Professional Practice: CLASSUM
Connect — Error Reporting & Feedback Loops
AI systems improve through feedback, but only when people feel comfortable providing it. This project examines how the absence of an integrated feedback pathway can quietly reduce participation—and why that matters for long-term system reliability and AI accuracy.
During Connect’s early rollout, the team prioritized core automation features and onboarding experience. As a result, an in-product error-reporting system was not implemented at launch. Instead, feedback was handled through alternative channels, primarily via direct communication with assigned human customer support.
This approach created a warm and reassuring first impression, especially during early adoption and testing phases. Initially, this seemed sufficient.
Over time, usage patterns began to change. As more users interacted with the system in real workflows, error reporting decreased—not because issues disappeared, but because providing feedback required extra effort outside the product. Users hesitated to report problems, often feeling that it was burdensome or disruptive to their primary tasks.
“I felt bad reporting the same error again.”
Academic Affairs Staff, University of Ulsan
“I didn’t want to bother you too often… so sometimes I just assumed it would be fixed eventually and let it go.”
Student Support Office Staff, Gwangju University
This drop in participation was subtle and easy to miss. However, its impact was significant: with fewer reported errors, opportunities for system improvement diminished, and AI accuracy plateaued. User research later confirmed that keeping feedback outside the product had unintentionally discouraged participation.
This project focused on how feedback location and visibility shape participation in human–AI systems. Rather than asking how to collect more feedback, the core question became:
How can feedback be made easy, expected, and safe—without making users feel responsible for fixing the system?
Key UX Decisions That Lowered Barriers
Decision #1
The strongest emotional barrier came from remembering later and the pressure of “making a formal report.” So I placed the report button directly in the answer bubble where users first notice the issue. This allowed users to act in the moment, without hesitation or internal negotiation.
Decision #2
When I reviewed previously submitted error reports, I noticed that users frequently referenced the original question and the AI’s response while explaining why the answer was incorrect. To support this reporting pattern, I designed the reporting view to display the original Q&A directly on the same screen. This allowed users to check the context immediately and write more accurate reports without navigating away or recalling details from memory.
Decision #3
All user-submitted reports flow directly into Connect’s Back Office, where the team can review them in real time. Users provide both the incorrect AI answer and the corrected version, giving immediate signals for improvement. Each report can be analyzed through the AI Error Analyzer—which identifies whether the issue occurred in the Data, Retrieval, Generation, Validation, or Output stage.
Before this project, reports came through scattered channels like KakaoTalk messages, spreadsheets, and QA notes. I consolidated these fragmented paths into a single reporting pipeline, eliminating manual tracking and enabling the team to run diagnostics instantly. This unified flow made the AI improvement process faster, clearer, and far more reliable.
Reporting Flow
User notices an incorrect response. User can find the “Report Error” button inside the answer bubble. User reviews the Q&A and submits the corrected version. The report flows directly into the AI improvement pipeline. This end-to-end flow was intentionally designed to feel light, quick, and unintimidating.
Reviving Participation and Strengthening the AI Learning Loop
After integrating an in-product error-reporting system, user participation recovered—not because more features were added, but because the emotional burden of reporting was reduced. Making feedback visible, accessible, and low-effort encouraged users to speak up again within their normal workflows.
As a result, consistent and high-quality user corrections began flowing back into the system. Within two months, the AI model’s accuracy improved from 87% to 93%, a 6% increase, and the project was recognized internally as a key contributor to that improvement.
*Accuracy measured two months before and after the release, reflecting the impact of renewed user participation.
Beyond model performance, the redesign significantly improved operational reliability. A previously fragmented process—spread across chat reports, manual spreadsheets, and pasted URLs—was consolidated into a single, structured pipeline:
This unified workflow removed friction for both users and the team, making the AI improvement loop smoother, faster, and sustainable over time.
How This Project Changed My Understanding of AI and UX
Before this project, I assumed that AI improvement was driven primarily by engineering efforts—through better prompting, tuning, and pipeline optimization. Designing the error-reporting experience challenged that assumption.
By lowering the emotional barriers around reporting, I revived a behavior the AI system depended on to learn. The model improved not because the technology changed, but because people felt comfortable participating again. This made it clear to me that emotional experience is not a surface-level concern—it is a structural part of how AI systems evolve.
Building a unified reporting pipeline reinforced this insight. Consolidating scattered feedback channels into one clear flow reduced friction for both users and the team, accelerating the improvement cycle and contributing directly to measurable accuracy gains.
This project taught me that:
Ultimately, this work went beyond interface design. It was about creating the conditions for people to speak up again—and recognizing how powerful that participation is in shaping an AI system’s growth.
