Fortune 100 client
2022-2024
Designing trust and actionability in AI-assisted decision systems
Problem
Why adoption, not model accuracy was the core problem
The AI model was working — it flagged anomalies and offered insights — but adoption was low. Engineers were stuck jumping between tools, didn’t trust the AI model outputs and hesitated to act on them. It wasn’t an algorithm problem — it was a UX one. We needed to make the AI interpretable, actionable, and actually usable in a real manufacturing environment.
Design goals
Design goals in a safety-critical context
Close the human-AI feedback loop
Capture human feedback to improve model accuracy over time and build a continuous learning system.
Reduce manual effort
Streamline repetitive investigation tasks by consolidating tools and reducing clicks for key workflows.
Build trust, drive adoption
Show tangible outcomes to build user confidence and integrate AI into real-world operations.
Constraints
Real-world tensions the design had to balance
Designing this AI-assisted investigation system meant working inside real manufacturing, safety, and model-maturity limits rather than ideal product conditions.
Real-time triage vs full analytical certainty
Engineers needed to decide quickly within live production timing, even without perfect certainty.
Single-unit clarity vs cross-unit pattern insight
Investigation began with one unit, while broader anomaly relationships required separate analysis.
Design decision 01
Making AI output interpretable through human feedback
Engineers were overwhelmed by flagged anomalies but lacked tools to validate and understand them. I designed clustering by signal pattern to help them spot trends across transmissions and added in-app feedback mechanisms so they could confirm or reject anomalies without needing ML expertise. I also introduced warranty correlation views to show when AI predictions matched real-world failures. These changes reduced manual work, improved model training, and made AI behavior more transparent and trustworthy.
Key Design Highlights
1. Anomaly Pattern Analysis
Helped engineers quickly spot recurring anomalies across units instead of comparing one by one.
2. Inline Feedback Loop
Let engineers mark anomalies as true or false directly in the workflow, improving model learning without disrupting investigation flow.
3. Warranty Correlation View
Connected AI-flagged anomalies to historical warranty outcomes, helping engineers validate whether predictions reflected real-world failures.
Design decision 02
Workflow redesign for enterprise efficiency
The AI identified problems, but the work to resolve them happened across disconnected tools. I redesigned the experience around the actual investigation workflow—centralizing review, escalation, and status tracking inside the AI tool to reduce context switching and manual effort.
Key Design Highlights
1. End-to-End Anomaly Review Flow
Brought anomaly investigation, review, and resolution into one seamless interface.
2. Built-in Hold Request System
Replaced spreadsheets and emails with traceable “Hold” workflows directly in-app.
3. Status Tracking Across Roles
Made engineering team progress visible and auditable — no more guessing or pinging.
Design decision 03
Designing for trust, adoption, and business confidence
I focused on making AI value visible and actionable by prioritizing urgent signals, highlighting outcome-based trends, and tying validated predictions directly into production workflows. We also integrated the tool into the plant's hold system, connecting AI actions to real outcomes. As a result, usage increased, trust improved, and the tool expanded to 8 plants.
Key Design Highlights
1. Action-Oriented Prioritization
Shifted the homepage from raw anomaly volume to urgency-based signals, helping engineers focus on what needs attention now rather than sorting through AI noise.
2. Outcome-Based Impact Trends
Used time-based trends and outcome metrics to make AI impact visible over time, emphasizing validated results instead of isolated predictions.
3. Production-Integrated Validation
Connected AI predictions to real warranty outcomes and integrated confirmed anomalies directly into production workflows, reinforcing trust and enabling decisive action.
Retrospective
