AI agents are already initiating financial transactions on behalf of people, and South African banks were not built for that. That was the central warning at Nedbank’s Innovation Day in Johannesburg last week, where banking and technology leaders said the industry’s biggest blind spot is not AI itself but who, or what, is on the other side of the transaction.
Every bank account, every credit product and every fraud-detection system in the country was designed for a human customer: a person with a face, a salary and a credit history. That assumption is now under pressure.
David Kerrigan, an author and technology analyst who lectures at Stanford’s Continuing Studies programme and consults for Mastercard and Enovation, said the problem runs deep. “Everything they’ve ever built was on the basic assumption that they were serving human customers,” he said.
Kerrigan told TechCentral that today, someone can open an app, instruct an AI assistant to book flights, compare insurance options and complete a payment, and never personally authorise each step. “The agent does it. The bank processes it. But the bank’s systems were not designed to know the difference.”
The first problem is identity. Banks verify humans using faces, fingerprints and ID numbers. An AI agent has none of those.
Chipo Mushwana, Nedbank’s executive for payments and technology, said a new approach is needed. “You need a different framework to assess the identity of an agent,” she said. “When was it created? Who is its owner? How long is it supposed to live for?”
Authentication for agents
Mastercard and Visa are already building authentication systems for agents, she said – work that matters because credit is the next complication. A bank granting a loan needs to assess the ability to repay, which has always meant looking at a person’s income. An entrepreneur running several revenue-generating AI agents does not fit that model.
“Maybe we need to create decisions that are a reflection of someone’s ability to pay at a point in time,” Mushwana said, “and not necessarily tied to traditional income cycles.”
Read: The AI model spooking the world’s biggest banks
Nedbank says it is already using agents internally to sharpen its own credit-scoring systems. The next step, Mushwana said, is building systems that can recognise when an agent is on the other side.
A bigger question is whether banks remain relevant at all as agents take over more of the transaction.
“Banks used to always worry about being top of wallet, about being the first card you use,” Kerrigan said. “But AI agents will check, for every purchase you make, what is the best payment method for this.”
Agents will have near-perfect price and product information and will route payments to whatever option saves the most. A retailer might offer a discount for an account-to-account payment over a card; a human would never notice, but the agent always will.
Mushwana said banks will not be bypassed entirely, because payments still need to be validated, traced and made reversible. “You need a function that validates the agents, validates the transaction, reports it as true,” she said. “Liability is held by someone.”
But the game has changed. The question is no longer which bank app is easiest to use, but which bank’s infrastructure agents will choose to route through.
South African banking is heavily regulated. The Financial Intelligence Centre Act – better known as Fica – requires identity verification; the National Credit Act governs lending. None of that legislation contemplated a non-human transacting party.
The gap is not unique to South Africa, Kerrigan said. “The legislation that they have was barely looking to catch up with AI, and it wasn’t ready then for agentic AI.”
‘Industry-wide evolution’
Ciko Thomas, Nedbank’s group managing executive for personal and private banking, acknowledged the issue. “As more autonomous capabilities emerge, the regulatory conversation will evolve, particularly around liability, consent and accountability,” he told TechCentral. “At this stage this is not a bank-specific issue, but an industry-wide evolution.”
On liability, Kerrigan said the law currently defaults to the human. “In most cases, the responsibility goes back to the human, not to the people who created the AI.” He pointed to work by Mastercard, a client of his, as an early attempt to close that gap.
Read: Investec’s contrarian AI bet: people over machines
The card networks have indeed made agentic payments a priority. Mastercard launched Agent Pay in 2025, issuing “agentic tokens” that bind a card credential to a specific AI agent, and in March added Verifiable Intent, an open, cryptographically signed record meant to tie a cardholder’s identity and instructions to whatever the agent ultimately does, so intent can be proven if a payment is disputed. Visa has set out a rival approach in its Trusted Agent Protocol, and both have since joined a wider industry effort to standardise how an agent proves who authorised it.
Thomas said Nedbank is generating 48 million personalised interactions using machine learning, has seen 70% year-on-year growth in AI-driven activity and has more than 2 000 AI agents in internal use. “Our ability to respond is not about catching up,” he said. “It is about accelerating from an established base.”
Mushwana was more direct about the real test. “All of us have access to the same capability. All of us have access to the same technology,” she said. “It’s just, is it valuable to our customers? Are they willing to pay for it?”
Read: Nedbank, Jumo bet on AI lending for the underbanked
The banks that answer that question first will not just win customers. They will decide how AI agents spend everyone else’s money. – © 2026 NewsCentral Media
