Visa processed 106 million charge disputes globally in 2025 — a 35% increase since 2019 — and has now launched AI tools to automate responses, triage, and back-office handling across merchants, issuers and acquirers. On the surface, this is obvious progress. Dispute resolution is a textbook automation candidate: high volume, repetitive, rule-bound. For a mega-merchant or large acquirer, Visa's tools will reduce headcount and cycle time. But here is what matters to UK regulated firms: Visa's tools optimise for throughput. They do not optimise for the regulatory burden that sits behind every dispute decision. Under FCA Consumer Duty PS22/9, every firm handling customer disputes must demonstrate fair treatment, not just faster treatment. The Payment Systems Regulator and the FCA both expect firms to be able to explain and justify individual decisions. An AI tool that processes disputes at scale without audit-ready decision logs, confidence scores, or human-review triggers is not compliant by default. It is compliant only if the firm using it builds that infrastructure around it.

This story is part of a wider pattern: large payments and fintech infrastructure providers are deploying generative AI and machine learning as if compliance is a downstream concern. Harvey and Luminance have made significant noise in legal discovery and document review, but both have encountered the same friction point when deployed in regulated environments: their speed advantage evaporates the moment you add the governance, explainability, and audit trail that the Solicitors Regulation Authority, FCA, or Financial Reporting Council require. Generative AI is excellent at guessing. It is poor at proving its work. And in financial services, proving your work is not optional. What we are seeing with Visa's announcement is the same mistake playing out at the infrastructure layer: the assumption that because an AI system works at scale, it works for regulated firms. It does not.

Trovix's view is that AI in regulated financial services must be built backwards from audit and compliance, not forwards from capability. A dispute resolution system should not be optimised for fastest response time per transaction; it should be optimised for zero ambiguity in decision-making and instant reconstruction of the reasoning behind any decision, two years later, for a regulator or customer complaint. This is why we built Trovix Sift to extract and classify document data with explicit confidence thresholds and human-review queues — not to replace human review, but to make it systematic and defensible. And why Trovix Watch monitors regulatory change in real time, because FCA guidance on dispute handling will shift, and AI systems that were compliant last quarter may not be compliant next quarter. Visa's tools will work well for Visa. They will cause problems for mid-market acquirers and issuers in the UK that bolt them on without the governance layer underneath.

If you are a mid-market payment processor, bank, or fintech operator in the UK, do not assume Visa's tools are turnkey. Before you deploy any third-party AI system into dispute handling, map your current decision-making process against FCA Consumer Duty, Payment Systems Regulator expectations, and your own complaints procedure. Build a decision log that captures why each disputed transaction was ruled in favour of merchant or customer. Test edge cases. Run a manual sample review across AI-resolved disputes every month. If you cannot do this today because your current process is already too fast and opaque, then your process is already non-compliant — and AI is going to make that visible to a regulator sooner or later.

Source: CNBC