AI use case

AI agent for card dispute intake

A customer facing AI agent that handles the "I do not recognise this charge" moment: it finds the transaction, separates suspected fraud from merchant disputes and simple confusion, explains the customer's rights and timelines, collects the details and evidence the rules require, and opens a correctly classified dispute case for the operations team.

By Len Debets · Last verified 27 September 2026 · 3 public deployments

2.3 million
Interactions handled
Klarna (organization claim).
USD 12,000 to USD 216,000
Indicative value per year
A card issuer with 1 million active cards. Worked example, see how it is calculated.

What problem does it solve?

Card disputes keep rising, and every one of them starts with a customer who is worried about money. Many are not fraud at all: Visa's Andrew Torre told CNBC that many disputes start with a cardholder who does not recognise a charge on the statement. Others are genuine fraud that needs the card blocked now, and others again are merchant problems (goods not received, a refund that never arrived) that follow the card scheme's dispute rules.

Intake is where much of the later cost and risk starts. A dispute filed under the wrong category, or without the details the network requires, can bounce between teams, miss deadlines and end in a write off. Timelines and refunds are regulated: in the US by Regulation E for debit cards and Regulation Z for credit cards, and in the EU by PSD2, which requires a refund of an unauthorised payment by the end of the following business day unless the provider has reasonable grounds to suspect fraud. A vague or wrong promise to the customer therefore becomes a compliance issue. The US Consumer Financial Protection Bureau has also warned that chatbots and highly scripted representatives may only recognise a dispute when the customer uses specific words.

How does it work?

  1. Recognise the dispute in the customer's own words. "I never bought this", "they charged me twice" and "my refund never came" are all disputes, whatever the phrasing or language.
  2. Find the transaction. In an authenticated session the agent lists recent transactions and enriches the one in question with the merchant's clear name, location and order details. Because many disputes start with a charge the cardholder does not recognise, some of them can be resolved right here, without a case.
  3. Triage. The agent separates three paths: unauthorised use (block the card, fraud claim), an authorised purchase that went wrong (merchant dispute), or confusion (explain and close). Signs of a scam, where the customer was tricked into paying, go to the scam team instead.
  4. Collect what the rules need. It asks only the questions the chosen path requires (dates, contact with the merchant, cancellation proof) and accepts receipts and screenshots, which document AI reads into structured fields.
  5. Set expectations from rules, not from the model. Timelines, provisional credit and next steps come from a deterministic rules engine per product and market, so the promise is right and auditable.
  6. Open the case. The agent creates the case in the dispute system with a suggested reason category and the evidence attached, for an analyst to confirm, and gives the customer a reference number.
  7. Hand over when it matters. Vulnerable customers, hardship, high values, repeat disputes and anything the agent cannot classify go to a human with the full summary.
Audience
Customer facing
Autonomy
Supervised agent
Adoption
Early adopters
Channels
Mobile app, Web chat, Phone and voice, WhatsApp

What is it worth?

Benchmarks are computed from the public deployments below: one data point per organization per KPI, with who made each claim.

Value benchmarks for AI agent for card dispute intake
KPIMedianReported rangeData pointsClaimed by
Interactions handledNot pooled
2.3 million
11 organization

Value drivers: Lower cost to serve, Customer experience, Compliance quality, Risk and loss reduction.

Indicative value

A card issuer with 1 million active cards

USD 12,000 to USD 216,000

Dispute intake and rework cost avoided per year

How this is calculated

Formula: activeCards * disputesPerCard * agentShare * minutesSaved * costPerMinute. The low scenario uses every low input, the high scenario every high input.

InputLowHighBasis
Active cards activeCards, cards1,000,0001,000,000The reference issuer.
Disputes raised per card per year disputesPerCard, disputes per card per year0.010.03Editorial assumption, replace with your own dispute volume.
Share of disputes started with the agent agentShare, fraction of disputes0.30.6Editorial assumption; depends on how prominent the digital route is in the app and on the phone menu.
Intake and rework minutes saved per dispute minutesSaved, minutes per dispute510Editorial assumption. Quavo reports that its clients cut handle time per assignment by nearly 30% (vendor claim, grade D), which supports a modest saving.
Fully loaded cost of a dispute agent minute costPerMinute, USD per minute0.81.2Editorial assumption, replace with your own fully loaded cost.

What it leaves out: Counts intake and rework time only. It leaves out write offs avoided through correct classification and on time filing, disputes prevented by explaining unfamiliar charges, the cost of running the AI and the integration work.

Who already uses it?

3 public deployments, strongest evidence first. Grades: A regulator or audit, B the organization itself, C vendor case study, D anonymous or estimate.

Visa

United States · Payments and cards · 2026

ProductionGrade B

In April 2026 Visa announced six new and enhanced dispute resolution tools. For issuers and acquirers they include Dispute Intelligence (predictive models that support case by case decisions, generally available), Dispute Doc Analyzer (AI summaries of merchant documents for issuer analysts and auto populated questionnaires for acquirers) and Visa Dispute Case Manager, which unifies dispute workflows from intake to resolution. Merchant tools cover pre dispute handling, generative AI representment responses and Compelling Evidence 3.0 to reduce friendly fraud. Several tools are still in pilot or planned for late 2026; no outcome figures were disclosed.

No outcome disclosed.

Klarna

Sweden · Payments and cards · 2024

ScaledGrade B

Klarna announced in February 2024 that its AI assistant built on OpenAI models had been live globally for a month as the first line of its customer service, handling refunds, returns, payment issues, cancellations and disputes in more than 35 languages across 23 markets. In 2025 the company said it had gone too far in replacing people and began recruiting human agents again so that customers can always reach a person; the assistant still handles the majority of inquiries. The record is useful precisely because it shows both the gain and the correction.

  • Interactions handled: 2.3 million, first month after launch
    "The AI assistant has had 2.3 million conversations, two-thirds of Klarna’s customer service chats"
    Claimed by: organization
  • Response time reduction: 82%, since launch, as reported in 2025
    "Since launch, response times have improved by 82%, and Klarna has seen a 25% drop in repeat issues."
    Claimed by: organization

Commonwealth Bank of Australia

Australia · Banking · 2024

ScaledGrade C

Commonwealth Bank built a central AI orchestration agent that reads the customer's intent and routes it to a conversational AI, retrieval over public content, a deterministic guarded path for regulated journeys such as fraud disputes, or a human specialist with the full context, on its messaging channel. It migrated nearly 700 chatbot topics and launched a generative AI banking chatbot in November 2024. Voice bots are a planned extension of the orchestration layer.

  • Containment rate: about 84.6%, May 2026, self service messaging
    "In May 2026, approximately 84.6% of self-service messaging interactions were resolved end-to-end in the messaging channel."
    Claimed by: vendor

How do you implement it?

A model agnostic playbook: what to prepare, the order to build in, and what goes wrong.

Data you need

  • Transaction history with enriched merchant names, locations and, where available, order details
  • Dispute rules per product and market (categories, required information, timelines, provisional credit)
  • Historical disputes with outcomes, to test triage and the suggested categories
  • Approved customer wording for rights, timelines and next steps

Systems to integrate

  • Card management and core banking (transactions, card block and reissue)
  • Dispute or case management system (case creation, status, documents)
  • Merchant data enrichment and network dispute services where the issuer uses them
  • Identity and step up authentication
  • Contact centre platform for handover with context

Complexity: Medium

The conversation is the easy part. The work is in reading transactions and merchant data in real time, encoding timeline and provisional credit rules per product and market, and writing clean cases into the dispute system that analysts trust.

  1. 1

    Map the intake paths

    From last year's disputes, list the real paths (fraud, not received, not as described, duplicate, cancelled subscription, unrecognised) with the questions and documents each one needs. This becomes the agent's playbook and your test set.

  2. 2

    Put the rules outside the model

    Encode timelines, provisional credit and eligibility per product and market in a rules service. The agent calls it and repeats its answer; it never works out a deadline itself.

  3. 3

    Resolve confusion first

    Show the merchant's clear name, logo, location and order details before offering a dispute. Visa's Order Insight service rests on the same idea: surfacing transaction details clears up confusion over legitimate charges before they become disputes.

  4. 4

    Make the case analyst ready

    Agree with dispute operations exactly which fields and documents a case needs, and have the agent suggest a category with its reasoning for the analyst to confirm, not apply.

  5. 5

    Route fraud and scams separately

    Unauthorised use triggers a card block and the fraud path; a customer who was tricked into paying goes to the scam team, because the rules and the customer's rights differ.

  6. 6

    Pilot on one product and one channel

    Start with debit or credit cards in the logged in app, compare case quality and rework against the human channel for a month, then widen.

Guardrails

  • Timelines, provisional credit and eligibility come only from the rules service, never from the model
  • Suggested reason categories are confirmed by an analyst before the case is filed with the network
  • Card numbers are tokenized and personal data masked before any text reaches a model
  • Automatic handover on vulnerability signals, hardship, suspected scams and high values
  • The agent never tells a customer a dispute will succeed

KPIs to instrument

  • Share of disputes resolved without a case (confusion explained), with repeat contact within 30 days counted as not resolved
  • Case completeness at first submission and analyst rework rate
  • Accuracy of suggested categories against the analyst's final category
  • Time from first message to case opened
  • Customer satisfaction on dispute conversations versus the phone channel

Human in the loop

Dispute analysts confirm the category, decide on provisional credit where rules leave room, and own every case once it is opened. Specialists take over for scam victims, vulnerable customers and complaints, and a weekly sample of agent opened cases is reviewed for completeness and correct triage.

Common failure modes

Wrong promises on timelines or credit
The model paraphrases a policy and gets a date or an amount wrong. Keep all such statements in the rules service and test them on every change.
Scam victims treated as disputes
A customer who authorised a payment under false pretences is pushed into a fraud chargeback that is unlikely to succeed. Train triage on scam signals and route them to specialists.
Dispute not recognised
The customer describes a problem in words the agent treats as a question, and no dispute is opened. The CFPB names this as a legal risk; test with real, messy phrasing.
Friendly fraud made easier
A frictionless flow invites false claims. Surface repeat patterns to analysts and keep evidence requirements proportionate to value.

What are the risks and rules?

EU AI Act

Limited risk (transparency)

A customer facing assistant must tell people they are interacting with an AI system (Article 50(1)). It triages and opens cases but does not evaluate creditworthiness (Annex III point 5(b), which in any case excludes systems used to detect financial fraud) or decide access to an essential service, so it is not high risk under Annex III.

Guidance

Controls to put in place

  • AI disclosure at the start of the conversation and a clear route to a person
  • Versioned rules service for timelines and credit, with change control and tests
  • Audit trail of every statement about rights, timelines and credit made to the customer
  • Analyst confirmation of the reason category before network filing
  • Monitoring of dispute outcomes and complaints by channel

Frequently asked questions

Can an AI agent decide who gets a chargeback?
It should not. The agent collects facts and suggests a category; an analyst confirms it and the card scheme rules decide the outcome. The issuer tools Visa announced in April 2026 follow the same split, with AI supporting analysts through predictions and document summaries rather than deciding cases.
How is this different from chargeback automation in the back office?
Intake is the customer conversation that creates the case. Chargeback and representment operations work the case afterwards: evidence packages, network filings and deadlines. Good intake makes the back office cheaper because cases arrive complete and correctly classified.
Do AI assistants already handle disputes in production?
Yes. Klarna says its AI assistant handles disputes alongside refunds and returns, and Commonwealth Bank routes fraud disputes through a guarded, deterministic path inside its AI orchestration. Published results cover customer service as a whole, not disputes alone.

How to cite this page

Blits.ai AI Use Case Library, "AI agent for card dispute intake", last verified 27 September 2026, https://www.blits.ai/ai-use-cases/card-dispute-and-chargeback-intake. Licensed under CC BY 4.0. Method: how we verify use cases.

Changelog
  • 27 September 2026: First published

Related use cases

Payments and cardsBanking

AI for chargeback and representment operations

AI that runs the dispute engine room for issuers, acquirers and merchants: it maps each dispute to the network reason code, gathers the matching evidence, assembles a network compliant chargeback or representment package, drafts the rebuttal, tracks every deadline and processes pre dispute alerts so a refund can be issued before a chargeback lands.

Deployments
2 public, best grade B
Autonomy
Supervised agent
BankingPayments and cards

AI agent for fraud alert confirmation with cardholders

A customer facing AI agent that contacts the cardholder as soon as the fraud engine flags a card transaction, in the channel they actually respond to, verifies them, asks whether they made the transaction and acts on the answer: releasing the block so a retry succeeds, or freezing the card and starting the fraud claim.

Deployments
5 public, best grade B
Reported fraud loss reduction
76%
Commonwealth Bank of Australia, organization claim
BankingPayments and cards

AI agent for account and card servicing

An AI agent that resolves routine account and card requests end to end, such as balances, statements, card blocks and replacements, PIN resets and limit changes, across app, web, messaging and phone, and hands anything sensitive or unusual to a human with the full context.

Deployments
2 public, best grade B
Reported containment rate
about 90%
DBS Bank, organization claim
BankingPayments and cards

AI scam intervention for instant payments

AI that talks to the customer when they are about to authorise an instant payment that looks like a scam: it combines the payee check and the risk score, asks targeted questions about the payment in plain language, explains the specific scam pattern, and holds, delays or escalates the payment to a human specialist when the risk stays high. Unlike fraud scoring, which stops payments the customer did not make, it protects customers from payments they are being manipulated into making.

Deployments
6 public, best grade B
Reported detection improvement
300%
Starling Bank, vendor claim
Cross industryBanking

AI agent for complaints recognition, investigation and response

An AI agent that recognizes when a customer interaction is a complaint, logs it against the regulatory definition, classifies its root cause and severity, gathers the evidence, drafts the acknowledgement and the response for a human handler to approve, and tracks every statutory deadline until the case is closed.

Deployments
2 public, best grade B
Reported time saved per task
about 5 minutes
Lloyds Banking Group, organization claim
Payments and cardsBanking

AI agent for payment initiation within a customer mandate

An AI agent that initiates and completes payments or purchases on a customer's behalf, within a mandate the customer set in advance (spending caps, allowed merchants or categories, a tokenized credential and rules for when to ask for confirmation), and then confirms and reconciles every transaction it made.

Deployments
7 public, best grade B
Autonomy
Supervised agent