How Worldwide Express automated freight exceptions with Amazon Bedrock

How Worldwide Express Automated Freight Exceptions with Amazon Bedrock

Worldwide Express, a major player in less‑than‑truckload (LTL) shipping, was losing $15 M of free cash flow each month to unresolved billing exceptions. Manual dispute handling was slow, error‑prone, and required a dedicated FinOps team to chase carriers one‑by‑one. By partnering with AWS and leveraging Amazon Bedrock, the company built a multi‑agent AI system that plugs directly into its Transportation Management System (SpeedShip). Within six months the solution reduced active exceptions by 75 %, accelerated cash‑to‑bank by four weeks, saved $1 M annually, and released a $10 M cash‑flow infusion.

---

The Core Challenge: Complex, Context‑Heavy Exceptions

In LTL freight, about 14 % of invoices fail the three‑way match (purchase order, goods receipt, carrier invoice). Exceptions fall into three buckets:

1. Line‑item mismatches – quantity or service‑code discrepancies.
2. Pricing variances – charges that differ from contracted rates.
3. Unexpected accessorials – lift‑gate, residential, or re‑weigh fees not quoted.

Resolving each case required a five‑step manual workflow: identify the type, research root cause, email the carrier, wait for a reply, and finally update SpeedShip. The variability of carrier tariffs and dispute procedures made pure rule‑based RPA insufficient; the process needed contextual judgment.

---

Designing a Multi‑Agent AI Framework

Architecture Overview

• Amazon Bedrock supplies foundation‑model inference via a single API, ensuring consistent governance and regional availability.
• Amazon EC2 hosts the orchestration layer and the custom agent runtime, allowing fine‑grained control over prompts, decision boundaries, and auditability.
• SpeedShip APIs provide the same data view a human analyst would see, enabling the agents to read contracts, shipment details, and proof‑of‑delivery documents.

Specialized Agents

| Agent | Responsibility | Key Inputs | Output | |-------|----------------|------------|--------| | Invoice Matching | Align invoice lines with shipment records | Shipment manifest, invoice line items | List of mismatched fields | | Pricing Variance | Compare billed rates to contract rates | Contract price list, carrier invoice | Valid/invalid flag with variance amount | | Accessorial Validation | Verify unexpected fees against delivery events | Surcharge policy, delivery proof | Approval or rejection of each fee |

Each agent carries an instruction prompt that defines its role, permissible actions, and required output format. Before making a recommendation, the agent pulls the relevant carrier’s tariff rules and SOPs from a centralized knowledge base, ensuring decisions are grounded in carrier‑specific policy rather than a generic model.

---

Human‑In‑The‑Loop and Continuous Improvement

During the initial rollout, a FinOps analyst reviewed every AI recommendation in the SpeedShip UI before it was committed. As model accuracy climbed above 99 %, bulk approvals were enabled, dramatically reducing manual effort. All reviewer feedback is logged; rejected cases trigger prompt refinements and knowledge‑base updates, creating a feedback loop that steadily expands automation coverage.

---

Observability and Business Impact

Worldwide Express built dashboards that track:

• Automation rate per carrier and exception type
• Error and override rates
• Resolution time from detection to final approval

These metrics revealed a 75 % drop in active exceptions and a four‑week improvement in time‑to‑cash. The financial upside included $1 M in annual savings and a one‑time $10 M cash‑flow release.

---

Practical Takeaways

1. Agent specialization works better than a monolithic model for heterogeneous dispute types.
2. Carrier‑specific knowledge bases keep AI decisions compliant and auditable.
3. Human‑in‑the‑loop is essential for early adoption; automated bulk actions can follow once confidence is proven.
4. Observability—real‑time dashboards and feedback logging—drives continuous improvement and ROI tracking.

By embedding Amazon Bedrock‑powered agents directly into their existing TMS, Worldwide Express turned a costly, manual bottleneck into a scalable, AI‑driven workflow, unlocking significant cash flow and setting a blueprint for other logistics firms seeking similar gains.