A 96% approval rate hid foreign-issued cards failing 40% of the time

An approval rate of 96% leaves a 4% decline rate across the portfolio ^[15]. If foreign-issued cards were failing about 40% of the time ^[9], that segment could have been no more than a tenth of transaction volume, because 40% of 10% already accounts for every decline on the book ^[16]. The true share was smaller still, since domestic transactions get declined too. According to Adarsh Naidu, who worked on the portfolio as an architect, a customer complaint exposed the problem and the aggregate number looked fine because the failure sat in one segment ^[11].

The feedback is one-sided by construction. An approved transaction eventually acquires an outcome: a fraud claim may arrive weeks later, or never, and either way an answer attaches to it ^[4]. A declined transaction usually produces no correction at all ^[4]. Naidu wrote that repeated retraining can make a system genuinely better at catching fraud while it stays almost completely blind on wrongful rejections, and that part of the apparent improvement can come from the model becoming more conservative in areas where nothing tells it it has gone too far ^[5].

A wrongly rejected customer never appears in the fraud-loss report. The cost shows up instead as lost revenue, lower repeat purchases, higher churn, or marketing spend to replace someone the business already had ^[6]. Those numbers sit on other dashboards and belong to other teams ^[6]. "The fraud team gets credit for improving its metric. The business absorbs the cost of the metric nobody was measuring," Naidu wrote ^[7].

In April 2026, Visa tightened the threshold in its Acquirer Monitoring Program from 2.2% to 1.5% of transactions across the United States, Canada, the EU and Asia-Pacific ^[12]. That is a cut of 0.7 of a percentage point, about 32% off the ratio a merchant is allowed ^[18]. Crossing it can bring per-transaction penalties and mandatory remediation, with an invoice and an owner attached ^[13].

The 40% figure required no new label, because it came from approval data the business already had, cut by issuer geography ^[11]. Naidu also recommends running a second fraud model in shadow mode alongside the live system, letting it score the same transactions and recording its decisions outside the payment flow ^[14]. His article does not say what false declines cost beyond that one portfolio; the quantified numbers in it are the 96% approval rate, the roughly 40% foreign-card decline rate, and the Visa thresholds ^[19].

So the choice in front of a payments owner this quarter is one of ownership. Either approval rate by segment is reported in the same review as fraud loss, with a name against it, or the tightening that the Visa threshold rewards runs unchallenged until a customer calls. On Naidu's portfolio, that took more than a year: "Our conservative fraud settings had been quietly pushing away international customers for more than a year," he wrote ^[10].