GenAI Dispute Authoring
at Visa Scale

Overview

Visa Recover (formerly Verifi) is Visa's post-purchase dispute management platform — the system that helps merchants defend against chargebacks, protect revenue, and resolve disputes with cardholders before they escalate to costly arbitration. When I joined the Post-Purchase product team, dispute authoring was a deeply manual, time-intensive process that required merchants and their representatives to compose detailed representment cases from scratch for each dispute.

I saw an opportunity to apply Generative AI not as a buzzword project, but as a genuine workflow improvement: reduce cognitive load, eliminate repetitive writing, and help merchants submit higher-quality cases faster — directly translating to better win rates and more revenue retained.

The Challenge

Dispute representment is a regulated, high-stakes domain. A merchant disputing a chargeback must submit a structured, evidence-based rebuttal letter that satisfies card network rules, addresses the specific dispute reason code, and tells a clear, compelling story. For a merchant processing thousands of transactions per month, this meant a significant authoring burden — often outsourced to expensive chargeback management vendors.

The key constraints were significant:

• PII sensitivity: Dispute data contains cardholder names, transaction details, and sensitive financial information — it could not be sent to third-party AI APIs without careful architecture.
• Regulatory compliance: Output had to satisfy card network rules and reason code requirements precisely.
• Trust and accountability: Merchants needed to review and own their submissions — fully automated output without human review wasn't acceptable.
• Enterprise reliability: This was a revenue-critical workflow. Latency and failure modes had to be handled with extreme care.

My Approach

I led the conception, design, and product definition of the GenAI authoring workflow from the ground up. The core design principles were:

1. PII-safe architecture first. I designed a data sanitization layer that stripped and tokenized sensitive identifiers before any content reached the AI layer, then re-injected them in the final output. This allowed us to use OpenAI's API without exposing cardholder data — a non-negotiable requirement.

2. Human-in-the-loop by design. The AI generated a first draft of the dispute rebuttal letter, but the merchant always reviewed, edited, and explicitly approved the final submission. The UX was designed to make the human feel in control, not supervised by a machine. This also created a natural feedback loop for improving prompt quality over time.

3. Reason-code awareness. The prompt engineering incorporated the specific dispute reason code, the merchant's transaction evidence, and card network guidance — ensuring generated content was relevant and structured correctly rather than generic.

4. Measurable outcomes from day one. I defined clear success metrics before launch: authoring time per case, case quality score (internal rubric), submission completion rates, and merchant NPS for the feature. This allowed us to iterate with data.

Results

The GenAI authoring workflow launched in production and delivered measurable outcomes:

Strategic Roadmap

Alongside the GenAI authoring launch, I built Visa Recover's 5-year strategic roadmap sequencing a broader set of AI-powered capabilities:

• Self-service onboarding agents: AI-guided merchant onboarding to reduce support burden and time-to-active.
• AI-powered support: Intelligent case routing and resolution suggestions for the merchant support team.
• ML case outcome prediction: Scored dispute cases by predicted win probability to help merchants prioritize representment effort.

This roadmap was the basis for a business case to expand Total Addressable Market by 95% and increase profitability by 150% over five years.

Platform Migration

Concurrent with the AI work, I also led a multi-year enterprise platform migration — coordinating 200+ dependency mappings across core features and complex payments ecosystem integrations, with a 50+ person cross-functional team spanning the US, Poland, and India. The migration was executed via phased rollout to ensure zero disruption to live dispute processing.

Key Takeaways

This project reinforced several principles I now apply across all AI product work:

• AI products succeed when they are designed around the human workflow, not bolted onto it.
• PII-safe architecture is not optional — it must be the starting point, not an afterthought.
• Human-in-the-loop UX builds trust, reduces liability, and creates improvement feedback loops.
• Measurable business outcomes must be defined before launch — not reverse-engineered after.