AI Concept

Designing adaptive AI summaries for the distracted state

Designing adaptive AI summaries for the distracted state

Designing adaptive AI summaries for the distracted state

0 to 1

Web

B2B

Enterprise

My Role

Lead UX Designer

Responsibilities

Early-stage concepting
Prototyping
Usability Testing

Team

I collaborated closely with a UX director, technical program manager, and the executive team.

Overview

Overview

Early-stage concept and experimentation for an AI-powered executive decision-making platform.

Early-stage concept and experimentation for an AI-powered executive decision-making platform.

Early-stage concept and experimentation for an AI-powered executive decision-making platform.

I was the lead UX designer responsible for leading the early-stage concept and experimentation for an AI-powered executive decision-making platform. The goal was to help executives quickly digest research insights and key metrics across multiple data sources through AI-generated summaries.

When internal priorities shifted, I continued building core concepts independently to validate fundamental assumptions about AI-generated summaries. Here are some areas that I focused on throughout the project.

I was the lead UX designer responsible for leading the early-stage concept and experimentation for an AI-powered executive decision-making platform. The goal was to help executives quickly digest research insights and key metrics across multiple data sources through AI-generated summaries.

When internal priorities shifted, I continued building core concepts independently to validate fundamental assumptions about AI-generated summaries. Here are some areas that I focused on throughout the project.

I was the lead UX designer responsible for leading the early-stage concept and experimentation for an AI-powered executive decision-making platform. The goal was to help executives quickly digest research insights and key metrics across multiple data sources through AI-generated summaries.

When internal priorities shifted, I continued building core concepts independently to validate fundamental assumptions about AI-generated summaries. Here are some areas that I focused on throughout the project.

01

Making it work with what already existed

Making it work with what already existed

Making it work with what already existed

02

Solving the executive vs. analyst problem

Solving the executive vs. analyst problem

Solving the executive vs. analyst problem

03

Seeing if they need to dig deeper later

Seeing if they need to dig deeper later

Seeing if they need to dig deeper later

Kickoff

Kickoff

No roadmap.
No requirements.

No roadmap.
No requirements.

No roadmap.
No requirements.

There wasn't a brief, roadmap, or set requirements. Just a Slack message to review a WIP demo before jumping onto the client call. So I started asking questions whenever I could during calls, sitting in on product meetings, and digging into the demo to figure out what we were actually building. Over a few weeks of wireframing and weekly design reviews, the real scope and constraints slowly came together.

There wasn't a brief, roadmap, or set requirements. Just a Slack message to review a WIP demo before jumping onto the client call. So I started asking questions whenever I could during calls, sitting in on product meetings, and digging into the demo to figure out what we were actually building. Over a few weeks of wireframing and weekly design reviews, the real scope and constraints slowly came together.

There wasn't a brief, roadmap, or set requirements. Just a Slack message to review a WIP demo before jumping onto the client call. So I started asking questions whenever I could during calls, sitting in on product meetings, and digging into the demo to figure out what we were actually building. Over a few weeks of wireframing and weekly design reviews, the real scope and constraints slowly came together.

Challenge

Challenge

“Executives don’t have time to look through all this. Think of it like a presidential briefing.”

“Executives don’t have time to look through all this. Think of it like a presidential briefing.”

The real brief came during a dashboard review, when a client said this. At the time, I thought simplifying navigation and surfacing key metrics in a dashboard was the answer. But that comment reframed everything. I wasn’t just designing for usability, I was designing for cognitive overload. It became our north star.

The real brief came during a dashboard review, when a client said this. At the time, I thought simplifying navigation and surfacing key metrics in a dashboard was the answer. But that comment reframed everything. I wasn’t just designing for usability, I was designing for cognitive overload. It became our north star.

Problem

Skimming wasn't optional. It's for survival. Users were primarily skimming to filter and triage.

Skimming wasn't optional. It's for survival. Users were primarily skimming to filter and triage.

Top reasons for skimming based on multi-selection responses from 17 participants.

Top reasons for skimming based on multi-selection responses from 17 participants.

Finding a specific piece of information

Finding a specific piece of information

76.50%

76.50%

Deciding if content is relevant

Deciding if content is relevant

64.70%

64.70%

Seeing if they need to dig deeper later

Seeing if they need to dig deeper later

47.10%

47.10%

Constraints

Constraints

01

Information overload was the norm.

I designed a component that could sit on top of existing documents and dashboards without breaking anything the dev team was already building

02

Skimming wasn’t optional. It was survival.

Same data, different needs. Executives wanted the high-level gist, while research strategists needed more granular details. So I built adaptive summaries that could switch between both.

03

There was no time for a learning curve.

I had to figure out ways to surface AI insights that would actually stick.

Ideation

Simplifying everything into a single sentence.

Simplifying everything into a single sentence.

After prompting AI to simplify dashboard wireframes, it kept shrinking to one sentence. That felt more powerful than expected.

After prompting AI to simplify dashboard wireframes, it kept shrinking to one sentence. That felt more powerful than expected.

Identifying the Gaps

One-size summaries doesn't fit all.

One-size summaries doesn't fit all.

As I explored existing AI-generated summaries, I noticed a pattern

Some were ultra-condensed into a single sentence, but I couldn't trace where information was being pulled from.

Others swung the opposite way where it was too long, hard to scan under pressure.

There didn’t seem to be a middle ground.

Nothing let users adjust the summary based on how much attention they had or what their role was. That became the core of the concept: design a summary that adapts to you—not the other way around.

As I explored existing AI-generated summaries, I noticed a pattern

Some were ultra-condensed into a single sentence, but I couldn't trace where information was being pulled from.

Others swung the opposite way where it was too long, hard to scan under pressure.

There didn’t seem to be a middle ground.

Nothing let users adjust the summary based on how much attention they had or what their role was. That became the core of the concept: design a summary that adapts to you—not the other way around.

Exploring interaction models

Exploring interaction models

I tested several approaches for adjusting summary length:

  • Vertical carousel (scrolling up/down)

  • Horizontal swiping

  • Tapping through levels

Despite using familiar patterns, it felt conceptually off. This interaction made the summary feel like flipping through pages, scrolling through a vertical feed, or reading a long scroll.

These interactions implied switching between different summaries. But that wasn't the experience I wanted to create. I wanted users to shape the same summary, not switch between versions.

I tested several approaches for adjusting summary length:

  • Vertical carousel (scrolling up/down)

  • Horizontal swiping

  • Tapping through levels

Despite using familiar patterns, it felt conceptually off. This interaction made the summary feel like flipping through pages, scrolling through a vertical feed, or reading a long scroll.

These interactions implied switching between different summaries. But that wasn't the experience I wanted to create. I wanted users to shape the same summary, not switch between versions.

Vertical scrolling

Gesture doesn't support long summary lengths that requires vertical scrolling for overflowing content

Horizontal carousel swiping

Limited visibility for all the supported summary lengths at first glance

Tap or clicking drilldown

Difficult to navigate in a non-linear order
Added too much friction

Landing on the slider

Landing on the slider

A slider mapped better to the mental model of modification like adjusting volume or brightness.

A slider mapped better to the mental model of modification like adjusting volume or brightness.

01

Tactile and fidget-friendly

02

Low cognitive effort

03

Higher affordance with minimal effort by design

Exploring visual metaphors

Exploring visual metaphors

An animated orb metaphor was my first direction—scaling with info density using WebGL shaders. But testing showed friction and visual fatigue. I scrapped it for simplicity: fewer animations, better legibility, and faster scanning.

An animated orb metaphor was my first direction—scaling with info density using WebGL shaders. But testing showed friction and visual fatigue. I scrapped it for simplicity: fewer animations, better legibility, and faster scanning.

Defining the summary length

Defining the summary length

How short is too short?

How short is too short?

To define the shortest readable length, I synthesized research from:

  • Nielsen Norman Group findings on reading behavior

  • Road signage standards for variable message signs (designed to be readable while moving)

The lower bound settled around 40–50 characters. Just enough to convey meaning without demanding focus. This aligned perfectly with the use case: critical signals in motion.

To define the shortest readable length, I synthesized research from:

  • Nielsen Norman Group findings on reading behavior

  • Road signage standards for variable message signs (designed to be readable while moving)

The lower bound settled around 40–50 characters. Just enough to convey meaning without demanding focus. This aligned perfectly with the use case: critical signals in motion.

Calibrating control levels

Calibrating control levels

How many lengths?

How many lengths?

I started with three levels (short, medium, long) but it felt too binary like choosing between presets. I introduced five levels to create smoother interpolation and flexibility.

I tuned each level's character count using a power law curve, then validated through distracted-state testing. If it didn't feel readable or approachable while multitasking, I adjusted again.

I started with three levels (short, medium, long) but it felt too binary like choosing between presets. I introduced five levels to create smoother interpolation and flexibility.

I tuned each level's character count using a power law curve, then validated through distracted-state testing. If it didn't feel readable or approachable while multitasking, I adjusted again.

Benchmarking percieved effort

Is this skimmable right now?

Is this skimmable right now?

Instead of standard usability metrics, I used a more nuanced evaluation:

  • Would I want to skim this in a distracted state?

  • Does this feel heavy or approachable?

This question guided every decision from interaction design to language length to visual affordance.

Instead of standard usability metrics, I used a more nuanced evaluation:

  • Would I want to skim this in a distracted state?

  • Does this feel heavy or approachable?

This question guided every decision from interaction design to language length to visual affordance.

Simulating skimming shorter versions for the distracted state
Simulating skimming longer versions in controlled environments

Defining the starting length

Low, medium, or high?

Low, medium, or high?

I realized that when I started on the low or highest level, it made the slider feel like there’s less affordance. By placing the user in the middle range, I felt that it gave more agency to either go shorter or longer from there and required less work.

I realized that when I started on the low or highest level, it made the slider feel like there’s less affordance. By placing the user in the middle range, I felt that it gave more agency to either go shorter or longer from there and required less work.

Pivot

Moving into AI-coded prototyping environments

Moving into AI-coded prototyping environments

Because Figma's prototypes are limited to linear, step-by-step transitions, they couldn't accurately simulate the fluid, single-swipe adjustments critical for this slider. To ensure our usability testing reflected the optimal user experience, I used AI-coding prototypes. This also allowed me to save significant time, as I could easily fork multiple variations for A/B testing or make global changes without the complex "stitching" of countless individual screens that Figma would require.

Because Figma's prototypes are limited to linear, step-by-step transitions, they couldn't accurately simulate the fluid, single-swipe adjustments critical for this slider. To ensure our usability testing reflected the optimal user experience, I used AI-coding prototypes. This also allowed me to save significant time, as I could easily fork multiple variations for A/B testing or make global changes without the complex "stitching" of countless individual screens that Figma would require.

Usability testing

A/B testing hypothesis around slider interactions, perceived trust, and skimming behavior

A/B testing hypothesis around slider interactions, perceived trust, and skimming behavior

I conducted usability testing with three separate cohorts. The first two cohorts were used for a mini A/B test to evaluate the impact of a key design decision: whether to provide users with direct access to "sources" and "topics" controls within the slider. The third cohort then tested a revised design that incorporated the key learnings and changes identified from the first two rounds.

I conducted usability testing with three separate cohorts. The first two cohorts were used for a mini A/B test to evaluate the impact of a key design decision: whether to provide users with direct access to "sources" and "topics" controls within the slider. The third cohort then tested a revised design that incorporated the key learnings and changes identified from the first two rounds.

3

3

Cohorts

Cohorts

2

2

Rounds of post-feedback revisions

Rounds of post-feedback revisions

Hypothesis A

Users would rely heavily on the slider if there were no topics or sources available

Users would rely heavily on the slider if there were no topics or sources available

Hypothesis B

Users would trust the topics and sources B variant more

Users would trust the topics and sources B variant more

Hypothesis C

Users would rely on topics and sources heavily

Users would rely on topics and sources heavily

Design A - Control
Design B

A/B TEST INSIGHT

Users felt lost and confused without extra contextual hints for both design variants

Users felt lost and confused without extra contextual hints for both design variants

Observation

One participant kept reading the title of the overlay multiple times while looking for other context clues.

Another participant looked instantly confused and frustrated when they opened the summary. They mentioned “I have no idea what I’m looking at."

“Hyper summary, hyper summary, hyper summary…"
Participant
"I have no idea what I'm looking at…"
Participant

Assumption

Participants were scanning and looking for information scents to understand what the document is about

Iteration

Added an eyebrow text with the source title to help guide users who may have low familiarity with the content.

Hinting at what the summary or source with the source title while adding minimal text.

A/B TEST INSIGHT

Users rarely used the slider interaction for both design variants

Users rarely used the slider interaction for both design variants

Observation

The majority of both cohorts didn’t use the slider at all, even though it was mentioned in the task prompt.

Users tended to focus on reading the middle length summarized text and moved onto to the next task.

Assumption

Slider lacked visual affordance since the slider lacked an active track and the slider head was small in size.

Users needed more relevance in high-stakes situation and guidance when it came to the slider level labels.

Iteration

Added an active track to the slider

Increased the height of the bar

Increased the size of the slider head to improve affordance.

Removed the instruction to use the slider in the task prompt to test standalone engagement in Cohort C

A/B TEST INSIGHT

Under time pressure, users skipped sources and started with topics

Under time pressure, users skipped sources and started with topics

Iteration B - Bottom sheet tab
Iteration C - Accordion

Observation

With only a few minutes to review the summary, most users skipped the sources and went to topics first, often before adjusting the summary length.

Despite bypassing sources, they still rated the summary as trustworthy.

Users who customized topics often struggled with order and visibility, often giving up.

Assumption

Users were using the topics first as a way to filter through the summary

Placing the topics closer to the summarized text may align better with their scanning behavior.

Users had a hard time seeing topics tags since it was small and easy to miss even though I had both the tags and the tab next to the slider controls.

Iteration

Moved topics out of the tab view and placed it as a accordion that sits right below the summarized text area

Added visibility of what topics are hidden due to compression

A/B TEST INSIGHT

Users preferred the medium length for perceived comprehension

Users preferred the medium length for perceived comprehension

Observation

Most participants preferred the medium or medium-high summary lengths, finding them easiest to understand.

Assumption

Users prefer a balance of brevity and detail, and can clearly distinguish between the 5 length levels.

Iteration

No changes made. Kept the medium length as the default position.

Testing design revisions

Dialing in revisions

Dialing in revisions

I realized that when I started on the low or highest level, it made the slider feel like there’s less affordance. By placing the user in the middle range, I felt that it gave more agency to either go shorter or logner from there and required less work.

I realized that when I started on the low or highest level, it made the slider feel like there’s less affordance. By placing the user in the middle range, I felt that it gave more agency to either go shorter or logner from there and required less work.

DESIGN REVISION FEEDBACK

Users lost their place when expanding the topics accordion

Users lost their place when expanding the topics accordion

Iteration C - Accordion
Iteration D - Horizontal Carousel

Observation

Expanding the topics pushed the summarized text out of view, leaving some users disoriented and switching back and forth between the two.

"I wish it was more interrelated with each other."
Participant

Assumption

Users need a quick way to connect topics with the summarized text.

Iteration

Replaced the topics accordion with a horizontal carousel that keeps topics visible alongside the summarized text.

Added an interaction that highlights the related section of the text when a topic card is selected.

Wrapping it all up

Wrapping it all up

Impact

Impact

Supporting deep dives around specific points

60%

60%

60%

of cohort 3 participants found overlays helpful for building trust and ensuring judgment accuracy.

of cohort 3 participants found overlays helpful for building trust and ensuring judgment accuracy.

Ramping up quickly on key topics

Ramping up quickly on key topics

42.5%

42.5%

42.5%

increase in cohort 3 participants who felt informed within 1 minute. Outperforming both the control (v1) and tab-based designs (v2).

increase in cohort 3 participants who felt informed within 1 minute. Outperforming both the control (v1) and tab-based designs (v2).

Building trust and traceability with inline sources

Building trust and traceability with inline sources

75%

75%

75%

of cohort 3 participants rated the summary a 4 or 5 . The highest trust score across all groups, with improved source discoverability correlating to higher trust.

of cohort 3 participants rated the summary a 4 or 5 . The highest trust score across all groups, with improved source discoverability correlating to higher trust.

Improved core feature slider engagement

Improved core feature slider engagement

3X

3X

3X

in participant engagement with the slider compared to earlier designs.

in participant engagement with the slider compared to earlier designs.

Reduced cognitive load

Reduced cognitive load

50%

50%

50%

of cohort 3 participants rated the summary as low effort, compared to just 20% in the control group.

of cohort 3 participants rated the summary as low effort, compared to just 20% in the control group.

Project reflection

Project reflection

01

01

01

Familiarity with content material

Familiarity with content material

For users who are unfamiliar with the source material, contextual tools like "topics" and source title become essential. These features help them build a foundational understanding of the content, acting as a mental model for navigating new information and information scent.

For users who are unfamiliar with the source material, contextual tools like "topics" and source title become essential. These features help them build a foundational understanding of the content, acting as a mental model for navigating new information and information scent.

02

02

02

Perceived effort

Perceived effort

A user's perceived effort is highly dependent on their context. A summary that seems easy to read in a focused state can feel overwhelming and too long when a user is distracted or in a hurry.

A user's perceived effort is highly dependent on their context. A summary that seems easy to read in a focused state can feel overwhelming and too long when a user is distracted or in a hurry.

03

03

03

Provide enough information scent

Provide enough information scent

Since users may not start at the beginning of a document, every entry point, including the summary, must provide enough context to clearly identify its source and purpose.

Since users may not start at the beginning of a document, every entry point, including the summary, must provide enough context to clearly identify its source and purpose.