In an enterprise climate obsessed with end-to-end automation, the payment dispute pipeline remains one of the few corners of fintech where purely autonomous AI may be an operational liability.
That’s largely because efficiency cannot come at the expense of accuracy. Most commercial AI tools operate with disclaimers warning users that answers may be inaccurate. In chargeback representment, however, platforms have no such margin for error. Dispute decisions are deterministic, governed by strict card network rules, and require evidence mapped precisely to specific reason codes. A flawed automated decision simply means a lost dispute and lost revenue for the merchant.
“Ultimately, we want to bring the right, correct, and compelling evidence to the fore on that particular case. You don’t want efficiency costing you cases or not being deterministic,” said Damo Sampathkumar, Chief Product Officer at Chargeback Gurus, a company focused on reducing fraud and chargebacks. “Most AI tools have disclaimers. We cannot afford to have disclaimers; we are the final stop.”
Rather than relying on end-to-end automation, platforms are finding that AI works best as an auditing tool—identifying incorrect reason codes, navigating blind spots in historical training data and packaging evidence specifically to pass the review systems used on the issuer side.
That dynamic is already apparent on the banking side, where dispute volume contends with federal compliance and relevant deadlines. In the U.S., specific rules for debit and credit issues often require immediate provisional credits while investigations play out.
Bridging the Divide
According to Phil Bruno of payments software provider ACI Worldwide, this pressure has created a divide between routine claims and complex judgment calls.
“A significant portion of issuer chargebacks can be automated and AI is providing greater input and suggestions to the judgment calls; for these “non-trivial” cases requiring judgment calls, a human in the loop is still required,” Bruno said. “The vast majority of chargeback claims are automated with little to no human review. An educated guess would be that about 70% fall into this category, especially for large issuers with the systems and case history — pattern-matching against known fraud signatures is inexpensive and fast.”
Low-value chargebacks rarely involve human review, Bruno said. If a dispute can’t be resolved automatically, the chargeback is simply refunded to the customer. The remaining 30%, which are typically high-value or ambiguous disputes like merchandise issues or unrecognized transactions, are routed to humans. At that point, AI assists the investigation before the case is entered into card network systems.
That split explains why merchant-side platforms would hold back on promising full automation.
While Chargeback Gurus is an AI-driven chargeback platform, the company carefully considers new AI elements before adding them to its tech stack.
“We don’t go by the concept of this is interesting, let’s go do it,” Sampathkumar said. “We look at the practical reasons and how it benefits the client.”
A Human in the Loop
Sampathkumar said Chargeback Gurus must consider many factors before augmenting its technology, beginning with client needs. These often differ by industry, the laws and regulations governing that industry, and a company’s goals and configurations.
“Fundamentally, we believe human in the loop is important here,” Sampathkumar said. “We want to use AI to augment their capabilities, not replace them.”
While there is clear boardroom pressure to adopt AI lest one falls behind, the inherent complexities necessitate that one looks before they leap.
“It’s wishful thinking to automate everything,” Sampathkumar said.
That limitation is made particularly clear when looking at how models are trained. Sampathkumar pointed to a common shortcoming of training AI on recent transaction data. If you’re looking at the last three months and only focusing on cases brought, you’re overtly ignoring a large swath of data. The AI never learns why a merchant chose not to dispute certain chargebacks.
“However great your model and training are, ultimately there are so many conditions around it in the industry that we need to ensure human in the loop is considered,” Sampathkumar said.
For platforms on the merchant side, AI brings clearer value when it comes to auditing and data reconciliation over pure decision-making.
Chief delivery officer Navin Sequeira noted that models can scan huge amounts of data to identify incorrect reason codes. He gave the example of AI being able to detect when evidence of physical, card-present behavior was logged, but the issuer applied a card-not-present reason code. It automatically identifies the claim as invalid; no further evidence required.
“What AI has opened up to us is the ability to synthesize comments coming from the customer, the ability to understand why the issuer has raised these chargebacks,” Sequeira said. “It’s possible to do it manually, but it would take a lot of time.”
The most important thing, according to Sequeira, is to ensure that manual corrections flow back into the model.
“The essential part is to not lose that change that’s being made to an AI-generated dispute package,” Sequeira said. “It’s about how you find the information you put back into the system for learning. That’s the crucial part.”
