Building an AI-powered distress detection system that cut counselor review time by 60%
Role
Product
Palo – a classroom-ready SEL curriculum with interactive activities that capture rich student behavioral and emotional data.
The problem
Palo’s interactive SEL curriculum generated rich student data – mood check-ins, scenario choices, and written reflections. While valuable, this volume of signals overwhelmed counselors, making it difficult to quickly identify which students needed attention. Insight existed, but action was slow.
Outcome
Reduced counselor review time by ~60%
All pilot counselors extended usage and referred peers
Contributed to expansion into 12+ school districts by the next term
Context: Counselors decide whether Palo succeeds
Although schools purchased Palo, counselors determined whether it was actually used.
They were responsible for hundreds of students, worked under constant time pressure, and were already stretched across administrative and emotional labor. Any solution that required analysis, interpretation, or extra effort was unlikely to fit into their daily workflow.
This reframed the challenge:
| The goal wasn’t to show more data – it was to help counselors act faster.
Designing with real stakeholders: Pilots with school counselors
The challenge: When more data slowed decisions
Palo’s interactive curriculum captured nuanced student behavior through:
Mood check-ins
Scenario-based choices
Written reflections
This created a rich picture of student well-being – but also surfaced hundreds of fragmented signals across classrooms and weeks.
The core challenge
| How might we surface meaningful risk signals from rich behavioral data – without adding cognitive load or slowing counselors down?
First attempt: A dashboard counselors didn’t return to
Our first response was a comprehensive dashboard that visualized emotional trends, engagement levels, and behavioral patterns across time. It looked informative – but it didn’t work.
After initial logins, counselor usage dropped sharply. The feedback was consistent:
Too many charts
Too much interpretation
Too much time required
What failed
| We optimized for completeness instead of counselor intent.
Counselors didn’t want another system to analyze. They wanted clarity that fit into an already overloaded day.
Design principle: Time is the real constraint
In high-stakes environments, time – not data – is the scarcest resource.
Counselors didn’t lack information. They lacked the time and cognitive space to interpret it. This led to a clear design principle that guided every decision that followed:
| If an insight can’t be understood and acted on in seconds, it doesn’t belong in the system.
This shifted our focus from visualizing data to reducing decision effort.
Fail fast: Testing insight delivery inside counselor workflows
Before redesigning the dashboard, we needed to validate one question:
| Could we surface meaningful signals without asking counselors to change how they worked?
To shorten feedback loops, we moved insight delivery directly into counselors’ existing workflow – their inbox.
I partnered with our researcher and SEL expert to send bi-weekly insight reports via email, iterating weekly based on counselor feedback. This allowed us to test:
Signal prioritization
Clarity of language
Trust in surfaced insights
What worked
| A simple, prioritized list of flagged students – with just enough context.
What didn’t:
Charts that required interpretation
School-level summaries that delayed action
Designing the system: Flags before frameworks
Once we validated that counselors responded best to prioritized signals, the challenge became designing a system that could reliably surface those signals at scale.
The goal wasn’t to explain student behavior. It was to alert counselors when attention was needed – quickly, safely, and with enough context to act.
We designed an AI-assisted pipeline that transformed raw behavioral data into distress flags, rather than scores or predictions.
How the system worked
Student inputs from Palo’s interactive curriculum – mood check-ins, scenario choices, and written reflections – were:
Tokenized and anonymized
Processed using researcher-defined prompts
Analyzed for unusual emotional or behavioral patterns
When a pattern crossed a defined threshold, the system generated a distress flag, adding the student to a counselor watchlist for review.
AI design constraints
From the outset, we set clear boundaries on what the system should not do:
No scores – to avoid false precision
No automated conclusions – counselors retained judgment
No loss of context – surfaced signals always linked back to student responses
These constraints helped build trust and ensured the system supported – rather than replaced – counselor decision-making.
Progressive disclosure: Start with names, earn depth
With the system in place, the UX challenge was deciding how much information to show – and when. Counselors needed to:
See risk quickly
Understand context when necessary
Avoid cognitive overload by default
We designed the interface around progressive disclosure, revealing information only as counselor intent deepened.
| The system should follow counselor intent, not data hierarchy.
How information unfolded
Key idea was clarity by default – and depth on demand.
Level 1 – Watchlist
A simple, prioritized list of flagged students. No charts. No scores. Just names that needed attention.Level 2 – Student context
On selection, counselors saw recent emotional signals and excerpts from student responses.Level 3 – Patterns over time
Deeper trends and correlations were available only when counselors chose to explore further.
Design-engineering tradeoffs under real constraints
As the system moved closer to production, design decisions were shaped by real-world constraints inside schools, not ideal conditions.
Most counselors accessed Palo on low-bandwidth Chromebooks, often on unstable networks. This directly impacted what the interface could reliably support.
Rather than forcing visual complexity, we made deliberate tradeoffs to protect performance, clarity, and trust.
Tradeoff
Key tradeoff we made:
Removed complex histograms
Multi-parameter charts added processing overhead and slowed page loads on school devices.Dropped weekly mood aggregation views
Aggregated trends were useful in theory, but delayed access to urgent, student-level signals.Replaced precision charts with a feeling scale
A simpler scale aligned better with how counselors discussed emotional states with students.
| We optimized for reliability and clarity over visual density – because broken insight is worse than missing insight.
The final experience: Clear signals, confident action
The final dashboard experience centered on speed, clarity, and trust.
Counselors landed on a prioritized watchlist showing only students who needed attention. From there, they could quickly review context, see recent emotional signals, and decide when and how to intervene – without navigating complex analytics.
The system supported counselor judgment instead of replacing it.
What counselors could do quickly
Identify at-risk students in seconds
Review contextual excerpts from student responses
Understand emotional patterns without interpretation overhead
Add notes and track follow-ups over time
Every interaction was designed to reduce thinking time and increase confidence in action.
Impact
The redesigned system significantly reduced the effort required for counselors to identify and act on student well-being concerns.
~60% reduction in counselor review time
100% of pilot counselors extended usage beyond the pilot period
Peer referrals from counselors to other schools
Contributed to expansion into 12+ school districts by the following term
More importantly, counselors reported spending less time searching for concerns and more time supporting students who needed attention.
Insights
What this taught me
In AI-driven products, legibility matters more than sophistication. If users can’t immediately understand why something surfaced, they won’t trust it.
Designing for action is fundamentally different from designing for insight – it requires ruthless prioritization and restraint.
In high-stakes contexts, users don’t want automation. They want support they can trust.
These principles now guide how I design systems where accuracy, ethics, and usability intersect.
More case studies
Nagent AI
Designing a visual AI agent builder for non-technical users
A systems-led exploration of reducing cognitive load in agentic AI creation
TaccoMacco Mobile App
Building a mobile reading platform – and learning why products fail
Product lessons on habit formation and distribution challenges.