US chargeback volume reached ~105M disputes representing ~$11B in 2024 — up from $7.2B in 2019 — and global volume is projected to grow from 238M to 337M transactions by 2026 (+41%). The all-in cost per dispute is $128 for merchants and the cost multiplier has climbed to $3.75–$4.61 per $1 of chargeback loss. A dispute automation program sized to last year's volume will be underbuilt within 18 months at current growth rates.

Reason-code distribution is the single most important input to model design because the decision logic, evidence requirements, and representment economics differ wildly across codes. Visa 10.4 disputes are eligible for Compelling Evidence 3.0 (CE3.0) defense using prior non-disputed transaction history — an AI-tractable retrieval problem. Visa 13.1 "not received" disputes turn on delivery confirmation and tracking artifacts. Treating all reason codes with a single policy means under-investing in the codes that dominate your portfolio and over-automating long-tail codes that do not justify the engineering.

The MRC 2024 Chargeback Field Report puts friendly fraud at 75–79% of total chargebacks, and 72% of merchants reported a year-over-year increase in 2024. Buyer's remorse alone drives 65.3% of friendly-fraud cases (Chargeflow 2024). Friendly fraud and true third-party fraud require opposite responses: third-party fraud should be conceded quickly with process improvement upstream, while friendly fraud is the exact scenario Visa CE3.0 and Mastercard First-Party Trust were designed to let merchants contest. Misdiagnosing the ratio means the automation either concedes winnable disputes or wastes evidence packaging on unwinnable ones.

Issuer-side automation is bounded by Regulation E's 10-business-day provisional credit and 45/90-day final resolution clocks, and by Reg Z's 60-day cardholder claim window — so the model must prioritize fast, compliant first-response over recovery optimization. Acquirer-side automation operates inside Visa VCR and Mastercard MCOM representment windows and is measured primarily on merchant win rate and net recovery. A single automation stack cannot optimize both without explicit role segmentation.

The Visa Acquirer Monitoring Program (VAMP) ratio effective June 2025 uses total disputes (TC40 fraud alerts + TC15 chargebacks) divided by total sales, with a 2.2% threshold that drops to 1.5% for North America, EU, and APAC from April 1, 2026. Merchants and acquirers above threshold face escalating fees and potential network termination. A rising chargeback ratio is both a regulatory risk and a signal that the automation program should prioritize intake triage and upstream fraud prevention, not just representment.

Adyen's July 2025 representment timeline tightening reduced the US/Canada window to 9 days and other regions to 18 days, with a new five-tier fee structure that penalizes slow responses. Manual teams typically need 3–7 days just to assemble evidence packages — which means any manual process now misses windows on a meaningful share of representable disputes. Same-day AI evidence packaging is no longer optional; it is a prerequisite for executing Visa CE3.0 defenses and Mastercard First-Party Trust challenges at portfolio scale.

A dispute pipeline spans three inference types — document intelligence, LLM rule interpretation, and scoring — and each has different latency, residency, and cost profiles. Splitting them across vendor APIs typically means every dispute artifact transits multiple cloud perimeters, which is both a PCI scope expansion and an audit-trail fragmentation problem. A single deployment fabric for all three inference types is the simplest architecture and usually the cheapest at steady state.