Worldwide Express hands disputed invoices to AI agents and cuts unresolved disputes by 75%

Worldwide Express, a US freight brokerage with about $5 billion in annual revenue, 130,000 customers and 75 LTL carriers (less-than-truckload, where one truck carries shipments from several customers), has handed disputed-invoice resolution to AI agents. The problem was mundane. About 14% of carrier invoices fail "three-way matching": an invoice has to agree with both the order and the proof of delivery.

Every failed invoice becomes a dispute, and until the dispute is closed, the money sits idle. About $15 million of the company's free cash flow was stuck in these "exceptions." With AWS and the integrator LTM, it built a multi-agent system on Amazon Bedrock and embedded it in its transportation management system, SpeedShip.

Six months later, the volume of unresolved disputes had fallen to $4 million.

Why carrier invoices don't match

Disputes fall into three types:

• Line-item mismatches. The quantity or service code on the invoice does not match the shipment record.
• Price variances. The carrier billed a rate different from the contracted one.
• Unexpected accessorial charges. Liftgate, residential delivery, reweighing. These charges were not in the original quote.

Previously, every case went through five steps: identify the type, find the cause in the vendor's records, write to the carrier or the warehouse, wait for a reply, enter the resolution in the system. An outsourced team did the work case by case, with phone calls. Each carrier has its own rates and its own dispute procedure, so the financial operations (FinOps) department was drowning in routine.

Classic robotic process automation (RPA) is a poor fit here: it works well for rigid scripts. This job requires reading a carrier's policy, matching a charge against what actually happened at delivery, and deciding whether the charge is legitimate for the amount billed. That is work with text and context, so the company chose LLM-based agents.

Three agents and a rulebook for each carrier

A dispute goes to one of three specialized agents:

• Invoice Matching Agent finds the shipment record the invoice refers to and flags what does not match.
• Pricing Variance Agent compares the billed price with the contracted one and decides whether the difference is justified.
• Accessorial Validation Agent checks whether additional charges are backed by real delivery events: whether a liftgate was used, whether the address was residential.

Each agent has instructions (role, decision boundaries, response format), a set of actions in SpeedShip and access to a knowledge base. The knowledge base holds each carrier's tariff rules, charge definitions, dispute procedures and permitted variance thresholds. Before making a decision, the agent pulls the documents for that specific carrier. That is why the same three agents work with all 40+ connected carriers, and no separate code is written for each one.

The engineering choices are sensible. The company built the orchestration itself on Amazon EC2 servers with autoscaling, and uses Bedrock only to call models through a single API. The routing logic stays with the team, and open-source agent libraries can be used. The agents reach SpeedShip through the same APIs as human analysts, so they see the same contract rates, shipment data and delivery confirmations. After a decision, the system writes a clear note: what it found and what it did. For disputes over money, that kind of explainability is mandatory.

Humans were not removed from the loop. At first, an employee confirmed every recommendation in a familiar interface. Once accuracy exceeded 99%, bulk confirmation became the main mode. Rejections are logged, and the team uses them to fix the instructions and knowledge bases. The dashboards track automation rate by carrier, errors by dispute type, resolution time and how often humans override the agent.

The numbers, and what to double-check

Reported results over six months:

• The agents close about 10,000 disputes a week, roughly half of the flow. The error rate is under 1%.
• Unresolved disputes fell from $15 million (June 2025) to $4 million.
• The order-to-cash cycle shortened by 4 weeks.
• Savings are over $1 million a year, the claimed first-year return is 60x, and the one-time cash-flow release is $10 million.

This comes from a post on AWS's own blog, and the numbers in it do not fully add up. The automation rate is given both as "about 50%" and as "more than 40%." The $10 million gain is described both as a one-time release of cash and as a weekly volume of invoices processed faster. The drop from $15 million to $4 million is actually 73%, not exactly 75%. The post names neither the model under the hood nor the cost of the calls, even though the project's economics are driven mainly by that cost. An "error" apparently means something the human reviewer rejected, so errors the reviewer missed during bulk confirmation are not covered by that figure.

Still, the approach looks transferable to logistics and to any company with a steady flow of invoices. Start with the most common dispute type and prove accuracy on it. Store counterparty rules as searchable documents rather than hard-coding them. Let the agent use the same APIs as humans, and keep a human confirming decisions until the metrics show stable accuracy. The company is already extending the approach to other parts of the order-to-cash cycle.