AI agents are already completing payments, booking flights and even comparing insurance policies on behalf of South Africans, and the country’s banking infrastructure is suddenly facing a dilemma it was never built to solve. At Nedbank’s Innovation Day in Johannesburg, senior technologists warned that the real blind spot isn’t the technology itself but the fact that traditional banking systems still assume a human sits behind every transaction.
Every credit product, fraud‑detection rule and KYC protocol in the nation was designed for a person with a face, a salary slip and a South African ID number. That human‑centric model is now being challenged by software that can act, decide and pay without ever showing a fingerprint.
David Kerrigan, a technology analyst who lectures at Stanford and consults for Mastercard, explained that banks are “processing transactions they can’t identify”. An app user can trigger an AI assistant to book a holiday, select the cheapest insurance and authorise the payment – all without ever confirming the human behind the request. The bank’s back‑end sees a legitimate payment, but it has no way to verify whether the initiator is a person or a piece of code.
Identity verification is the first crack in the dam. Traditional banks rely on biometric data – face scans, fingerprints and ID numbers – to confirm a customer’s identity. An AI agent, by definition, possesses none of these markers. Chipo Mushwana, Nedbank’s executive for payments and technology, argued that a new framework is required: “When was the agent created? Who owns it? How long is it meant to operate?”
Authentication for AI agents in South African banking
The pressure to adapt is already prompting global card networks to develop agent‑specific authentication. Mastercard’s 2025 launch of Agent Pay introduced “agentic tokens” that bind a card credential to a particular AI, while its Verifiable Intent protocol creates a cryptographic record linking the cardholder’s consent to the agent’s actions. Visa’s competing Trusted Agent Protocol follows a similar logic, and both schemes have joined an industry consortium aimed at standardising how an AI proves it has been authorised.
Below is a snapshot of the emerging token‑based solutions and their key attributes:
| Feature | Mastercard – Agent Pay | Visa – Trusted Agent Protocol |
|---|---|---|
| Token type | Agentic token (AI‑specific) | Agent token (cryptographically signed) |
| Link to cardholder | Verifiable Intent record (signed consent) | Trusted Authenticator linking AI to cardholder |
| Dispute resolution | Cryptographic proof of intent for charge‑backs | Auditable trail for liability attribution |
| Adoption timeline (SA) | Pilot phase 2025, rollout 2026 | Pilot phase 2025, rollout 2026 |
The table highlights how both networks are converging on a model that couples a human’s identity with the AI that acts on their behalf, offering banks a clearer path to compliance.
Adapting these standards will be crucial for South African lenders, especially as credit products become entangled with autonomous agents. A bank that grants a loan today evaluates income, employment history and personal risk. If an entrepreneur runs multiple revenue‑generating AI agents, the traditional income‑based model no longer fits. Mushwana suggested a shift toward “decisions that reflect the ability to pay at a point in time”, decoupling credit assessment from conventional salary cycles.
Banks also have to ask whether they will remain relevant as agents increasingly dictate the most cost‑effective payment route. Kerrigan pointed out that AI agents will always choose the cheapest method, whether that’s a direct account‑to‑account transfer, a tokenised card payment or a new fintech offering. For retailers, this means discounts for specific payment channels could become invisible to human shoppers but obvious to their bots.
South Africa’s regulatory landscape adds another layer of complexity. The Financial Intelligence Centre Act (FICA) mandates identity verification, while the National Credit Act governs lending practices. Neither piece of legislation contemplated a non‑human transacting party, leaving a regulatory vacuum that could expose banks to unforeseen liability.
Industry‑wide evolution drives new standards
Ciko Thomas, Nedbank’s group managing executive for personal and private banking, stressed that the challenge is not isolated to one institution. “As more autonomous capabilities emerge, the regulatory conversation will evolve, particularly around liability, consent and accountability,” he told TechCentral. The consensus among industry leaders is that the shift is inevitable and that the focus should move from catching up to accelerating from an existing AI‑enabled base.
Nedbank’s internal data underscores the pace of change: the bank now generates 48 million personalised interactions via machine learning, records a 70 % year‑on‑year growth in AI‑driven activity, and operates over 2 000 AI agents across its services. These figures illustrate that AI is no longer an experimental add‑on but a core component of daily banking operations.
| Metric | Nedbank (2025‑2026) |
|---|---|
| Personalised AI interactions | 48 million |
| YoY growth in AI activity | 70 % |
| Internal AI agents in use | 2 000+ |
| Projected AI‑enabled transactions 2026 | 12 million* |
*Estimate based on current growth trends.
The data confirms that banks are already deeply embedded with AI, making the transition to agent‑aware systems a logical next step rather than a disruptive overhaul.
Mushwana summed up the commercial reality: “All of us have access to the same capability and technology. The real question is whether it delivers value that customers are willing to pay for.” In a market where fintechs and traditional banks vie for the same digitally savvy populace, the ability to securely and transparently handle AI‑initiated transactions could become a decisive competitive edge.
South African consumers are likely to feel the impact soon. As AI agents negotiate the best price, select the cheapest payment method and execute the transaction in milliseconds, the banks that can verify, trace and reverse these payments will earn trust. Conversely, institutions that cling to human‑only verification risk being bypassed by smarter bots that can slip through the cracks undetected.
The transformation is already underway; the next few years will determine whether South Africa’s banks can rewrite the rulebook fast enough to stay in the driver’s seat of a rapidly automating financial ecosystem.