop. So rather than blow up scope, we proposed the export feature to provide value quickly by enabling users to analyze off-platform.

User flow

Informed by the user flow, I delved into wireframes to rapidly ideate. The goal was to enable investigation of cluster terms at scale, which our research defined to mean:

  • Greater space for users to browse terms

  • Clear indication of which terms have been selected

  • Deeper context to determine which terms are most relevant

Wireframes

Process

Refining to high-fi

Process

Refining to high-fi

Process

Refining to high-fi

Process

Refining to high-fi

To explore these ideas, I led several more usability tests, this time with Figma prototypes. We learned people felt less cognitive load with a dialog, but thought it was tedious to switch back and forth between the visualization and selection state.

It's also critical to my process to solicit feedback early and often. I brought this project to design critiques with other designers, and meetings with company leadership.

Through iteration, with design decisions informed by testing and feedback, I designed a solution that features cluster terms in a non-modal dialog, and an updated selection state to show selected terms and the actions users can take with them.

High-fidelity

Impact

Client response and metrics

Impact

insights gained and sought

Impact

Client response and metrics

Impact

Client response and metrics

We shipped this feature in time for the strategic client’s use on the 3 million document case, and received positive feedback. In addition, the PM and I defined several metrics for collection to continue monitoring the response to the improvement:

  • Click rate for the cluster term dialog to see how often users are relying on analysis via on-platform tools

  • Frequency at which users are exporting terms

☝️

v2 idea

If exporting terms turns out to be a really common, I'd advocate for a v2 to explore how we can enable those workflows (data analysis/visualization of terms) on-platform.

☝️

v2 idea

If exporting terms turns out to be a really common, I'd advocate for a v2 to explore how we can enable those workflows (data analysis/visualization of terms) on-platform.

☝️

v2 idea

If exporting terms turns out to be a really common, I'd advocate for a v2 to explore how we can enable those workflows (data analysis/visualization of terms) on-platform.

☝️

v2 idea

If exporting terms turns out to be a really common, I'd advocate for a v2 to explore how we can enable those workflows (data analysis/visualization of terms) on-platform.

This project was a learning experience; as I learned through user interviews, simply showing users the data and output from AI isn't enough. When designing for complex data visualizations or system outputs, it's essential to build trust into the system and make that data actionable.