Small charges can be easy to miss, and challenging them can take more effort than the amount seems worth. DoNotPay is exploring whether an AI system can find those costs and take steps to recover money or stop unnecessary payments.
In a test described by the company’s CEO, Joshua Browder, GPT-4 and Auto-GPT reviewed personal financial information and handled tasks ranging from subscription cancellations to refund requests. The examples show how automation could turn a forgotten charge into an action, with consequences for both consumers and the organizations receiving their claims.
Looking for money that slips through the cracks
DoNotPay describes itself as “the world’s first robot lawyer.” Its platform is aimed at practical tasks such as appealing parking tickets, canceling auto-renewing subscriptions and seeking other remedies. One feature focuses on “hidden money”: payments a person may be owed, or recurring costs they may no longer want, that go unnoticed or seem too time-consuming to pursue.
Browder said he gave Auto-GPT access to his bank accounts, credit reports and emails, then assigned it a broad instruction: “find me money.” In the reported test, GPT-4 had found about $218 so far. About $81 of that total came from subscriptions the system identified as unnecessary and offered to cancel.
The figures describe one person’s trial, rather than a general result for users. Still, they illustrate the idea behind the tool: an AI assistant could search across records and make it easier to act on small, dispersed expenses that might otherwise be ignored.
From identifying a charge to canceling it
Finding a subscription is only one step. The system was also designed to carry out some of the follow-up. For gym memberships, the article says it could send digitally signed cancellation letters through a U.S. Postal Service API, or communicate online with customer service representatives and submit a digital cancellation.
This changes the role of a finance assistant. Instead of simply pointing out a recurring payment, an automated system may be able to contact the business and complete the request. That can reduce the effort required from a customer, especially when the amount at stake is too small to justify a long manual process.
DoNotPay was developing a chat system using GPT-4 and Auto-GPT for its platform. The planned access points described in the article were a ChatGPT plugin on the DoNotPay website and iMessage. Those channels would bring the process into familiar interfaces, though the examples remain tied to the company’s test and plans.
Refund requests and tougher negotiations
The test went beyond recurring charges. Auto-GPT found a bill of about $37 for in-flight Wi-Fi on a flight from London to New York. After asking Browder whether the service had worked properly, the system drafted a refund request when he said it had not. A bot sent the letter to the airline, and Browder said the money was returned within 48 hours.
The account suggests a sequence that AI could handle: notice a charge, ask a relevant question, prepare a complaint and send it. Browder also said GPT-4 negotiated with Comcast after outages, securing a $100 refund and a 20 percent discount on his bill for the next three months. He contrasted that with GPT-3, which he said had been more accommodating during negotiations.
These examples make the potential value concrete. A customer does not have to remember every charge or compose every message from scratch. If a system can identify a reasonable claim and handle the communication, more people may pursue refunds or cancellations that they would otherwise leave alone.
More claims could change the customer service burden
The same automation could increase the volume of disputes facing businesses and government agencies. If people can use AI to prepare appeals, complaints, negotiations or lawsuits with less effort, organizations may receive more requests. A process that once screened claims through the time and persistence needed to make them could become easier to repeat at scale.
The article points to a possible response: organizations may also use AI systems to answer or contest claims. That could set up a cycle in which automated tools make it easier to challenge decisions, while other automated tools handle the resulting requests.
For consumers, the promise is practical: overlooked payments may become easier to find, and routine follow-up may take less work. But the reported outcomes come from Browder’s test, and they do not establish how the system would perform across other accounts or disputes. The larger question is how companies and agencies will respond if pursuing a small claim becomes as simple as asking an assistant to do it.