top of page

Amazon Return Internship 

Group 1739327776.png

UX Designer | September 2023 - December 2023

I integrated generative AI into Amazon Seller Central’s Associate-facing product, used by 10,000 Support Associates to address cases by 2 million active Amazon Sellers. Collaborating directly with science and tech stakeholders, I designed a time-sensitive new LLM-powered case summary feature, consolidating critical info such as seller intent and primary identifiers including ASINs, to reduce handle time and improve Associate accuracy metrics.

Designs were grounded in 3 rounds of
moderated user research I conducted with 18+ Associates worldwide, developing IA and user workflows, and delivering a series of Figma high-fidelity mockups to the product team for launch in January 2024.

Role: Project Owner, Lead UX Designer 
Stakeholders: Product Manager, 2 Software Development Engineers,
UX Researcher, 2 Senior UX Designers, Data Scientist

Timeline: 14 weeks, Project launch in January 2024

final version (2).png
Figma
User Research
Usability Testing
User Personas, Workflows, Journeys
Information Architecture
Competitive Research
Wireframing
Interactive Prototyping
Stakeholder Management
Problem Breakdown

1. Problem Breakdown

There are currently 1.9 million active Amazon sellers using Seller Central to run their businesses every day. Sellers create cases when encountering issues with selling on Amazon which are regularly addressed by around 10,000 Amazon Associates.

Group (11).png

1.9 Million
Active Amazon Sellers

Vector (6).png

9324
Active Amazon Associates

Group 1739327763.png

2.58 years
Average Associate Tenure

Support Associates address cases through the internal platform Paragon which consists of more than 17 different widgets. 
Associates currently diagnose the seller issue through a tedious multi-step process:

Review the initial contact to identify the case intent

Look through every correspondence entry (there can be around 40 for a multiple reopens case) including case annotations and subsequent contacts to understand the current status and resolutions

Access the case and customer info to copy the seller and case ID for troubleshooting, and checking the account health, status, owner, and more

Check the related items widget to access completed paragon workflows and tickets

Action on the case: Replying to the seller requesting information, transferring the case to the correct department, creating a ticket or a Paragon workflow, or pulling an Andon cord if they need further help

Frame 1739327759.png

Currently, Associates dedicate considerable amounts of time and effort to diagnose case intent (21% of Associate handle time) by scanning through numerous correspondence entries.

The
accuracy and efficiency of the case-handling process is hindered by information overload across all these different Paragon widgets and the lack of easy referencing ability for critical information.

Consolidating the seller issue, primary identifiers, and relevant correspondence, within a new LLM-powered Issue Summary widget, will improve the ease of diagnosing the seller issue without needing to individually review each contact. 

This new Paragon feature will
reduce handle time by providing Associates with a clear and concise summary that they can continuously reference as well as minimize human error, improving Associate reliability and low adoption rates.
 

Group 1739327777.png

The primary problem above can be broken down into 3 pain points aligning with design goals to address them. 

Group 1739327786.png

1. Project Logistics

Project Role & Stakeholders

I served as the project owner and primary designer working directly with multiple stakeholders including the product and tech team, coordinating weekly reviews to recieve feedback on design mockups, IA, and research findings.

  • Vandana Soundera Raj (Product Manager)

  • Pavithra Aravindan (Associate-facing UX designer)

  • Allison Kapps (Intern Mentor: UX Designer)

  • Minwoo Kim (Intern Manager: UX Designer)

  • Larcombe Teichgraeber (UX Researcher)

  • Tech Team (2 Software Development Engineers) 

Project Scope

The project experiment is anticipated to launch in January 2024 with science and tech work being currently completed. The project
will roll out to certain groups of Associates after socializing the concept with supervisors to ensure Associates are not overwhelmed. 

Project Success Metrics 

Adoption Rate

Outcome Accuracy

Average Case Handling Time

To evaluate the Issue Summary project, we will measure these 3 metrics. Establishing baseline metrics before the project's launch, we will assess these indicators post-experiment and post-launch to quantify the impact on associate performance. This iterative process fosters continuous improvement.

Design Process (Reference Links) 

Context & Impact 
Current Experience
Problem & Solution
Project Stakeholders
Project Scope

UI Research 
Competitor Research 
Information Architecture 
Moderated User Research 
Key Research Insights

Wireframes & Sketches 
Initial Design Explorations 
Stakeholder Alignment
Top Design Iterations 

Stakeholder Reviews 
Usability (A/B) Testing 
Key Research Insights 

Design Research 
Design Iterations 
Unmoderated User Research
Key Research Insights 
Final Design Walkthrough  

Group 1739327789.png
Group 1739327754.png

2. Research & Exploration 

Understanding existing problem space 

To understand the existing research and work, I met with stakeholders numerous times to analyze the Associate Northstar future visions, Associate AXO framework (new standards for improving the Associate experience), and research reports from extensive studies done worldwide (by UX researchers and PMs).

Frame 1739327760.png

UI Research (Dribbble, Pinterest, Google)

Before designing, I wanted to explore existing conventional UI patterns used for summaries within complex platforms, to create an engaging and more modern experience. I analyzed dashboard summaries, tab summaries, carousel designs, side panels, and card designs, evaluating their pros and cons and use cases. Some principles and insights that stood out to consider are specified below.

Group 1739327778.png

Dashboard Summary

  • Breakdown main elements

  • Saving/ pinning components

  • Strong separations between sections

  • Main metrics scrollable at the top, engaging users


Tab Summary Designs

  • Less screen territory and info overload

  • Break steps/ information into simple processes


Carousel Designs

  • Show completion progress at the top

Competitor Research (Microsoft Learn, Cresta AI, CCAI, Talkdesk)

Next, I studied several competitor platforms to find effective elements to incorporate. One example includes Microsoft’s Customer Service Summary dashboards which consistently use engaging and purposeful graphics and visuals, provide dashboard customization for agents, and feature insights and recommended topics at the top of the screen for the Associate. Wheras smaller companies like Cresta AI provide AI generated summaries based on conversations, which are divided into 3 clear, distinct categories: reason, resolution, result.

Group 1739327780.png

Microsoft Learn: Customer Service Summary Dashboards

 

  • Graphics and visuals, engaging and purposeful

  • Customization for agents

  • Tabs to navigate through the dashboard

  • Rigid but organized structure with distinct sections

  • Top insights discovered included at the top

  • Recommended topics for the associate to address popular cases

Information Architecture

After getting inspiration, I listed all the potential components for the Issue Summary, reorganized into a hierarchy to consider the overall widget layout. I used this diagram to later identify which features would be crucial for the new design and which were future possibilities such as showcasing related cases or a transcript summary.

Group 1739327782.png

Moderated Baseline User Research

I began with conducting moderated user research with 5 Associates and supervisors to inform Issue Summary designs, recruiting primarily tenured associates with 2+ skills (such as FBA, IIDP) and 2+ channels.

Research Goals 

  • Determine the most critical identifiers and case/customer attributes for handling a case

  • Understand the current process for identifying seller intent to determine the location of the widget

  • The benefit of potential new features to the summary (such as CIM messages, status tags, and related open cases)

  • Understand how review tactics vary by case state (first contact, reopen, transfer, PMA) if at all, to determine if content should vary accordingly 

image 150.png

5 Active
Amazon Associates

image 151.png

1 hour 
Sessions

Frame 1739327761.png

Recorded & Moderated via Chime & Figma

2+
Skills (FBA, IIDP, FFT, etc)

2+
Channels (Email/ Phone)

4/5
Tenured (2+ years) Associates

Research Activities

Goal: Determine what information Associates find most valuable in their ideal summary. Going through every case attribute (such as case ID, case status, business phone), Associates ranked what they would find “critical,” “nice to have,” “neutral,” or “not necessary” and explained their rationale. After each piece of information, Associates were asked to compare the priority across case states (such as first contact and multiple reopens).

V3 activity (updated after 1st session).png

In this card sorting activity, the goal is to understand how much manual effort is currently required to find specific pieces of information (such as ASINs or case intent) to determine where LLM automation would be the most beneficial.

Actvity 2_ Final (unlocked).png

Research Analysis

After receiving Associate insights, I transcribed and categorized the research by priority of features across all participants to identify trends using sticky note clustering and numerous tables. 

Group 1739327787.png

Baseline Research Insights

Next, I wrote a 12-page report summarizing all of the relevant trends and design recommendations including anecdotes from Associates. The report cannot be disclosed due to personal information. Here are the primary 4 insights. 

Insight 1 | Centralizing Key Identifiers

Frame 1739327762.png

5/5

Associates stated that ASINs and other identifiers are essential in addressing the SP issue to access troubleshooting tools, locate FNSKUs, and for Paragon workflows. Currently, the primary pain point in locating ASINs is parsing through numerous correspondence entries and annotations.
 

  • Identifiers change and evolve throughout the case which makes them difficult to search for.

  • Associates also regularly encounter multiple identifiers and must ensure every seller intent is adequately addressed so the case is not reopened.

Design Recommendations

  • Centralize relevant ASINS and critical identifiers at the top of the summary

  • Pre-populated and constantly updated from correspondence

  • Copy/paste functionality for easy referencing

Insight 2 | Summary Content & Format 

5/5

All Associates take their own notes while addressing cases (using external notes app, plugin, or clipboard), and the common pieces of information they include:

  • Seller ID, Case ID

  • Relevant help links

  • Relevant tickets and SOPs

  • Summary of what has been done

Associates frequently encounter cases with multiple intents where it’s difficult to monitor investigations

“Always need seller ID, have to scroll down and up several times. Have to copy and paste it in my annotations. Case ID. Summary of the issue. Identifiers IDs, Batch ID... Steps taken for what I did. Any link from any tools that I use.” (A1).

Design Recommendations

  • Validates displaying multiple case intents and showcasing the latest main events

  • Explore how to clearly differentiate each intent and its resolution with colors and smaller chunks of text

  • Explore different positions for the Summary Widget including (1) side-bar panel (summary is always visible to reference), (2) floating ingress (opens as a modal), or (3) top of the page (first item seen and can be pinned)

Insight 3 | Categorizing Case Information  

Group 1739327739.png

Critical Information
All 5 Associates identified the 10 features (Case ID, Seller ID, Queue, Status, Owner, Account Health, Account Status, Marketplace, Business Phone, SIV Status) as definitively “critical” and “nice to have” due to their necessity to be continually referenced during the case handling process to access tools, create tickets, distinguish between cases, and launch workflows.

Not Critical Information
7 features (Tenant ID, Created Date, Reason Code, Severity, Action by date, Services, Phone Support) were unanimously identified by Associates as not necessary to immediately see in the Issue Summary since Associates are unaware of their functionality, have no preference, or have never heard/used them before.

Uncertain Information
5 features (Brand Registry, Staging, 2SV Enabled, Nickname, Type) have conflicting opinions across different Associates, so require further research to determine value or should be included with specific design changes such as adding status tags for brands and links to the brand registry tools.

Design Recommendations

  • Highlight critical attributes in the summary so that they are glance-able and easy to access

  • Integrate the case and customer widget to minimize information overload

  • To prevent losing information before more extensive research, make the items visible in a collapsed option

Insight 4 | Future Additional Summary Features 

Group 1739327751.png
Group 1739327752.png
Group 1739327753.png

4/5

Associates believe CIM messages contain important information yet none frequently use them because the content is rarely relevant to their case and the current UI format does not keep attention


“Really easy for us to just ignore this part because we don't really like pay attention to it. I can say that I have used this section here once in the last three or four months because it's not like user-friendly.” (A3).


Associates are aware of the status tags (i.e., Frustrated, 911 Urgent), but stated they always ignore them since all cases are approached in the same unbiased manner, regardless of severity and sentiment.


“I'm always going to provide paraphrasing and acknowledge their frustration even if they don't have the tag. The process is going to be the same for all of the cases” (A4).

Associates commonly refer to past resolved cases to follow successful resolutions. No current way for Associates the process is correct, allowing for repeated mistakes and increasing AHT.

5/5

5/5

Design Recommendations

  • Alerts section with relevant CIMs, case statuses, and tags that Associates can quickly scroll through

  • Surfacing recommended related cases that are more trustworthy since approved as successful resolution examples

  • Considered in the future to address the Action and Respond case-handling stages

3. Design & Iteration 

Sketching & Lo-Fi Mockups 

In parallel with conducting research, I began sketching and creating multiple rounds of iterations on approaches for the Issue Summary. Here I focus on some of the most important designs and decisions.

Group 1739327754.png

Initial Design Explorations: Product Doc Review

In the first product doc review with leadership (prior to research findings), my design consisted of the LLM-generated content displayed in a paragraph with the 2 widgets (case and customer info and correspondence) collapsed within it.

Group 1739327790.png

Post Research: Stakeholder Alignment

The new research findings guided what information to include, the location of the Issue Summary and the visual hierarchy of the content, labeling, visuals and colors, etc. After reviewing the research report multiple times with stakeholders, I was able to align on the priority of each new feature.

Group 1739327789.png

Top Design Iterations

From research, I was able to derive a 3-step associate case handling flow which will be applied in the following designs.

Understanding high-level case details

Read what occured in the case: LLM generated summary 

Use key identifiers to troubleshoot or learn more about the case

Step 1 | Understanding high-level case details

V1_ 2 column_ general info on the side, correspondence outside.png

Associates can now quickly see:

  • Shows top-10 most relevant case details so that they are glance-able and easy to access

  • Organized by priority starting with the seller and case ID (used 80% of cases)

  • Visual breakdown (colors to recognize status) and copy/ paste functionality to minimize information overload

  • Ability to “show more” (view less important case details) and easily spoof Seller Account

Other Iterations 

Group 1739327791.png

Version 2: Horizontal Card Design

  • Easy separation of information (case vs customer) with icons, minimize information overload

  • Takes more screen territory

V5_ Tab design.png

Version 4: Tab Design

  • Separation between stages (seller intent vs looking case info)

  • Less screen territory since content is separated 

  • Less discoverable, more click time

V4.5_ Horizontal design, info on bottom.png

Version 3: Case Info integrated with summary

  • Less screen territory (no separation between text)

  • Clear linear flow to follow (top to bottom)

  • Easy integration of links ​

V2_ 2 column_ Case info and identifiers on the side, correspondence within (1).png

Version 5: Side-bar case info

  • Easy to reference information

  • Takes more vertical screen territory

  • No easy linear flow to follow

Step 2 | Reading LLM Generated Summary

Associates next read the LLM summary of the issue:

  • Summary blurb (3-4 sentences, 100-150 words) broken into 2 simple sections: “case intent” and “issue context”

  • Breaking down chunks of text, Associate easily knows primary issue and relevant correspondence/ main events

  • Multiple intents could be displayed with a number indicator and separation between issues

Group 1739327792 (1).png

Other Iterations 

Frame 1739327794 (1).png

Version 1: Main actions broken down 

  • Summary is always 3-4 sentences so not necessary to breakdown into every sentence

  • Date is not relevant to resolve the case

Version 2: Stages to the seller issue

  • Easy to understand steps in the process (current, past, future) but not tech-feasible

  • Table format to clearly show relevant correspondence but requires more work

Version 3: Multiple intents numbered

  • Not scaleable in the future and too much emphasis placed on numbers over content

  • Text is broken down too much

Step 3 | Reviewing key identifiers to troubleshoot

Associates next troubleshoot the issue:

  • Easily access relevant ASINs and key identifiers as they change and update throughout the case (don’t need to scroll through correspondence entries: primary pain point)

  • Click to copy functionality for easy troubleshooting with 3 methods: Copy individual ASINs, Copy All, hyperlink

Group 1739327794.png

Other Formats

Vertical sidebar design (pros and cons):

  • Continuously reference critical information

  • Linear flow in information presentation

  • Easy access to “transfer, reply” actions

  • Less readability, correspondence is squished

  • Only linear reading flow (might miss lower info)


Horizontal top of screen design:

  • First item that is surfaced and always seen before addressing case

  • More conventional with existing UI

  • Can be pinned to the top of the screen

V2_ Widget format.png

4. Evaluating Designs

Stakeholder Reviews & Co-Designs Feedback

After receiving feedback from every stakeholder and reviewing designs at UX team syncs, I made changes for the next
Product Doc review with L8 leadership.


Changes are shown below including specifically editing color and font thickness to increase contrast for accessibility, creating hierarchy within the text by adding blue links to listing pages, bold texting to emphasize keywords, and splitting Key Actions taken into bullets to improve skimming and readability.

Group 1739327796.png

Usability Testing Moderated Research

I conducted A/B testing with 6 tenured Associates to evaluate the top 2 proposed concepts. The research focused on understanding the usefulness and completeness of the summary information and the following goals. 

Research Goals

  • Utility of Issue Summary Info; How easy it is to diagnose intent

  • Ideal information placement aligning with the case handling process by testing 2 differing layouts

  • Test readability and discoverability of new feature placement

  • Understand Associates' perceptions towards trust and verification of the summary content. Observe behavior around accessing correspondence and why.

image 150.png

6 Active
Amazon Associates

image 151.png

1 hour 
Sessions

Frame 1739327761.png

Recorded & Moderated via Chime & Figma

2+
Skills (FBA, IIDP, FFT, etc)

2+
Channels (Email/ Phone)

5/5
Tenured (2+ years) Associates

Concept A: Side Panel Design 

The first concept includes a side panel design with general case information, the main summary content in the center, and key identifiers below it, followed by case notes. The scenario text displayed in these prototypes is also LLM generated, taken from initial science experiments, to ensure the accuracy of tasks and a realistic experience for Associates.

Concept B: Card Design 

The second concept was a card design dividing general case information and placing it underneath the LLM summary. It also experiments with a different breakdown of summary text and a different case notes interaction.

Usability Testing Research Findings 

Insight 1 | Prototype Comparison

The majority of Associates prefer Concept A because of its ease of comprehension, improving accessibility of case information, easy referencing of identifiers, overall user-friendliness, and structure. 

“I think is really great. With aged cases, we kinda have a hard time to understand the case. We have to go through several correspondences through several annotations. I think it will be way easier to understand.” A1


“My overall impression is that it is a lot nicer than the version we have now. The information, it's a lot more accessible, it's really nice. The summary of the actions taking, it's really helpful. The second thing that I see it is really good is the option of copy all the ASINs. So you don't have to copy all them manually... the cases will be a lot easier to handle with this interface.” A3


“I think with this kind of information, it will save a lot of time researching and it will save at least five minutes.” A5

Reasons Associates did not prefer Concept B: the text, details, and visuals were overwhelming and intimidating, making it more difficult for them to easily find content.

“The overall layout, we have the information exactly the same as the previous one, but I dunno, it's not as user friendly... The graphics, it's like more shocking to the user. The other one is a lot more friendly compared to this one. It seems like it is a lot more information than it actually is so it kind of intimidates the user.” A3

Prototype A was ranked higher for every EASE metric 
(ease of diagnosis, quickly interpret information, overall satisfaction, clarity, how often used, and confidence)

Screenshot 2023-12-05 at 12.43 1.png

Insight 2 | Case Attributes

Case and Customer Information
Associates were aligned with the chosen hidden fields and located the new “show more” link.

Associates preferred no separation between the seller and case info since the most useful fields should be surfaced first.

Group 1739327797.png

6/6

Case Notes Feature

Summary-specific case notes were continuously confused with annotations and believed to be the same.
Pros: Beneficial to serve as a (1) feedback mechanism to verify key events, (2) scrapbook of messages for future Associates, (3) supplement missing info, (4) concise summary of annotations
Cons: Overlaps with established annotations and requires more handle time to create and look through.

4/6

6/6

4/6

3/6

Insight 3 | Key Actions Taken 

Frame 1739327799.png

6/6

Associates unanimously identified case intent and identifiers as the most important sections. Associates required individual Copy, Copy All, and hyperlink functionality.

Associates independently mentioned that SOPs, workflows, tickets and tools used are items they are missing and their reason to go to correspondence.

6/6

Insight 4 | Trust and Verification

6/6

All Associates still refer to correspondence. Relating to lack of trust, top reasons include: 
(1) to ensure there is no missing information
(2) verify the key actions taken content
(3) in case past Associates entered incorrect information

Trust regarding the Issue Summary was divided.

Associates believed the info was manually generated and lost trust when informed it was machine-generated

Associates knew it was AI-generated, but still needed to verify the “Key Actions Taken” content

Associates learned it was AI-generated, which increased trust because it would minimize human error probability

2/6

3/6

1/6

Based on this research, our goal changed from immediately increasing trust to make it easier for Associates to verify the information (since they will always verify regardless).

5. Additional Feature

Sources: Design Research

We considered options to ease verification such as showing confidence scores or "Approved by Associates" tags. In the end, we decided on integrating sources for the key actions taken.

I started by exploring how platforms today are showcasing sources for AI-generated content such as Google’s Bard with highlighting and previewing sources, ChatGPT with in-line citations, messaging platforms with thumbnails and blurbs from the source, and Bing surfacing the full source link.

Bard
Click highlighted text
Hover to see source link

Web ChatGPT
Inline sources links [1]

Messenger & WhatsApp
Thumbnail with title
Blurb excerpt, quick links

Bing
Excerpts of news, source, date
Full source link

Sources: Design Explorations

Drawing from the research, I explored different methods of displaying sources such as source previews. in line citations, and lists.

Concept 1 | Hover to see source preview 

Frame 1739327800.png

Pros: Can quickly access preview, don’t need to click on full correspondence entry

Cons: Highlight, color changes, and bolding creates too many distracting elements, requires more clicks to “show sources”

Concept 2 | Sources below each event

Frame 1739327801.png

Pros: Can access sources for each key action taken, stack different source types

Cons: Not scaleable for multiple actions, more screen territory (clutter taking away from key events)

Concept 3 | In-Line Sources

Frame 1739327802.png

Pros: Blends into key actions text, leads directly to correspondence, intuitive pattern

Cons: Blue links overlooked within the text

Concept 4 | Expand Card designs for sources

Frame 1739327803.png

Pros: All in one location, easy to locate to get a general overview

Cons: Harder to verify each key action taken, screen territory when expanded

Sources: Follow-up Unmoderated Research 

To determine the best method of displaying sources to improve Associate trust in the new widget, I conducted follow-up unmoderated A/B testing with 7 Associates worldwide in 3 different sites. I tested usability for the new sources features, determined how often sources will be used to verify summary content, and validate the Issue Summary designs on a larger scale.

Research Goals

  • Test usability of sources concept (success rate and ease of Associate identifying where the info comes from)

  • Determine ratings for clarity, satisfaction, and completeness of source to identify missing/ confusing content

  • Determine how often sources will be used to verify summary content regarding trust.

  • A/B test 2 concepts to determine how sources should be associated with key actions. All sources in one location OR do Associates need to verify each detail.

image 151.png

~30 min
Sessions

image 150.png

7 Active
Amazon Associates

3 Countries
Worldwide (IN, CH, MNL)

Frame 1739327761.png

Recorded & unmoderated via usertesting.com

Concept A: Hover for source preview 

In the first concept, hovering on the blue hyperlinked number next to the key action opens up a source preview. Clicking on view full source link leads to correspondence.

Screen-Recording-2023-12-04-at-6 (2).gif

Concept B: Cards listed for all sources 

In the second concept, all the sources are located in one place which is revealed after clicking on “show sources”. This is to determine how sources should be associated with key actions. Do associates need to see a holistic view of where the information is coming from or do Associates need to verify each detail?

Screen-Recording-2023-12-04-at-6 (3).gif

Sources: Top Research Insights

Insight 1 | Prototype Comparison

Majority of Associates preferred Concept A and it ranked higher for EASE metrics (easy to find, clarity and satisfaction, confidence).

Top reasons included: 

  • (1) Appreciated the automatic redirect down to correspondence

  • (2) Hyperlinks and the blue color allowed them to easily identify where to click

  • (3) Can quickly confirm each individual action because of the location of the hyperlink

“The first reason is very clear to understand. It's very easy to locate because it's beside that key point and it's highlighted with a hyperlink. So it's very easy to click there and to understand the content.” A5

Group 1739327758.png

Reasons Associates did not prefer Concept B:

  • (1) Trouble with locating the “show sources button”

  • (2) Confused when having to look through multiple source previews to find the relevant one

  • Majority of Associates also stated they felt less confident

“Before we had the hyperlink beside the key point, it was easy to understand. Here it's like so many sources, we cannot understand which one is the first to see or the second to see. I'm confused whether to click second or third or fourth one.” A5

Group 1739327758 (1).png

Insight 2 | Source Preview Content

Associates said the source preview content was not enough to address the case, and they still needed to access the full source.

Group 1739327801.png

6/6

7/7: Highlight of source information
4/7: “View full source button”
2/7: ALL the information visible in the source preview

Insight 3 | Trust and Confidence

Associates knew the Summary was machine generated

7/7

5/7

“Very confident” (Rating 5) in the Summary when aware that it is machine-generated

  • More efficient (in capturing info like ASINs)

  • Easily surface relevant information

  • “By letting machine into this process, it would make tedious work, less tedious.” A7

2/7

Less confident (Rating 3, 4) when aware it was machine-generated

  • Summarizing incorrect information without human awareness to determine correctness

Sources: Next Steps

  • Proceed with Prototype A for the rationale of easier to verify information

  • Improve source discoverability through changing hyperlink numbers to “[See Source]” or [“”]

  • Continue conducting further unmoderated testing

  • Change the color and style of “View full source” button to improve discoverability

  • Unprioritize date and contact type and focus on the source highlight with bold keywords and more content (3 lines versus 2)

Frame 1739327642.png

6. Final Designs

Final Figma Handoff Document

Consolidating usability testing feedback, unmoderated testing, stakeholder feedback, co-design reviews for the Figma Handoff, considering primary use cases, edge cases, error/empty states, multiple case intents, and ideal spacing.

Frame 1739327804.png

Considered overflow behavior for brand registry information, ASINs, categories of identifiers, considering modals, in-line expansions, and tooltips weighing their role in the case handling workflow

Frame 1739327763.png

Final Design Walkthrough

Competitor Research
User Research
Design Iterations
Evaluating Designs
Additional Feature
Final Designs
Usability Testing
Unmoderated Testing
74.png

Imagine Sarah is a tenured Associate with skills in FBA, IIDP, and more, and works primarily with email cases. She receives a promotions multiple reopens case. The new issue summary is immediately surfaced at the top so it easily draws her attention while correspondence is collapsed by default.

75.png

She first glimpses over the general case information to make sure the statuses of the case and account are normal.

76.png

All the general case information she needs is at the top so she doesn’t need to press show more but is immediately aware of how to do so.

Rectangle 3471349.png
77.png
78.png

She then easily copies the seller ID over to add to her annotations to access paragon workflows and troubleshooting tools in the future.

She also receives confirmation of the information being copied

79.png

Sarah then moves onto reading the summary content. She immediately understands the case intent - the coupon creation error and what has progressed in the case so far.

80.png

She clicks on the link to the corresponding listing to check for apparent issues and appreciates this convenience.

81.png

After looking at the listing, she reviews the key identifiers detected and clicks on one of the identifiers which links straight to the ASIN troubleshooting tool.

82.png

She appreciates that she doesn’t have to look through correspondence entries, the initial contact, or annotations to access the ASINs, as well as doesn’t have to individually navigate to this tool.

83.png
86.png

Sarah finds the issue with the ASIN classfication but she knows she can always spoof the seller account with the “view seller account button.”

Sarah then copies all the ASINs to add to her annotations, saving her the trouble of individually copying or locating these ASINs.

87.png

She finally opens up correspondence, just to annotate what information is needed from to the Associate and request a reply

88.png

Without having to look through the correspondence which has more than 40 entries, Sarah is glad she is able to accurately and efficiently address the seller’s issue through using the new Issue Summary,

Reflection

Overall Reflection

During my second internship at Amazon, some unique takeaways included learning about the complexity of the Associate Enterprise problem space which greatly differed from my experience on the Seller side last summer. Since this platform is used by a more niche group of experienced professionals (10,000 Amazon Associates), the interfaces and surfaced features were more unintuitive and overwhelming, with technical jargon and large amounts of content to parse. In contrast, Amazon Seller Central, catering to sellers globally with different levels of experiences and backgrounds, designs are prioritized to always remain simple with limited options and broken into short steps. This project required problem-solving and extensive competitor research into other AI platforms to determine how to minimize information overload and visually represent page content. I had the opportunity to conduct 3 rounds of research and testing with 18 Associates (within 14 weeks) which was crucial to accurately addressing their pain points and workflows, and not changing aspects too much from what Associates are accustomed to.


Working within my amazing UX team again, I was given ownership of this project as the lead designer collaborating directly with several cross-functional stakeholders, grasping the importance of frequent design reviews, and incorporating feedback from varied perspectives. Leading the project from start to launch, I experienced a full-time delivery schedule, creating precise workback plans, delivering detailed implementation specs, and scheduling weekly syncs for alignment on the priority and feasibility of launch features. Also working in the constantly changing, fast-paced, AI space, I had numerous time-sensitive product and tech deadlines to manage to successfully launch within 3 months in January 2024

Also here is a Seattle-themed Peccy (Amazon's Mascot) I drew!

final version (2).png
Project 1

Project Overview

Project Description 

As a part of the CCAI UX Team, my first project was redesigning the Agent Assist Conversation Profile (CP) configuration process, integrating new LLM capabilities (CP further explained below in Context). As the solo project designer, with an end-to-end design process, I led UX for integrating new time-sensitive generative AI features in order to provide customized real-time suggestions to call center agents.

Project Scope & Impact 

I delivered implementable solutions directly to the engineering team alongside implementation specifications within 7 weeks. Considering the growing AI space with fast turnaround, designs are currently being launched as a new component of the Cloud Enterprise Agent Assist product. This product is currently used by large-scale customers including Verizon and Accenture.

Project Stakeholders

I coordinated and led weekly review presentations evaluating wireframes and hi-fidelity mockups with cross functional stakeholders (Product Manager, 2 Front-End engineers, 1 back-end engineer, UX researcher, UX product owner) and tested prototypes with potential users

bottom of page