← Index

Goodwill

Relay: reimagining the donation processing workflow with voice

Role

Product Designer

Timeline

6 months

Skills

UX Research · Interaction Design · UI Design · Prototyping · Voice Interfaces · Usability Testing · Service & Workflow Design

Team

Brittany Jain, Will Pan, Casey Potrebic, Yusen Zhang, Holly Zhu

Partner

Goodwill of Southwestern Pennsylvania

01

Project Overview

Problem

Goodwill processors inspect, categorize, price, and tag millions of donated items each year. But repetitive software interactions interrupt this physical workflow, requiring processors to move between clothing racks and shared computers and repeatedly enter information into a legacy POS system.

These inefficiencies slow processing and limit how quickly donated clothing can reach the sales floor.

Goal

Design a faster processing experience that reduces cognitive load and physical effort while helping processors tag clothing quickly and accurately.

Current process at a glance
01
Sort
Inspect, categorize, price
02
Rack
Fill the rack, hold decisions in memory
03
POS
Push rack across the floor, wait for a station
04
Tag
Print and attach tags item by item
05
Sales floor
Item available to shoppers
02

User Research

Understanding processing from the inside

We started by immersing ourselves in processors’ day-to-day work. Across 4 store visits, 8 processor interviews, 8 think-aloud sessions, and 36+ items processed ourselves, we observed not only what processors did, but where their workflow created unnecessary effort.

4
store visits
8
processor interviews
8
think-aloud sessions
36+
items processed ourselves

Five points of friction emerged

01
Decisions happen before data entry.

Processors often decide how an item should be categorized and priced while initially handling it, but have to remember those decisions until they reach the POS later.

02
Shared computers create bottlenecks.

Multiple processors may be working at once with only a limited number of softlines desktops, forcing processors to wait for a station.

03
Processing requires unnecessary movement.

After filling a rack, processors push it across the production floor to enter items into the POS, adding travel time and congestion.

04
The existing POS requires repeated interaction.

Logging a single item requires processors to navigate multiple screens and make many taps.

05
The workflow creates backtracking.

By the time an item is ready for the sales floor, a processor may have handled it three separate times: sorting, logging, and tagging.

Key insight

The opportunity wasn’t simply to make data entry faster. It was to eliminate the separation between making a decision and recording it.

Current workflow, annotated
Click to zoom — the current processing workflow, annotated with the five points of friction.
03

Problem Statement

How might we reduce cognitive load and physical effort while helping processors tag clothing faster and more accurately?

Our research led us to one guiding principle:

Let processors enter information at the moment they make a decision—without leaving the rack.

04

Strategy

What if processors didn’t need their hands to use the POS?

Because processors’ hands are constantly occupied inspecting, hanging, and tagging garments, we explored ways of capturing information without interrupting their physical work.

We compared three input modalities with processors using storyboards: smart glasses, voice, and tapping on a tablet.

Example of the storyboards used in testing
ModalityMental effortPhysical effortPreference
VoiceLowerLowerStrongest
Tablet tappingHigher than voiceHigher than voiceMiddle
Smart glassesLowLowLowest

Voice emerged as the strongest direction. Processors perceived it as requiring less mental and physical effort than tapping, while smart glasses received the lowest preference ratings.

Could voice actually work in a noisy Goodwill backroom?

We tested microphone types and transcription tools in noisy, warehouse-like environments. Even in a large, echoing basketball court with people playing, our strongest transcription & microphone combination achieved approximately 98% accuracy, demonstrating that voice was technically viable.

Later, we evaluated cloud and local speech-to-text models using Goodwill’s actual vocabulary. Deepgram achieved 97.6% accuracy, while Vosk provided a 90.8% accurate local fallback, giving us a path toward both performance and reliability.

Speech-to-text accuracy on Goodwill vocabulary
Deepgram · cloud97.6%
Vosk · local fallback90.8%

So we determined voice could capture the processor’s decision. Now we had to figure out how to get that information into Goodwill’s existing system.

05

Explorations

Exploration 1 — Replace the legacy POS

Our first instinct was to redesign the experience entirely.

The existing Solutions interface required processors to navigate multiple screens and repeatedly tap to enter a single item. We explored a simpler, voice-first interface that could replace those interactions.

But Solutions wasn’t just an interface. It was connected to inventory data, production tracking, checkout, and other backend systems.

Replacing it would mean replacing critical infrastructure—not just redesigning a screen. So we decided to work with the existing system instead of replacing it.

Exploration 2 — Integrate directly with Solutions

If we couldn’t replace Solutions, we wanted to connect our interface directly to it. That path quickly hit another constraint: Solutions provided no direct or public API access.

No API access

No public or direct API connectivity available

Rather than treating that constraint as a dead end, we reframed the problem:

If processors can operate the POS by tapping buttons, could software tap those buttons for them?

We built a browser proof of concept using UI automation. A processor could speak an item’s information, and the system translated that input into the interactions required by Solutions. The prototype proved that we could automate the legacy workflow without requiring API access.

UI automation proof of concept — a spoken command drives the tagging UI automatically.
Voice inputInterpret item informationUI automationSolutions POS

Exploration 3 — A voice-first replacement interface

Next, we combined voice and automation into a dedicated fullscreen interface. Conceptually, it worked. Technically, it didn’t.

For UI automation to reliably interact with Solutions, the legacy application needed to remain in the foreground. Our fullscreen interface competed with Solutions for focus, breaking the automation.

Windows focus conflict
Relay fullscreen
Needs the screen
focus
Solutions
Needs foreground access for automation

Only one application can hold focus, so the automation stopped.

Pivot — From replacement interface to companion

Instead of covering Solutions, we designed a lightweight widget that sits directly on top of it.

The widget gives processors access to voice controls and feedback while allowing Solutions to remain visible and operable underneath. Processors no longer have to switch between applications, and the UI automation can continue working in the background.

Fullscreen — broke automation
Widget — Solutions stays live
06

Redesigning the Workflow

Solving the interface was only part of the problem. Our original research showed that processors interacted with the same garment three separate times: sort → log → tag.

Touch 1
Sort

Inspect, decide, hang on the rack

Touch 2
Log

Handle again at the POS to enter it

Touch 3
Tag

Handle a third time to attach the tag

We asked whether Relay could eliminate that backtracking entirely. Instead of processing items in batches, we tested a tag-as-you-go workflow, where processors completed each item from beginning to end before moving on to the next.

The result surprised us.

Batch vs. tag-as-you-go
Batch processing
SortSortSortLogLogLogTagTagTag
Relay — tag as you go
SortLogTagSortLogTag
25%
less processing time than batch processing
3 → 1
garment touchpoints

That finding changed Relay from an alternative input method into a new processing workflow.

07

Final Designs

Introducing Relay

Relay is a voice-operated tagging assistant that allows Goodwill processors to log and tag clothing while continuing to physically process it.

Rather than asking processors to adapt to a new system, Relay sits alongside Goodwill’s existing POS and automates the interactions that previously required their attention.

A processor speaks an item and Solutions fills in the details underneath.
01

Process directly from the rack

A Windows tablet sits beside the processor's station, eliminating repeated trips to a shared desktop.

Tablet beside the processing rack
02

Speak while handling garments

Processors can say item details aloud while inspecting and hanging clothing, capturing decisions at the moment they're made instead of remembering them for later.

Listening state — transcript & recognized details
03

Automate the legacy POS

Relay translates spoken information into UI actions inside Solutions, allowing processors to continue handling garments while the system performs repetitive POS interactions on their behalf.

Relay recognizing details, Solutions updating itself
04

Tag as you go

Once an item is logged, its tag can immediately be printed and attached. The processor completes the garment before moving to the next one, reducing touchpoints from three to one.

Printing and attaching a tag immediately

Together, these features allow a processor to complete an entire rack without leaving their processing station.

08

Retrospect

Impact

We tested Relay in-store for 23 hours as part of a summer that included 9 store visits, 15 research experiments, and 18 prototype iterations.

23
hours of in-store testing
9
store visits
15
research experiments
18
prototype iterations
Pilot result
2.12×

increase in production speed — a 112% increase in items processed per hour.

Before1.00×
With Relay2.12×

items processed per hour, single-processor pilot

More importantly, the processor using Relay immediately understood the value of removing those repetitive interactions:

I love it. This is great, this is going to make us faster.

— Pilot test processor

Lessons Learned

Designing for constraints can reveal the product.

No API access and Windows application limitations initially felt like barriers. Ultimately, they pushed us toward the widget model that made Relay compatible with Goodwill’s existing technology rather than dependent on replacing it.

Improving an interface isn’t always enough.

Some of our largest efficiency gains came from questioning the workflow around the interface. Moving from batch processing to tag-as-you-go reduced both time and physical handling.

Meet users where decisions happen.

Processors were already categorizing and pricing garments while handling them. Relay became valuable when we stopped asking them to translate that thinking into a separate digital workflow later.

How the design evolved
VoiceFullscreen UIUI automationWidgetTag-as-you-goRelay

With more time, I’d like to

I would expand the pilot across multiple processors and stores to understand whether the production improvements hold across different working styles and environments.

I’d also test Relay with a broader range of accents, explore multilingual voice input and automatic translation, and make the system configurable across additional product categories and pricing structures.

Finally, as processors become more comfortable trusting automation, I’d explore stronger feedback and error-prevention mechanisms to ensure that increased speed doesn’t come at the expense of tagging accuracy.