Cost Responsibility Allocation Platform

Overview:

Designing an Agentic AI workflow for warranty claim classification and human validation.

Ford processes thousands of warranty claims that must be reviewed to determine whether repair costs are the responsibility of the manufacturer, the supplier, or require additional investigation. Traditionally, this process involved significant manual effort, inconsistent decision-making, and lengthy review cycles.

As the Lead Product Designer, I designed the end-to-end user experience for an AI-assisted Cost Responsibility Allocation platform that combines large language models, machine learning, and Human-in-the-Loop validation to automate much of this workflow while maintaining expert oversight.

Role:

Lead Product Designer

Responsibilities:

To tackle this complex challenge, I implemented a user-centered, collaborative design process:

UX strategy
User research
Journey mapping
Workflow design
Agentic AI interaction design
Human-in-the-Loop (HITL) workflow design
Information architecture
Figma design and prototyping
Error state design
UX writing
Test case creation
Stakeholder presentations
Cross-functional collaboration with AI engineers and product owners

Ford Enterprise AI Model

The Challenge:

Warranty engineers needed a way to:

Reduce manual review of thousands of claims
Automatically generate failure modes using AI
Classify claims into responsibility categories
Validate AI decisions with SMEs
Measure AI performance before production use
Support both end-to-end and modular workflows

The existing process required multiple disconnected tools and extensive manual analysis.

Goals:

Design a platform that would:

Reduce manual effort
Increase consistency
Keep humans involved for high-risk decisions
Allow independent execution of each workflow stage
Provide transparency into long-running AI jobs
Build trust through explainable AI

Design Process:

Research

I worked closely with:

Warranty engineers
SMEs
AI engineers
Product owners

To understand:

Existing review processes
Manual pain points
Decision logic
Required approvals
Workflow dependencies

Key UX Challenge:

he biggest design challenge wasn't building another dashboard.

It was designing an experience where:

AI performs most of the work
Humans intervene only when necessary
Users always understand the current workflow state
Long-running jobs remain transparent

This required creating an agentic workflow instead of a traditional application flow.

Solution:

The platform consists of five connected workflows.

1. Failure Mode Generation

Users upload warranty claims.

The AI:

analyzes technician comments
groups similar claims
generates failure modes
produces confidence scores
creates an executive summary

If confidence is sufficient, the workflow pauses for SME review.

2. Human-in-the-Loop Review

Rather than approving every claim individually, SMEs review generated failure modes.

The interface allows experts to:

assign responsibility categories
edit AI-generated labels
add comments
save decisions

Once completed, the workflow automatically progresses to the next stage.

3. Claims Binning

Using the approved failure modes, the system classifies every warranty claim into:

Ford
Supplier
Inconclusive

This stage performs high-speed semantic matching across the entire dataset.

4. Sampling

Instead of validating every classified claim, the platform generates a statistically representative sample using stratified sampling.

This preserves the original distribution of responsibility categories while dramatically reducing SME effort.

SMEs validate only the sampled claims.

5. Metrics & Evaluation

The final workflow compares AI predictions against SME decisions.

The system automatically calculates:

Overall accuracy
Precision
Recall
F1 score
Category-level performance
Confusion matrix

This enables the business to determine whether the AI model is ready for production.

UX Decisions:

Human-in-the-Loop Status

Instead of displaying generic loading screens, I designed a dedicated Human-in-the-Loop state.

Features included:

Progress tracker
Human Review indicator
Workflow status
Estimated completion
Resume support

This helped users distinguish between AI processing and human review.

Modular Workflow Entry

Rather than forcing users to complete every step sequentially, I designed each workflow to operate independently.

Users can begin with:

Failure Modes
Claims Binning
Sampling
Metrics

provided they supply the required inputs.

This increased flexibility for different teams and use cases.

Long-Running Job Management

Because some AI processes take 10–15 minutes (or longer), I designed a workflow centered around persistent jobs rather than individual pages.

Users can:

leave the application
return later
resume exactly where they left off
monitor workflow status

Explainable AI

Trust was a major design goal.

Rather than presenting only AI decisions, the interface surfaces:

confidence scores
generated failure modes
supporting evidence
evaluation metrics
human validation results

This increased transparency and confidence in AI-generated outputs.

Outcome

The platform provides a scalable framework for AI-assisted warranty analysis by:
Reducing manual claim review
Standardizing responsibility classification
Incorporating human validation where needed
Supporting modular workflows
Measuring AI performance before deployment
Increasing transparency through agent status and explainable AI

What I'm Most Proud Of

Designing an intuitive experience for a highly technical AI system.

Instead of exposing users to complex orchestration, machine learning models, and multiple backend agents, I translated those processes into a workflow that feels simple, transparent, and predictable. By introducing clear workflow stages, Human-in-the-Loop checkpoints, resumable long-running jobs, and modular entry points, I helped create an enterprise AI experience that is approachable for business users while supporting the needs of data scientists and engineers.