← Back to Portfolio Machine Learning

AI-Powered Document Review

Partnering with Data Science to build ML models that analyze customer photos in real-time, reducing manual review time and accelerating appointments

Role Lead Product Designer
Team Data Science, Product, Engineering
Timeline 6 months, Launched Autumn 2025
ML Document Review System
3-6s
Processing time vs 4-6 hour manual review
85%+
ML accuracy rate
+8%
Appointment completion increase
+6%
Buy rate increase

Manual Review Creates Friction

Customers upload odometer photos during the appraisal process, but associates had to manually review every single image—creating a 4-6 hour bottleneck before appointments could be scheduled

The Challenge

Associates were spending significant time reviewing odometer photos to catch the most common errors: blurry images, incorrect readings (trip vs actual mileage), and wrong image types. This created wait time for customers and slowed down the entire appointment flow.

When document validation took hours, customers would lose momentum and were less likely to complete their appointments. The longer the wait, the lower the conversion.

Opportunity Solution Workshop

We ran an opportunity solution workshop to identify the highest-impact interventions to reduce document review time and improve conversion

Key Hypothesis

If we could automate validation for the most common rejection reasons (blurry images, incorrect odometer readings, wrong image type), we could:

  • Save associate time by auto-flagging obvious issues
  • Alert customers faster when documents need resubmission
  • Increase appointment set rate by reducing wait time
  • Increase buys by keeping customers engaged in the flow

The Approach

We partnered with the Data Science team to design our own machine learning models that could analyze customer images as they upload, catching errors in real-time and saving both customer and associate time.

Mapping Questions & Flow

Used Miro to collaborate with Product and Data Science, identifying key questions about the ML process and mapping the validation flow

Understanding the ML Process

Miro board with process questions

Sticky notes exploring: How does the model work? What's the accuracy threshold? When do we route to associates?

Mapping the Validation Flow

Miro board with flow diagram

Customer uploads photo → ML analyzes → Confidence check → Auto-approve or route to associate

Customer Upload Flow

Designed the mobile experience for customers to upload odometer photos, with real-time feedback when ML validation is complete

Upload Screen

Odometer photo prompt

Processing State

ML analyzing image

Validated

Ready for appointment

ML-Powered Odometer Verification

Partnered with Data Science to build custom machine learning models that analyze customer photos in real-time, catching errors at upload and accelerating the entire flow

Real-Time Analysis

ML model analyzes odometer photos as customers upload them, extracting mileage readings in 3-6 seconds—dramatically faster than the 4-6 hour manual review queue.

Computer Vision

Automated Error Detection

Model identifies the three most common rejection reasons: blurry images, incorrect readings (like trip vs actual mileage), and wrong image types—catching issues before associates even see them.

Smart Validation

Intelligent Routing

Only flags uncertain cases for associate review. When the model is confident (85%+ accuracy), documents are auto-approved. Associates focus their time on edge cases that actually need human judgment.

Smart Routing

Measured Results

Clear correlation between faster document review and improved business outcomes

Speed & Efficiency

3-6 seconds
Processing time per document
85%+
ML model accuracy rate

Customer Impact

+8%
Appointment completion increase
Faster feedback
Real-time validation vs hours-long wait

Business Outcomes

+6%
Buy rate increase
Production scale
Thousands of validations daily

What I Learned

What Worked

  • Cross-functional partnership with Data Science

    Building trust with the DS team early meant we could design the ML model collaboratively, not just hand off requirements. This led to better accuracy and faster iteration.

  • Starting with the highest-volume error types

    Focusing on the top 3 rejection reasons (blurry, incorrect reading, wrong type) meant we could validate the model quickly with measurable impact.

  • Intelligent routing, not full automation

    Only auto-approving when the model was confident (85%+) meant associates still reviewed edge cases, maintaining quality while dramatically reducing overall review time.

What I'd Do Differently

  • More customer research upfront

    We focused heavily on the associate side (manual review pain) but could have done more research on customer expectations around upload feedback timing.

  • Earlier production testing with smaller rollout

    We tested extensively in staging but could have de-risked the launch with a beta group of customers before full production rollout.

  • More visibility into model decisions

    Associates wanted to understand why the model flagged certain images. Building in explainability from the start would have increased trust in the system.