About 14% of those invoices didn’t match, the case study said. The line items were wrong, the price differed from the contracted rate or an unexpected fee appeared, such as a liftgate charge. An outsourced team worked the pile one case at a time by phone and email. By June 2025, the unresolved cases held $15 million in frozen cash flow.
Three artificial intelligence agents built on Amazon Bedrock now work those disputes. Within six months, the backlog of active exceptions had fallen from $15 million to $4 million, the case study said. That freed about $10 million in cash flow, cut time to cash by four weeks and saved more than $1 million a year.
Worldwide Express, part of ShipStation Global with about $5 billion in systemwide revenue, earned a 60-fold first-year return on the project, according to the case study. It doesn’t say how that return was calculated.
Three Agents Split the Disputes by Type
Each exception goes to one of three agents, the case study said. An invoice-matching agent finds the shipment record and flags mismatches. A pricing-variance agent compares the carrier’s charge to the contracted rate and decides whether the gap is valid. An accessorial-validation agent checks whether extra fees, such as liftgate or residential delivery charges, match what happened at delivery. Before deciding, each agent reads the carrier’s tariff, surcharge definitions, dispute steps and the company’s variance thresholds. The same three agents cover every carrier without carrier-specific code.
The agents reach the company’s SpeedShip transportation system through the same interfaces its analysts use, the case study said. They see the same contract rates, shipment details and proof of delivery. After each resolution, the system writes a plain-language note on what it found and what it did. At first, human reviewers approved every recommendation. Bulk approval became the default once accuracy passed 99%. Rejections are logged and used to refine the agents’ prompts and reference material.
“With AWS, we’ve automated more than 40% of weekly billing exceptions,” Nathanael Pearson, Worldwide Express’ vice president of FinOps, said in the case study. The results section puts the figure at about 50%, or roughly 10,000 exceptions a week, with an error rate under 1%. The project went from an assessment in June 2025 to production in under six months.
Unilever Automates the Order End of the Same Cycle
At the other end of the order-to-cash cycle, Unilever’s AI agents cut initial order processing from 15 to 20 minutes to about 90 seconds and settle refund claims twice as fast in pilot markets, PYMNTS reported last week.
Middle market finance chiefs are moving the same way. According to the PYMNTS Intelligence report “Time to Cash: What Two Years of Data Say About Faster Cash and Rising Risk,” CFOs now automate about 73% of their accounts receivable processes, up from 55% a year earlier. The report surveyed 100 U.S. CFOs at companies with $250 million to $2.5 billion in annual revenue in July 2026. Integration complexity is the obstacle they cite most to automating the rest, named by 90%, up from 67% in 2025, the report found.
Worldwide Express avoided much of that integration work by plugging its agents into the interfaces its analysts already use, and it is extending the approach to other order-to-cash processes, the case study said. “Today, we release approximately $10 million more in customer billings each week,” Pearson said.
For all PYMNTS AI coverage, subscribe to the daily AI Newsletter.
