AI isn’t all about automating core business functions at Fortune 500 companies. Small and medium-sized businesses can also use AI to optimize, economize, and in some cases compete more effectively against much larger rivals.
An Austin, Texas-based vegan cheese-maker called Rebel Cheese used it to level the playing field against a larger supplier. Specifically, the company developed a small system of AI tools to help it claw back overcharges from a major shipping carrier.
The company is perhaps best known for winning a $750,000 investment from Mark Cuban, money it used to grow Rebel Cheese into what it says is now a $20 million business. Cuban recently spoke about the company’s crafty use of AI on stage at the Convergence AI event in Dallas.
The Problem
Rebel Cheese ships tens of thousands of orders of perishable, handcrafted vegan cheese across the country. The holiday season is by far its busiest period. “Q4 is all hands, heads down, get it out the door, make sure customers are happy,” the company’s cofounder, Kirsten Maitland, wrote in a recent blog post. “There’s no time to stop and analyze anything.”
After this past holiday season, Maitland took a look at the company’s bank account, and something seemed off. Rebel Cheese had just had its best holiday season ever, yet the numbers didn’t reflect it. So she started digging to find out where the profits were leaking away. She discovered that the company had paid $250,000 more for shipping than planned.
Hiring new employees to research and fix the problem wasn’t in the cards. So Maitland turned to Anthropic’s Claude. “I handed it a year of invoices and a contract,” she tells Fast Company in an email exchange, “and it found patterns I would have needed a forensic accountant to surface, which would have been time-consuming and expensive.”
She says the carrier’s shipping invoices run hundreds of pages per week, with fees layered inside fees. “Most shippers don’t have the time or tools to audit them,” Maitland says. For the carrier, the complexity was not a bug but a feature, and a profitable one.
Her analysis turned up several causes, not just one. “Some were our fault, like significant weight overages on our packages, which we could fix,” she says. “The rest were on the carrier: They had put a custom contract in place for us. Under that contract, any package bulging or weight overage triggered drastic price spikes.”
By the time Maitland sat down with representatives from the shipping company, she had analyzed a year’s worth of data and could show them exactly which contract clauses were doing the damage. The biggest overcharges mapped to a new weight limit the carrier had implemented, but not communicated, in early 2025. Their response was: “Well, you should have caught it.” She vowed never to let that happen again.
The build
To build the actual programs that read the invoices and request refunds, Maitland used Manus, an AI orchestration layer that coordinates work among various agents and subagents, using different models for different tasks. Maitland says she also tested Bolt, Lovable, and Relay, but found that Manus handled the job more easily and accurately.
After a lot of experimentation and discovery, she was able to architect a system that automated a data-heavy auditing process that had previously required manual review of tens of thousands of shipments. The build unfolded in four distinct phases:
1. Standardizing the “Truth.” The process began with data preparation. Maitland created two simple comma-separated value (CSV) templates. A “Zone Data File” contained Rebel Cheese’s negotiated contract rates, while a “Transaction File” contained weekly invoice data. This gave the AI a clear structure for comparing “what we should pay” against “what we were actually billed.”
2. Designing the Blueprint. She uploaded example invoices, Rebel Cheese’s carrier contract, and a presentation detailing the results of her Claude-assisted investigation into the overcharges to Manus. Rather than immediately asking the AI to “build a tool,” Maitland first used it to generate a comprehensive “Requirements and Design Document.” The document served as a technical blueprint, laying out the business logic for “fuzzy weight matching” and methods for flagging discrepancies. The step ensured the AI understood edge cases like fuel surcharges and weight brackets before a single line of code was written.
3. Building via Orchestration. Maitland then asked Manus to build a tool based on the blueprint document. Her prompt began: “I need you to build a standalone, single-page web application that acts as a Carrier Billing Discrepancy Detection Tool (works for any carrier — UPS, FedEx, USPS, or your specific shipping partner).” She stipulated that the tool should flag every shipment where the actual charge exceeded the contracted rate by more than ten cents. Those overcharges would then be sent to the carrier alongside a request for credit for every discrepancy it could not justify.
4. Continuous Auditing and Strategic Insights. Once the tool flags overcharges, Maitland feeds the data back into Claude, which analyzes the logs for higher-level patterns, such as shipping zones where costs are spiking. That transformed the tool from a simple invoice checker into a permanent recovery system.
The Result
The system Maitland built audits all shipping charges by comparing carrier invoices against Rebel Cheese’s contracted rates. It flags every discrepancy and generates a report that’s sent directly to the shipping carrier, requesting credit for every overcharge that can’t be justified. Maitland says the carrier has approved and credited every claim the system has submitted so far. She pays about $200 per month for her Claude and Manus subscriptions, and says the company is now saving between $1,500 and $4,000 every week.
Now that Rebel Cheese has gained experience with, and trust in, AI automation, the company is already using the technology for other core business functions, Maitland says. She built an agent that monitors the fundraising pipeline, researches VCs, and prepares her for investor meetings. She also built a site that handles inbound donations and partnership requests. Another tool uses historical data to draft ads.
“The bigger shift was realizing this is what closes the gap for companies our size,” she says. “We don’t have an engineering team. We don’t have a data analytics team. A few years ago, I would have had to hire a consultant . . . now I can do the work myself in an afternoon.”
