Why did Uber Freight use multiple AI models instead of a single solution for document processing?

Marchevsky explained that the ensemble approach using multiple specialized models was intentional. Rather than forcing one model to handle all document types, Uber Freight found it more effective economically and operationally to use different tools optimized for specific problem spaces within document processing. This avoided over-reliance on any single model and achieved the highest precision across tens of thousands of documents processed daily.

What is shadow mode testing and why does Uber Freight use it?

Shadow mode is a parallel testing environment where new AI models run alongside legacy systems without making actual decisions on live data. This allows for direct A/B comparison between the old approach and proposed new approach. Marchevsky emphasized that this methodology is critical for validating that a new system actually performs better before rolling it out to production, reducing the risk of deployment failures.

What does the 95% AI pilot failure rate tell us about freight industry challenges?

Marchevsky cited research showing that 95% of AI proofs of concept fail due to production issues, not model performance. In freight specifically, the cost of failure is higher than most industries due to impact on consumer supply chains. The lesson is that successful AI adoption requires operationalization from day one, involving the teams that actually run processes rather than having engineers develop solutions in isolation.

Why is the cost of failure higher in freight AI compared to other industries?

SONAR's Julie Van de Kamp noted that freight's failure cost is significantly higher than consumer-facing AI use cases. Errors in logistics directly disrupt supply chains and consumer delivery. This creates a naturally higher bar for adoption and requires human oversight as a safety backstop. Additionally, freight involves unpredictable, unbounded environments where large language models perform less reliably than in controlled scenarios.

What should shippers look for when evaluating AI vendors?

Marchevsky recommended focusing on three areas: (1) Data governance practices—ensuring your data doesn't become vendor property, (2) Environment-specific benchmarks rather than generic benchmarks, as traditional metrics don't predict performance in your unique setting, and (3) Evidence of production deployment capability, not just model building. Uber Freight built its own benchmarking systems to understand how models perform in their specific environment.

Technology & Innovation

High Impact

Uber Freight CTO: Why 95% of AI Pilots Fail and How to Win

Jul 20, 2026

Get tomorrow's supply chain signal

Daily supply-chain brief. Free, unsubscribe anytime.

The signal

Val Marchevsky, CTO of Uber Freight, shared critical lessons on AI adoption in freight at the Supply Chain AI Symposium 2026, warning that the industry must focus on solving fundamental business problems rather than chasing trendy applications. The rise and fall of the AI wrapper economy demonstrates that companies building shallow solutions on temporary gaps in technology capabilities disappear quickly, while those addressing core operational pain points endure. Uber Freight's success with document processing—reducing manual work by over 50%—came not from a single AI model but from an ensemble approach using specialized tools fitted to specific problems within the broader workflow.

Marchevsky emphasized that freight faces uniquely high stakes for AI implementation due to the cost of failure in consumer supply chains, where human oversight remains essential. The company's adoption of shadow mode testing—running new models in parallel with legacy systems without making live decisions—allows for rigorous A/B comparison before production deployment. A sobering finding cited by Marchevsky: 95% of AI proofs of concept fail due to production issues, not model performance, highlighting the critical importance of involving operational teams from day one rather than relegating AI development to isolated technical groups.

For supply chain professionals evaluating AI solutions, Marchevsky's guidance is clear: scrutinize data governance practices, demand environment-specific benchmarks rather than generic metrics, and prioritize vendors demonstrating a systematic approach to production deployment. The future of freight AI lies not in replacing existing systems but in building agentic software that integrates with legacy infrastructure, a more pragmatic path that acknowledges the fragmented, complex reality of modern logistics operations.

#production-deployment

#shadow-mode-testing

#ai-adoption-strategy

#operational-efficiency

#freight-innovation

#logistics-transformation

Frequently Asked Questions

What This Means for Your Supply Chain

Simulation Suggestion

this month

What if document processing errors increase by 5%?

Simulate a scenario where the accuracy of automated document processing (proofs of delivery, bills of lading) decreases from current high precision levels to 95% accuracy. Model the impact on manual review workload, processing time, and labor costs across a typical monthly document volume of 10,000+ documents.

Run this scenario

Simulation Suggestion

this month

What if legacy system integration failures prevent 20% of identified AI use cases from reaching production?

Simulate the operational impact if integration challenges with fragmented legacy systems prevent successful deployment of 20% of AI use cases identified in internal hackathons. Model the cost of maintaining parallel manual processes, staff retraining requirements, and revised ROI calculations for the remaining 80% of successfully deployed use cases.

Run this scenario

Simulation Suggestion

strategic

What if production deployment of new routing optimization AI takes 6 months instead of 3?

Model the scenario where extended shadow mode testing and iteration delays deployment of an AI-powered truck routing optimization system by 3 months. Estimate the opportunity cost in fuel savings, fleet utilization improvements, and competitive advantage lost during the extended testing period.

Run this scenario

Go deeper:

Supply Chain Intelligence: The 2026 Buyer's Guide

The four-layer stack (twin, simulation, sandbox, intelligence) and the framework for evaluating platforms.

Get the daily supply chain briefing

Top stories, Pulse score, and disruption alerts. No spam. Unsubscribe anytime.

The signal

Frequently Asked Questions

What This Means for Your Supply Chain

Related Articles