There was almost nothing to learn from back then, and four years later that hasn’t changed much. Kaggle has no shortage of credit risk and fraud scoring datasets and notebooks. I couldn’t find one on disputes.

So here’s my theory. Disputes sits right in the middle of product, engineering, analytics and operations. It can’t be filed under any one of them, so it doesn’t get its own career path, its own courses or its own blog posts. This post is my small attempt at a fix: disputes from the issuer’s side, from the analytics seat, with a few ideas on where AI fits.

Why bother?

Because it’s big and getting bigger. Visa says it processed 106 million disputes globally in 2025, up 35% from 2019 [1]. In the US, the CFPB reports that cardholders disputed $9.8 billion in credit card charges in 2024 [2], and the single most common reason on general-purpose cards was a cancelled recurring charge, at roughly 40% of disputes [3].

Almost everything I found on dispute analytics is written for merchants: how to win more chargebacks, which reason codes to fight. The issuer’s side, where the customer calls first and the bank has to decide, gets far less airtime.

The basics, quickly

When a customer has a problem with a charge on their statement, they contact their card issuer. From that moment the bank has to investigate and give an answer, and it’s bound by regulation. For US credit cards, Regulation Z gives the issuer two billing cycles, and no more than 90 days, to resolve a billing error after the customer notifies them [4]. Get it wrong and you’re looking at fines and some very uncomfortable conversations with regulators.

Four parties are involved:

• The issuer (the customer’s bank)
• The merchant whose charge is being disputed
• The acquirer (the merchant’s bank)
• The network (Visa, Mastercard, Amex)

That’s the usual picture. Amex often plays more than one of those roles itself, but let’s not go down that road today.

There are two ways to handle a dispute.

Chargeback. This runs through the card network, and each network has its own rulebook. The issuer files on the customer’s behalf, the merchant can fight it, and if nobody backs down it can go all the way to arbitration, where the network decides who’s liable. It feels a bit like a court case, and it costs money: a few dollars for a routine chargeback, up to several hundred in arbitration fees [5]. So you have to be picky about what you file.

Reasonable investigation. Sometimes the issuer can settle it on its own, without the network. Maybe the merchant already issued a credit. Maybe the claim doesn’t hold up: false claims, someone trying to game the system. If the bank has solid evidence, it can make the call itself.

Two disputes, two very different clocks

Dispute A: “I never got my order”

A customer buys something and it never shows up. They try the merchant, get no reply, and two months after the purchase they call their issuer. It’s still inside the chargeback window, so the bank files a chargeback and waits for the merchant’s response period to run.

If the merchant stays silent, the money comes back through the network, the customer keeps their refund, and the merchant takes the loss. Clean and quick.

It gets slow when the merchant answers with proof of delivery. Now someone has to judge the evidence. If it looks solid, the issuer rebills the customer, which just means the charge goes back on their account. But the customer may have more to give, say a photo showing the exact delivery time and an empty front porch. Then the issuer can go to pre-arbitration, the last negotiation round before the network steps in. The merchant either accepts or pushes back. If it accepts, the customer keeps the refund. If it pushes back, the issuer still has a choice to make: rebill the customer, write the charge off, or take it to arbitration, where the network picks a winner.

Most of the days in this case are spent waiting: for the merchant, then pre-arb, then arbitration.

Dispute B: “I don’t recognize this charge”

A customer calls about a charge from months ago and says they don’t recognize it. That’s past the chargeback window (windows vary by network and reason code), so the network route is closed. All that’s left is the issuer’s own investigation, and it ends one of two ways: rebill the customer, or write the charge off.

The days here go to waiting for the customer’s documents, and to reviewing the case together with the other disputes that came from the same fraud event.

Same facts, different investigator

Here’s the part that bothers me most. In both cases, the result can depend on who picks up the file.

In Dispute A, one investigator sees the proof of delivery and rebills. Another goes back to the customer, asks whether they can send anything more, and ends up with the porch photo that wins the case.

In Dispute B, a customer says their card was stolen and the charge is fraud, but they’ve kept using that same card for months. One investigator declines the claim on that alone. Another looks closer and notices that on that one day, the charge came from a merchant the customer had never used, or an online order started in a different city. That raises suspicion, and the case gets written off in the customer’s favor.

Same case, different outcome. This isn’t about anyone being careless. Guidelines can only be so specific, and thoroughness comes down to experience and time. I’d love to see that variance go away, and that’s exactly where I think AI belongs.

Why it’s hard: four what-ifs

These are made up, but anyone who has worked in this space will recognize the shape of them.

Volume. Say an issuer sees $10B in monthly spend and, to keep the math easy, 0.25% of it ends up disputed. That’s $25M of cases, and the team is staffed for exactly that. Then one month a breach or a viral scam pushes it to 1%. That’s $100M, and every single case still has a regulatory clock running. Is the team ready? Is the budget?

A well-meaning rule. A product team ships automation that resolves in the customer’s favor any dispute under $1, for first-time disputers only. Reasonable, cheap, and it saves investigator time. Now imagine a lot of people finding out about it at once. Rules built for a normal day meet a very abnormal one, and the hit lands in the P&L.

Intake. A system problem means disputes from certain regions don’t get captured for a few days. Customers tried to dispute and the system dropped it. If a regulator finds out, that’s the kind of failure that draws heavy fines. What do you do on Monday morning?

The double credit. A merchant issues a credit to the customer and then rejects the chargeback after the regulatory window has closed. Does the customer end up credited twice? Who absorbs that?

There are a lot more of these kind of cases only an issuer side can bring up.

Who works on disputes?

A quick tour, because the point is how many teams touch a single case. ( feel free to skip if you already know the each teams responsibilities)

• Operations are the backbone. Investigators work each case against its regulatory timeline, including the ones that slip past automation. Their constant problem is capacity against deadlines.
• Engineering builds intake, case systems and data flows. Good data in, good analytics out. Bad data in, well, you know.
• Product sits between ops, engineering, strategy and leadership. They turn ideas into requirements, and every change needs Legal and Compliance sign-off. Every rule change is also a risk change.
• Strategy and analytics build the reporting, find the gaps in the process, and recommend fixes.
• Controls own the regulatory side: building and maintaining the controls that make sure every guideline is met. In disputes, that job is never optional.

Five teams, one case. No wonder it’s hard to say who owns disputes.

Where I’d bet on AI

I’d look at four places, following the life of a case.

1. Intake. An AI agent that answers the call or chat, asks the right questions for the customer’s actual concern, captures their intent accurately and writes a clean summary. Everything downstream depends on getting this step right.
2. Investigation. Two jobs here. First, file chargebacks automatically with the evidence gathered from the customer, asking them for more only when it’s actually needed. Second, propose declining a claim when the evidence supports it: look at the customer’s profile and transaction history, check for patterns that point to first-party fraud, and write up the recommendation in clear language with the proof attached.
3. Reporting. Build the charts for leadership, flag where performance is slipping, and keep comparing against industry trends.
4. Regulatory monitoring. Spot breaches early and propose control improvements before a regulator does it for you.

The thread running through all four is consistency. I want every case to get the same thorough look a great investigator would give on their best day, backed by the same knowledge. Grouping related disputes from the same fraud event, so they’re reviewed together, is another slow step in fraud cases that AI could take over.

Keeping a human in the decision isn’t a courtesy. A decline affects a real customer, and the reasoning has to be explainable to them and to a regulator. The AI does the legwork, and the investigator makes the call: agree and add their own comments, ask for more evidence, go for a chargeback, change the fraud type, flip the fraud indicator, or write it off.

The risk with “AI recommends, human approves” is rubber-stamping, so I’d build a few guardrails in:

What’s already out there

None of this is science fiction, and I’m not the first to think of it. Visa has announced AI tools for issuers, including Dispute Intelligence and a Dispute Doc Analyzer that summarizes merchant documents for reviewers [1]. Airwallex uses an LLM to categorize disputes and fill in its Visa submission forms, and reports 40% less processing time [6]. Vendors like Decisionly sell dispute automation to issuers [8]. Casap describes AI agents covering claim intake, evidence gathering, fraud scoring and chargeback filing for financial institutions [7]. The results these companies publish are their own claims, so I’d read them with a grain of salt, but the direction is clear.

What I find most interesting is the judgment layer: how an AI agent reaches a recommendation, how it explains itself, and how it works alongside an investigator. End-to-end investigation by AI agentsis still early and nowhere near fully developed, and I think that’s where the biggest opening is.

The product I’d build

Picture an issuer plugging into one platform through a few simple MCP connections, instead of a multi-year integration project. From there it handles everything in one place: intake, investigation, resolution, customer feedback and analytics. The honest catch is that to do the job it needs access to the issuer’s own transaction and customer data, so security, permissions and audit trails aren’t features. They’re the ticket to entry.

How I’d know it works

Four numbers I’d watch:

• Recovery rate: how much of the money we’re entitled to get back actually comes back.
• Reassertion rate: how many customers reopen the dispute because they’re unhappy with the outcome.
• CSAT and NPS: whether customers felt they were treated fairly, whatever the result.
• Average days to resolution, measured against the regulatory clock.

A tool that speeds things up but pushes reassertions higher isn’t working. These four have to move together.

Where this leaves me

Disputes is a real problem, with real money, real regulation and a real customer on the other end of every case. It deserves more attention, and I expect it gets it as AI takes on the repetitive parts and leaves the hard calls to people.

If you work in disputes, I’d love to hear which scenarios keep you up at night. If you’re building in this space, I’d like to talk.

Sources

[1] Visa, “Visa Unveils New Services to Modernize Dispute Resolution Process” (April 1, 2026): https://usa.visa.com/about-visa/newsroom/press-releases.releaseId.22261.html

[2] CFPB, The Consumer Credit Card Market (2025 report): https://www.consumerfinance.gov/data-research/research-reports/the-consumer-credit-card-market-2025/

[3] WTOP, “Canceled subscriptions are driving credit card disputes” (reporting CFPB data): https://wtop.com/news/2026/08/canceled-subscriptions-are-driving-credit-card-disputes/

[4] CFPB, Regulation Z, 12 CFR 1026.13 (billing error resolution): https://www.consumerfinance.gov/rules-policy/regulations/1026/13/

[5] Infosys BPM, “Charging up for chargebacks” (arbitration costs): https://www.infosysbpm.com/offerings/business-transformation-services/insights/documents/charging-up-for-chargebacks.pdf

[6] Airwallex, “We’re using AI to process your corporate card disputes faster”: https://www.airwallex.com/ca/blog/ai-corporate-card-dispute-processing

[7] Tearsheet, “How Casap is rethinking dispute resolution with agentic AI”: https://tearsheet.co/awards/how-casap-is-rethinking-dispute-resolution-with-agentic-ai/

[8] Fintech Global, “Episode Six teams with Decisionly on dispute automation”: https://fintech.global/2026/07/07/episode-six-teams-with-decisionly-on-dispute-automation/