ScalePoint Research Updated 14 October 2026
Get in touch

Updates by email

Get an email when new research is out

Sign up to hear when a new test, study or article is published.

Process map › C3 Disputes, claims and complaints

Merchant chargebacks

When a customer disputes a card payment, the merchant has a few weeks to decide whether to fight it and, if so, to prove the payment was valid. This profile describes how that work is done today and how well AI suits it, drawing on ScalePoint Research's own testing and a scan of 37 vendor tools. The economics behind it, from ROI to vendor pricing, are in the chargebacks report described at the bottom of the page.

Merchant chargebacks: on this page
  1. Our verdict
  2. What it is
  3. Scale and volumes
  4. The typical steps
  5. What makes it hard
  6. AI fit and our test results
  7. Vendor landscape
  8. Get the chargebacks report
Who does itMerchants, and platforms selling as merchant of record
Usual teamPayments operations, fraud or customer service
Starts withA dispute notice from the acquirer or PSP
Set byCard scheme rules and acquirer deadlines
1Our verdict · AI readiness

High: AI suits the work, and tools are on sale

Most of a chargeback case can be done by machine, and the biggest gain comes from running the whole case, not a single step. Tools that claim to do this are on sale now.

Based on ScalePoint Research's analysis of the process, a scan of 37 vendor tools and one model test of the fight-or-accept decision. The rating will be reviewed as new evidence comes in.

Fit of the workHigh

Code takes in and files cases, machine learning scores them, and language models choose the evidence and write the response. People set thresholds and handle conflicting records.

Agent fitHigh

A case runs through eight steps and four or more systems, so an agent working the whole case adds the most.

Tools availableHigh

37 tools on sale. 10 claim an agent that works a case end to end. None of these claims is tested yet.

How the ratings work

AI readiness is ScalePoint Research's view of how well AI suits a process today. It combines three parts, each rated High, Medium or Low. ScalePoint Research's own test results are reported in section 6 and will be used to check these ratings as more processes are tested.

PartHighMediumLow
Fit of the workA machine could do most steps, using rules, machine learning or language modelsA machine could do some stepsAlready run by code, or mostly a person's call
Agent fitA flow across several steps and systems that an agent could run end to endAn agent could prepare the work for a person to decideA single step or calculation
Tools availableSeveral tools on sale claim to do it with AI todayTools on sale, with little AI in themFew or no tools on sale

2What it is

  • A reversed payment. A cardholder asks their card issuer to reverse a payment, for example because they don't recognise it, the goods never arrived or a subscription wasn't cancelled.
  • Who pays. The issuer takes the money back from the acquirer, and the acquirer takes it from the merchant.
  • The merchant's choice. Accept the loss, or contest it by sending evidence that the payment was valid.
  • The cost. Each chargeback costs the sale, the goods and a fee. A high dispute rate can also bring scheme fines4 and, in the end, the loss of the ability to take card payments.

3Scale and volumes

261m → 324mchargebacks a year worldwide, 2025 to 2028, forecast by Datos Insights for Mastercard2,3
106mdisputes handled by Visa alone in 2025, up 35% since 2019. Visa's rules call a chargeback a dispute1
45%of merchant chargeback volume is fraud, first and third party combined. The other 55% are non-fraud disputes3
50%of merchants handle chargebacks fully in-house. 34% outsource chargebacks but keep fraud in-house, and 16% outsource both3

4The typical steps

  1. 1Receive the dispute
  2. 2Gather the facts
  3. 3Fight or accept
  4. 4Build the evidence
  5. 5Write the response
  6. 6File by the deadline
  7. 7Follow the outcome
  8. 8Report and fix causes
How each step is done in most firms now:CodeRulesMLLLMPersonEmerging

What AI could take on at each step is in section 6.

Before it becomes a chargeback. Many merchants now catch disputes earlier. Alert services from Visa (Verifi) and Mastercard (Ethoca) warn the merchant when a cardholder complains to the issuer, so the merchant can refund first. Order data can also be shared with the issuer at the enquiry stage, through Visa Order Insight or Ethoca Consumer Clarity7, so the cardholder recognises the purchase. Some merchants set rules to refund certain disputes automatically.

  1. Receive the dispute

    The acquirer or PSP sends the chargeback through its portal or an API, with the transaction, the amount, the scheme's reason code and the response deadline. A case is opened in the merchant's chargeback tool or ticket queue.

    Now:Code

  2. Match it to the order and gather the facts

    Find the original order and pull the records that matter: payment and authentication data, delivery tracking, customer emails and chats, refunds already given, account logins or usage. Integrations do the lookups where they exist. Elsewhere a person searches each system.

    Now:CodePersonLLM (emerging)

  3. Decide whether to fight or accept

    Check whether a refund was already made, whether the evidence answers the reason code, and whether the amount justifies the effort. Rules set value thresholds and auto-accept obvious cases. Some vendor tools score the chance of winning with machine learning. The rest is a person's call.

    Now:RulesMLPersonLLM (emerging)

  4. Build the evidence pack

    Select the documents the reason code calls for. For a "not received" claim that means proof of delivery. For a fraud claim it can mean the customer's earlier purchase history, under Visa's Compelling Evidence 3.0 rules.5 Reason-code checklists drive it, and vendor tools pick evidence from past win and loss data.

    Now:RulesMLPerson

  5. Write the response

    Write a short letter that explains, point by point against the reason code, why the payment was valid, and refers to each piece of evidence. Most teams fill templates per reason code. Some now use AI drafting tools.

    Now:RulesPersonLLM (emerging)

  6. File before the deadline

    Upload the response through the acquirer's portal or API. A late or incomplete response loses automatically.

    Now:CodePerson

  7. Follow the outcome

    The issuer reviews the response. If the cardholder keeps disputing, the case can go to pre-arbitration, where the merchant decides again whether to continue, and then to arbitration by the scheme, which carries a further fee for the losing side.

    Now:CodePerson

  8. Report and fix causes

    Track win rates, reasons and dispute ratios, and pass recurring causes back to the teams that can fix them, such as unclear billing descriptors, slow delivery or weak fraud checks.

    Now:CodePerson

Typical set-up

  • Large merchants and marketplaces: hybrid. In-house work connects order, delivery and customer data, which feeds a vendor chargeback platform or service. A few build the whole flow themselves on their PSPs' dispute APIs.
  • Mid-sized merchants: vendor stack. A chargeback tool or service connected to the PSP, often paid as a share of what it recovers.
  • Small merchants: the PSP's own tools. Responses are filed by hand in the PSP's dashboard, or through automation the PSP offers, such as Stripe Smart Disputes.
  • Where AI sits today. Machine learning is common inside vendor tools for scoring and evidence choice. Language models are the newest layer and are mostly used to draft responses.

5What makes it hard

  • Rules set the frame. Card scheme reason codes decide what evidence counts, and deadlines are tight: around 30 days for the acquirer at Visa and 45 at Mastercard6, less for the merchant.
  • True fraud or friendly fraud. Many chargebacks come from genuine customers who forgot a purchase or didn't recognise it. The right response differs.
  • Scattered evidence. Order, delivery, payment and customer contact data sit in different systems owned by different teams.
  • Volume spikes. A delivery failure can leave a team with many cases due on the same day.
  • It varies by business. Marketplaces may handle disputes for their sellers, subscriptions see more cancellation disputes, and travel firms face disputes long after the payment.
ScalePoint Research view

6AI fit and our test results

StepWhat could do itWhat stays with a person
Receive
Code
None. Cases arrive through acquirer and PSP feeds.
Gather the facts
CodeLLM
Code does the lookups. A model reads customer emails and chats. A person fills gaps where a system has no link.
Fight or accept
MLLLM
Sets the value thresholds and takes cases with conflicting records (study S3, below).
Build the evidence
RulesMLLLM
Checks the pack on high-value cases.
Write the response
LLM
Spot-checks letters.
File
Code
None. A portal upload or API call.
Follow the outcome
CodeLLM
Decides on pre-arbitration for larger amounts. A model can prepare the case.
Report and fix causes
CodeLLM
Acts on recurring causes with other teams. A model can group causes from case notes.
The whole case
Agent
An agent can run steps 2 to 6 as one case, calling code, models and people in turn. A person signs off where the value or the records call for it.

What our test found

  • The decision step works. In study S3, Claude Opus 5.5 made the right fight-or-accept call in 42 of 44 merchant cases at low effort and 43 at high.
  • The cheaper model trailed. Mistral Medium 3.5 got 37 with reasoning on and 33 with it off.
  • Case value matters as much as accuracy. The cheaper model still lost less money than Opus at low effort, because its misses fell on low-value disputes.
  • Messy records are the weak spot. Conflicting records were where even Opus slipped, so those cases should go to a person. Read study S3 →

What our tests add. ScalePoint Research tests how well general AI models do the work, and what each case costs. That sets a baseline: a tool that charges a share of what it recovers should do clearly better than a well-prompted general model running at a known cost per case.

Confidence: medium. Based on one ScalePoint test on synthetic cases, published market data and vendors' own claims. No independent before-and-after results were found.

7Vendor landscape

ScalePoint Research scanned 37 tools a merchant can buy or switch on for chargebacks, checked on 2 November 2026. They fall into five groups, with very different amounts of AI in them.

  • AI-first chargeback tools (6). Start-ups built around automating the dispute response, such as Chargeflow and Justt. Most claim to build and file responses with no manual work, and they usually charge 15 to 30% of what they recover.
  • Fraud platforms with a dispute module (10). Fraud prevention vendors such as Forter and Signifyd that have added dispute handling. AI use ranges from a full dispute agent to rules with a recommendation layer. Pricing is bundled with the fraud contract.
  • Tools built by payment providers (11). Most leave the response to the merchant. Stripe, Shopify and Worldpay are the exceptions, with automatic responses. Merchants pay a fee per dispute, and Stripe's AI tool also takes 30% of each dispute won.8
  • Card scheme tools (8). Visa and Mastercard alerts and order data that stop disputes before they become chargebacks. These are rules, not AI, and Visa's generative AI drafting tool for merchants is still a pilot.1
  • Chargeback services (2). Analysts plus software, sold as a managed service and quoted case by case. How much is automated is hard to tell from outside.
10 agent in production4 copilot9 unclear14 no AI

Who claims an agent in production

These 10 products are described by their vendors as working a case end to end, from evidence to filing, and have paying customers. These are the claims ScalePoint Research wants to put to the test.

  • AI-first tools: Chargeflow, Justt, Kloutit, ChargePay, Disputifier
  • Fraud platforms: Forter (Dispute Management), Signifyd (Chargeback Recovery)
  • Payment providers: Stripe Smart Disputes, Shopify Payments, Worldpay Disputes Defender

Based on each vendor's public material, not tested. Only Justt, Forter and Signifyd also describe making the fight-or-accept call. Worldpay's claim conflicts with older material that credits a partner for the product.

Vendor counts, AI ratings and prices come from the vendors' own published material, summarised. All vendors in this cluster →

8Get the chargebacks report

This profile gives the free view. The chargebacks report goes further: it sets out a reference set-up for handling chargebacks with AI today, then works through the business case and the implementation against it.

What's inside

  • A reference set-up. How a chargebacks flow would be built today, step by step: what code, machine learning, language models and an agent each do, the data each step needs, where people sign off, and for each step whether to build, buy or use what the payment provider offers.
  • Impact against the reference set-up. Money recovered, staff time saved and running cost per case for a reference merchant, and how that changes with volume and dispute value.
  • Implementation. The effort and time to put the set-up in place, and the controls a regulated firm needs.
  • Cost of mistakes. What a wrong call costs, and where people should stay in the loop.
  • Vendors mapped to the set-up. Which of the 37 tools cover which steps of the reference set-up, and how they charge.

Deciding which process to start with across your operations? That's what the AI business case service is for.

Related

Process mapC3 Disputes, claims and complaints VendorsWho sells software for this cluster Study S3Includes a test on this process Acquirer disputes and cardholder disputesThe same dispute seen from the acquirer and the issuer

Sources

  1. FStech: Visa rolls out suite of AI-powered dispute resolution tools, April 2026
  2. Mastercard: What's the true cost of a chargeback in 2025?
  3. Ethoca (Mastercard) and Datos Insights: 2025 State of Chargebacks report, pages 4, 12 and 21
  4. Visa Acquirer Monitoring Program fact sheet, 2025
  5. Checkout.com: Visa Compelling Evidence 3.0 explained
  6. Chargebacks911: Mastercard chargeback time limits
  7. ChargebackHelp: Order Insight and Consumer Clarity compared
  8. Stripe: Dispute management and Smart Disputes

Vendor counts and pricing come from a ScalePoint Research scan of vendor sites and app store listings, checked on 2 November 2026. A description of how the process typically runs, with ScalePoint Research's view on AI. Practice differs by firm, acquirer and scheme. Status as of November 2026.