Enterprise AI for Lead Time Visibility


A global retailer offers a wide assortment of home improvement products across 2,000 global retail locations. The company attempted to predict inbound lead times by utilizing a 13-week rolling average-based approach. This rolling average-based approach is aggregated by lane at the factory-distribution center level and does not consider purchase order specific characteristics (e.g., SKU, supplier, carrier). Additionally, with recent supply chain challenges and variability in supply, the customer made inbound order lead times fixed to mitigate changes in inventory levels. As a result, lead times were either manually updated by replenishment analysts or buyers placed orders and manipulated order inputs (e.g., adding the order to a program or changing vendor delivery SLAs) to expedite deliveries.

The following challenges continued to impact the company’s ability to anticipate order delays and gather real-time predictions:

  • Statistical outliers were difficult to detect. Invalid or missing timestamps, and poor quality of carrier data were challenging to mask, making it difficult to train AI algorithms.
  • External indicators of delays were hard to predict without access to Automatic Identification Systems (AIS) data to build port congestion features for AI algorithms.
  • Data scientists had no way to monitor or segment models that were generating rolling average lead times across the customer’s supply network.
  • Reuse of use case artifacts (e.g., models, data aggregations, and predictions) was not possible with existing process of ad-hoc data views, scattered notebooks, and brittle endpoint solutions.

As a result of these challenges, the global retailer faced an inventory backlog at port and distribution centers, transportation fines due to emergency shipments, and higher labor costs due to over-staffing.


C3 AI configured the C3 AI Supply Network Risk application across two machine learning models: static, one-time predictions for estimated delivery date upon PO creation, and continuously updated ETAs once containers leave a port and are in transit. For the end-to-end model, a one-time prediction was generated at the PO Creation date to predict what date the PO would arrive at the destination customer distribution center. For the port-to-port model, a daily prediction was generated to predict the remaining time for a container to arrive within the network while in transit.

The application and models were scoped for three United States port clusters, comprising roughly 60% of total import volume. Models were evaluated both using pre-COVID data and more recent data, with consistent results throughout both time periods.

The joint team built a multi-screen user interface to visualize AI-based insights and delayed orders based on severity of delay. The users can investigate order milestones both geospatially and temporally to view the intended path of the order and the underlying nodes that are causing the delay. The company’s data science team can monitor and manage the models generating the predictions through the Model Ops screens configured for the project.

About the Company

  • $150+ billion annual revenue in 2022
  • 2,000+ stores worldwide
  • 65+ distribution centers worldwide
  • ~500,000 employees


Project Objectives

  • Unify data from 12 enterprise data tables and three external data sources (e.g., Automatic Identification Systems (AIS) data from vessels, weather, and container congestion at ports)
  • Apply machine learning algorithms to predict inbound order lead times and daily ETAs while orders are in transit
  • Configure the C3 AI Supply Network Risk user interface and expose AI insights and unified analytics to end users


Project Highlights

  • 4.5 years of historical data integrated, comprised of 2 billion rows from 12 Big Query tables and three external data sources
  • Built an extensible data model with 30+ C3 AI logical objects
  • Developed 200+ timeseries analytics for machine learning models and application UI
  • Configured and tested 480+ machine learning model permutations to predict inbound lead times
  • Configured the C3 AI Supply Network Risk application user interface


in potential annual economic benefit for imported orders
in potential annual economic benefit when scaled to domestically-sourced orders
improvement in lead time predictions for the estimated delivery date compared to retailer’s baseline
improvement in daily lead time predictions for the estimated delivery date once an order was in transit

Solution Architecture


By using the C3 AI Supply Network Risk application, the global retailer is able to:

  • Generate up to $30 million of potential annual economic benefit across imported products ($100 million when scaled across domestic orders), based on three value drivers:
    • Reduced inventory levels with minimized replenishment uncertainty
    • Reduced transportation costs and fines with avoided emergency shipments
    • Reduced excess labor costs
  • Improve lead time accuracy compared to rolling average baseline, including:
    • 55% of end-to-end (E2E) lead time predictions
    • 25% of port-to-port (P2P) lead time predictions
  • Leverage a data foundation that enables rapid design, development, and deployment of a suite of AI-enabled supply chain applications in the future (e.g., assortment optimization, demand forecasting, warehouse optimization, and inventory optimization)

Proven results in weeks, not years

Get insights into C3 AI’s capabilities, enterprise AI best practices, and highest-value use cases.
Gain insights into the C3 AI Platform's capabilities, its model-driven architecture, and test it against your company's sample data set.
Identify a high-impact business problem and collaborate with the C3 AI team to rapidly build an AI application that solves it.
Scale and deploy a tested C3 AI application into production. Incorporate user feedback and optimize algorithms to drive maximum economic value.

Want more information?

Contact us