An AI agent can handle a payment dispute end to end, provided end to end means intake, evidence gathering, deadline tracking and customer communication, with a human or an approved rules engine making the decision. The clocks are fixed by regulation: under Regulation E a bank must determine whether an error occurred within 10 business days, or provisionally credit the customer and take up to 45 days. The workload justifies the build: US cardholders disputed $9.8 billion in credit card charges in 2024, and 40 percent of general purpose card disputes were cancelled recurring transactions, the kind of case an agent can take cleanly.
Key takeaways
Automate intake, classification, evidence collection, deadline tracking and status updates. Keep provisional credit decisions, fraud determinations and the final outcome with a human or a rules engine your compliance team approved.
Reg E: 10 business days to investigate, 45 days with provisional credit, 90 days for point-of-sale and foreign-initiated transfers. Reg Z: acknowledge a billing error within 30 days and resolve it within two complete billing cycles, never later than 90 days.
Merchants get 7 to 21 days to respond to a chargeback depending on the card network, and a missed deadline is an automatic loss, per Stripe's dispute documentation.
Decision rule: if a regulator will audit a date to the day, compute it with deterministic code. Let the language model decide what the customer is disputing and gather the evidence.
Scale: $9.8 billion disputed on US credit cards in 2024 (CFPB), and global chargeback volume forecast to reach 324 million in 2028 (Datos Insights with Mastercard).
What does "end to end" honestly mean for an AI dispute agent?
End to end means the agent owns every step the customer sees and every handoff between your systems, and it never decides the outcome.
Intake. Recognise the dispute, verify identity, identify the transaction, ask the questions that set the dispute type, and record the notice date that starts the clock.
Orchestration. Open the case, map it to a reason category, request the evidence that category needs, pull account history, and route the file to the right queue.
Status and communication. Tell the customer what happens next and by when, answer "where is my dispute" on any channel, send required notices, and chase missing information.
Closure. Once a person or an approved rule decides, explain the result, reverse or confirm any provisional credit, and seal the record.
Missing from that list is adjudication. The agent never forms an opinion on whether the dispute should be upheld. It gathers what the decision needs and hands over a complete file. That boundary is what a compliance team can approve, and the one most vendor pages blur.
Which side of the transaction you sit on also matters. A bank or fintech handles its own customer's claim under Reg E or Reg Z. A merchant receives a chargeback from the network and decides whether to fight it. The sections below cover the issuer side first, where the regulatory deadlines live, then the merchant side.
Which deadlines govern a dispute, and who sets them?
Three rule sets apply: Regulation E for debit cards and electronic fund transfers, Regulation Z for credit card billing errors, and card network rules for the chargeback itself. The same transaction can run two clocks at once.
Rule | What starts the clock | Deadline |
|---|---|---|
Reg E, consumer notice | Periodic statement sent | 60 days for the consumer to report the error |
Reg E, investigation | Notice of error received | 10 business days to determine whether an error occurred |
Reg E, extended investigation | Provisional credit issued within 10 business days | Up to 45 days from receipt of notice |
Reg E, special cases | Account opened within 30 days; point-of-sale or foreign-initiated transfer | 20 business days replaces 10; 90 days replaces 45 |
Reg Z, consumer notice | First statement reflecting the billing error | 60 days for the consumer to notify the creditor |
Reg Z, acknowledgment | Billing error notice received | Written acknowledgment within 30 days |
Reg Z, resolution | Billing error notice received | Two complete billing cycles, never later than 90 days |
Card network, merchant response | Chargeback notification | 7 to 21 days depending on the network; no response means the merchant loses |
The Reg E figures come from 12 CFR 1005.11, the Reg Z figures from 12 CFR 1026.13, and the network window from Stripe's dispute documentation. Two details trip up automation projects. Reg E counts business days, so the agent needs a bank-holiday calendar. And the clock starts at notice, not at completeness. The CFPB's electronic fund transfer FAQs state that an institution "may not delay initiating or completing an investigation pending receipt of information from the consumer". An agent that holds a case open until the customer uploads a receipt is building a compliance problem, not solving one.
How big is the dispute workload, and what is at stake per case?
Disputes are steady, high-volume work with a few hundred dollars riding on each case. The CFPB's 2025 Consumer Credit Card Market Report found that in 2024 cardholders disputed $9.8 billion in credit card charges, resulting in $5.9 billion in chargebacks. Consumers disputed about $3.03 of every $1,000 spent and received $1.63 back. The average credit on a general purpose card dispute was $197.
The mix matters more than the total. For general purpose cards, cancelled recurring transactions made up 40 percent of disputes in 2024, "services not received" 21 percent, and "cancelled/credit not issued" 18 percent. Four in five disputes are therefore about subscriptions, cancellations and undelivered services, where the facts sit in merchant records and account history, and an agent can document them in full. Unauthorised-use claims are the smaller share where a human fraud reviewer stays in the loop.
Globally, the trend is up. Datos Insights, working with Mastercard, expects chargeback volume to reach 324 million in 2028, with a total value of $41.69 billion, and estimates that fraudulent chargebacks, including first-party fraud, account for 45 percent of volume.
One honest gap: there is no reliable public figure for the fully loaded cost of handling a dispute by hand. Use your own handle time and analyst cost, and count the status-check contacts each dispute generates, which are often the larger cost.
What should the AI decide, and what should code compute?
Split the work by how it will be audited. A language model is good at reading a messy message and working out what the customer is actually disputing. It is the wrong tool for producing a date a regulator will check. The working rule: if a regulator will audit it to the day, it should be computed.
The model decides:
Whether the message is a dispute at all, or a question about an unfamiliar merchant descriptor that an explanation will close.
Which category applies: unauthorised use, duplicate charge, cancelled recurring payment, goods or services not received, amount differs from receipt.
Which transaction the customer means when they describe it loosely, what follow-up questions settle it, what evidence the category needs, and how to summarise the file for the analyst.
Deterministic code computes:
The notice date and time, stamped from the system clock, not extracted from text.
The 10-business-day determination deadline against a bank-holiday calendar, and whether the account is new enough to move to 20 business days.
Whether the transaction is point-of-sale or foreign-initiated, which changes 45 days to 90, read from the transaction record rather than the conversation.
The provisional credit due date, the Reg Z 30-day acknowledgment date and the 90-day outer limit, and which notice template is due at each step.
Threshold checks: dollar amount, prior dispute count, anything else that forces human sign-off.
The halves connect through a structured handoff. The model produces a classification with required fields; the code rejects it if a field is missing and sends the agent back to ask. Every computed value is logged with its inputs and rule version, so an examiner can see that case 4471 got the same arithmetic as case 4470. A generated date can be right 99 times and wrong once, and the once is the one you get asked about.
The same split governs what the agent says. The safe statement is "we have opened an investigation and will complete it, or give you a provisional credit, by 18 September". The prohibited statements are structural, not tone guidance: no promise of an outcome, no confirmation that a charge is fraud, no description of provisional credit as a refund. Those belong in outbound guardrails that block the message before it sends, and in pre-launch simulations that try to extract the promise.
Can an AI agent automate merchant-side chargeback representment?
Partly, and the honest end point is submission, not the win. On the merchant side the issuer has already filed the chargeback. Stripe's documentation is blunt: you have "usually 7 to 21 days, depending on the card network" to respond, you get one opportunity to submit, and if you miss the deadline you automatically lose. Stripe's own Smart Disputes feature automates evidence collection and submission for eligible disputes.
An agent's job here is to read the reason code, pull the order, delivery, login and customer-communication records that code requires, assemble them in the processor's format, and submit before the window closes. Fight or accept is a rule the merchant sets: below a dollar threshold, accept; above it, fight when the evidence checklist is complete; refer the rest to a person. The issuer decides who wins, so no merchant-side tool can honestly claim to resolve a chargeback end to end.
The larger opportunity for a merchant is upstream. Cancelled recurring transactions are 40 percent of general purpose card disputes. A support agent that can cancel the subscription, confirm it in writing and issue the refund inside the conversation removes the reason to call the bank. That is a support automation problem, covered in how to let AI take actions in backend systems safely.
What this looks like in practice: a worked example
A customer of a US fintech opens the app chat at 9.40pm and types "there's a charge from a company I've never heard of". Here is how the flow runs on Lorikeet, which combines a natural-language workflow with deterministic structured workflows in one interaction.
Intake. The natural-language workflow recognises a possible dispute, runs the fintech's own identity step-up, and pulls the last 30 days of transactions. It asks whether the customer recognises the merchant under its trading name and whether they authorised a related purchase. They did not. The agent classifies a possible unauthorised transaction and records the notice timestamp.
Computation. A structured workflow reads the account open date and the point-of-sale flag from the transaction record, runs the business-day calendar, and returns the determination deadline and the provisional credit due date, written to the case file with their inputs.
Communication. The agent gives the case number, the date by which the fintech will complete the investigation or provisionally credit the account, and confirms no police report is needed for the investigation to begin. An outbound guardrail blocks a draft containing "you'll get this back" and the agent restates it as "the investigation will decide the outcome".
Orchestration. The case opens in the dispute platform with the classification, the computed dates and a card-block request. Because the category is unauthorised use, a human fraud reviewer is assigned. Where a category calls for merchant contact, a Team of Agents sub-agent can email or call the merchant and post the result to the case.
Status. Three days later the customer phones. The voice agent, on the same case, gives the status and the same dates with sub-second responses, without asking the customer to repeat anything.
Decision and closure. The reviewer approves provisional credit on day six. The agent sends the notice, explains the credit is provisional and reversible if the charge proves valid, and later communicates the final decision. Coach reviews 100 percent of dispute conversations afterwards.
Published proof that regulated, action-taking workflows hold up: Carmoola, an FCA-regulated UK car finance company, resolves 60 percent of inbound conversations end to end on Lorikeet, and TapTapSend supports remittance customers across 35+ send and 80+ receive markets in 15 languages, which is what intake looks like for a global customer base. Pricing is per resolution, about $0.80 for chat, email or SMS and about $1.20 for voice, unresolved or unsatisfactory tickets costing nothing; see pricing. Security posture, including SOC 2, PII redaction, RBAC and US, AU and UK data residency, is on the trust page.
The limitation to state plainly: Lorikeet does not adjudicate disputes and does not file chargebacks with the card networks. It runs intake, computation, orchestration and communication, and hands a complete file to your dispute system and your analysts. A tool that scores evidence against scheme rules and raises the case with the networks is a back-office dispute product and a separate purchase. To see the split applied to your own reason categories, book a demo and bring your hardest cases. More on the regulatory side is in Reg E dispute compliance for AI and the financial services page.
What still needs a human
The list is short and should stay short, because each item carries money movement or regulatory liability.
Provisional credit above a threshold, and every denial. The agent prepares the recommendation and executes after approval.
Fraud and account takeover determinations. A pattern of unauthorised claims, a recent credential change or a device mismatch goes to a fraud reviewer with the file attached.
Represent or concede on the merchant side above the threshold. Spending the single submission is a judgment call.
Anything trending towards a complaint or a regulator enquiry. An adverse decision the customer contests routes to a named owner.
Mixed cases. Part unauthorised use, part billing error: a person confirms the category before filing, because the wrong category loses a winnable case.
The rules themselves. Someone owns the business-day calendar, the thresholds and the notice templates, and reviews the QA output from Coach for drift.
Automating disputes well means taking the intake, the chasing, the date arithmetic and the status replies off the analysts and leaving them the calls that carry liability. The test of any agent is whether it can show you, for one case, which steps the model took, which values the code computed, and where a person signed off.
