Assistive AI for Feedback
Year
2025 - 2026
Company
JPMorganChase
Role
Project Lead
Embedding assistive AI into an enterprise review workflow to help 300,000+ employees turn vague comments into specific, constructive, and actionable performance review feedback at scale.
© JPMorganChase Assistive AI
Context
for better feedback
JPMorganChase supports the growth of 300k+ employees, and one important part of that growth is helping employees give each other meaningful feedback that not only recognizes what went well, but also offers constructive guidance and clear next steps.
Over a three-month exploration, I helped shape this feature from a basic AI rewrite idea into a more thoughtful assistive AI writing experience — one that supported quality, trust, and human ownership in a sensitive feedback workflow.
The ask
Raise feedback quality
Leadership wanted quality improvement, not submission volume. The bar wasn't "more" — it was "meaningfully better."
The constraint
High-trust AI
Performance reviews are sensitive. AI needed to enhance human judgment without replacing it or undermining employee ownership.
The outcome
More actionable feedback
The experience helped employees write feedback with clearer context, observable behaviors, and practical next steps.
The scale
Built for 300k+ employees
The solution was designed to fit into an enterprise review system and scale across a large, complex organization.
Problem
Not a submission problem.
A quality problem.
Employees were already submitting feedback. The bigger issue was that submitted feedback was not always useful.
Praise without specifics
Unclear context
Missing impact
No actionable next step
Judgmental wording
Discovery
for the workflow
The existing flow was simple: select an employee, write feedback in a text input box, and submit.
For this project, the key design question was not simply where to add AI, but what kind of AI pattern made sense. I explored whether this should be a conversational AI experience or an assistive AI experience embedded directly into the feedback flow.
Path A
Conversational AI
A conversational AI model could support open-ended coaching, but too heavy for a simple submission flow. Employees would need to leave the writing moment, choose their intent, and translate the conversation back into the form.
Write feedback
Click "Refine"
Chat open
Choose prompt
Iterate
Apply revised feedback
Selected path
Embedded Assistive AI
An assistive AI model is lightweight, contextual, and fit the workflow better. It allowed employees to write first, then use AI to refine their own input against predefined good feedback criteria.
Write feedback
Click "Revise"
Accept or edit
Submit
Quality framework
Defining what
Before designing the comprehensive AI interaction, we needed a clear quality framework. The AI refinement was based on four criteria, the same lens an experienced manager would apply.
Exploration
Exploration 1
Single popover with explanation
A small popover appeared after users clicked "Revise". It summarized how AI improved the feedback across three qualities: behavior-based, constructive, and actionable.
+ Lightweight and close to the input field.
– Became dense when explanations were longer.
Design learning
A compact popover worked well for quick feedback, but it did not leave enough room for deeper explanation. For sensitive writing tasks, transparency needs more space.
Exploration 2
Guided revision menu with targeted goals
Instead of explaining changes after the fact, this direction let users guide the AI first. The menu offered preset goals like more specific, more constructive, or add a clear next step.
+ Increased user control while keeping the interaction lightweight.
– Added an extra decision step and less transparency into exactly what changed.
Design learning
Preset options reduced the effort of prompting AI and made the interaction feel more intentional. However, users still needed a clearer way to review what changed afterward.
Exploration 3
Multi-step popover with sentence-level highlights
This concept broke the revision into step-by-step suggestions. The AI highlighted specific parts of the feedback and explained each improvement one at a time.
+ Made AI logic more traceable.
– Too click-heavy and slowed users down.
Design learning
Detailed explanations improved trust, but too much step-by-step interaction added friction. Users preferred a faster way to scan the reasoning and move on.
Final solution
A side sheet that supports
The final direction used a side sheet for AI refinement. This gave the AI enough space to explain its analysis while keeping the original feedback visible in the main form.
Original feedback remains visible
AI analysis is criteria-based
Suggested refinement is separated
User must choose to accept
Human ownership is preserved
Impact
Since launch, the feature helped employees move from vague comments to feedback that was clearer, more useful, and easier to act on.
Satisfaction score
positive rating based on thumbs up/down feedback
AI adoption
used Refine before submitting
Refinement acceptance
accepted or edited AI suggestions
