Group CEO Adrian Gore shared the figure at Discovery's interim results presentation, framing the bank's AI and machine learning investments as technology that is "not flashy" but delivers measurable, real-world results. Discovery has made these investments significant enough to include AI as a separate business unit in its reporting.

The 80% reduction is the headline, but the mechanism underneath it is what makes this worth unpacking, especially coming a week after we covered Capitec's Pulse AI delivering an 18% drop in call handling times. Two SA banks, two different AI applications, both producing auditable results.

Most fraud detection focuses on the sender. Does this transaction match your spending patterns? Is the amount unusual? Is your device behaving differently?

Discovery Bank's system adds a second layer by scrutinising the beneficiary across its entire ecosystem. When you initiate a payment, the AI checks whether other clients with your profile typically pay that beneficiary and whether that specific beneficiary has shown suspicious activity or been flagged elsewhere in Discovery's network. If people like you don't interact with that recipient, or if that recipient has accumulated risk signals across the bank's client base, your transaction's risk score goes up.

If the score crosses a medium threshold, you receive a Trust Alert explaining why the payment was flagged. You can review and decide whether to proceed. If the score hits a higher threshold, the transaction is blocked outright.

Bank CEO Hylton Kallner explained that this happens on every single transaction, individually and in real time, across the 9.5 million data events Discovery Bank processes daily. The system covers EFTs, real-time clearances, and PayShap payments.

Kallner identified false positives as the core design challenge. Blocking a legitimate school fee payment or a mortgage transfer is often more damaging to the customer relationship than a missed fraud attempt. The entire system is calibrated to minimise false positives without sacrificing protection.

This is a design philosophy, not just a technical detail. Discovery chose to build a tiered system (warn, then block) rather than a binary one (block everything suspicious). The Trust Alert creates a deliberate pause, forcing the customer to reconsider without removing their agency. For social engineering scams, where victims believe they are making legitimate payments, that pause may be the difference between losing money and catching the deception in time.

The 80% figure was presented at an earnings call, where the incentive is to highlight wins. A few things are worth noting.

Discovery hasn't disclosed the absolute EFT fraud volume before October 2025, so the baseline is unknown. An 80% reduction from a small number is different from an 80% reduction from a large one.

Discovery also serves a high-LSM, low-risk client base of 1.1 million clients with R21 billion in deposits. The behavioural data per client is rich, which makes the AI's job easier. Whether this approach would produce the same results at a mass-market bank processing higher volumes of lower-value transactions with less data per customer is an open question.

The beneficiary analysis is also a network effect play. The more clients Discovery has, the more data it has on each beneficiary, and the smarter the risk model becomes. At 1.1 million clients, the data density is solid for a focused demographic. But catching sophisticated cross-bank fraud syndicates that operate across multiple institutions requires a different kind of intelligence, the kind that Orca Fraud is building by embedding fraud detection across banks, telcos, and payment providers simultaneously.

Discovery's approach sits alongside Capitec's Pulse as evidence that SA banks are moving past "AI strategy" announcements into deployed systems with measurable outcomes. The common thread: neither system is a chatbot or a generative AI experiment. Both are purpose-built tools solving specific operational problems, call centre efficiency in Capitec's case and payment fraud in Discovery's.

For the customers who kept their money this year because a Trust Alert made them pause, it's the most valuable feature the bank has shipped. Gore's framing was deliberate: in a world obsessed with flashy AI, Discovery built something that quietly works.

This news first appeared in our 17 March ‘26 newsletter on CrabaRide long-distance ride-share trips.

See how Capitec is using AI to predict why you're calling in Capitec Pulse AI. Read how Orca Fraud is building cross-platform fraud intelligence for 70+ countries. And explore how AI is reshaping the continent in Africa's AI adoption.