For two decades, banks have treated the mobile app as the centre of the customer relationship. The client logs in, sees the bank's view of their financial life, chooses from the bank's products and completes the transaction inside the bank's interface.
A personal AI agent changes the starting point. A client may ask one assistant to understand spending across several banks, compare savings rates, move surplus cash, find credit or choose the best way to pay. If that assistant can reach authorised services, compare structured offers and prepare transactions for the client to approve, the bank no longer controls every interaction that leads to a transaction.
The bank still holds the regulated capability. It holds deposits, assesses credit, settles payments, detects fraud, resolves disputes and protects the account. What becomes contestable is the interface, the recommendation and the routing decision.
Andrew Baker, Group CIO at Capitec, writing on his personal blog, calls this the great agentic unbundling. In "The Great Agentic Unbundling: When the Agent Becomes the Interface, Banks Will Have to Compete for Every Transaction" (11 October 2026), he argues that agents may increasingly plan, agree and pay without anyone opening a banking app, that banks lose their default position as a result, and that they will have to compete for each transaction. He is careful about it. He notes that the protocols involved are at different levels of maturity, that they do not yet form a universal consumer agent network, and that agent-to-bank negotiation is a forward-looking scenario rather than a present market.
It is a provocation worth taking seriously. This eight-part series tests it against architecture, security, liability, regulation and the practical realities of South African banking. Our central conclusion is a little more cautious than the headline:
Banking apps will not disappear suddenly. Their position as the compulsory front door can weaken as authorised agents become another channel through which customers discover, compare and use financial services.
Kestrel Bank and its customers are fictional, and their products, figures and systems are illustrative. Kestrel is a mid-sized South African retail bank, headquartered in Johannesburg, with about 1.2 million retail clients. It offers a transactional account, a 32-day notice deposit, a fixed deposit, personal loans and a credit card, through a well-used app and modern APIs that sit in front of an older core banking platform. Naledi, a physiotherapist in Gqeberha, is one of its clients. The series follows Kestrel's leadership and architects as they work out what an agent-mediated future would mean for them, and Naledi as the client it would happen to.
This series is strategic and architectural analysis, not legal, regulatory, investment or financial advice.
The relationship can survive while the interface moves
Naledi's salary is paid into her Kestrel account. She keeps her savings at another bank, which we will call Bank B, and her credit card at a third, Bank C. Today that fragmentation costs her effort. She has three apps, three sets of credentials, three fee structures and no single view of her money. Moving cash between them is a chore, so she mostly doesn't, and that friction quietly favours Kestrel as her primary bank.
Now imagine, as a scenario, that Naledi uses an independent personal finance assistant. With her permission, it could:
- aggregate her authorised balances and transactions across all three banks;
- explain the total fees she pays across providers;
- monitor savings rates and their restrictions;
- recommend which account should fund a particular payment;
- prepare a transfer or a product application;
- ask her to approve the exact action;
- track whether it settled correctly.
Naledi still owns her accounts. Each bank still authorises access and enforces its own transaction rules. What changes is where she expresses intent: in a conversation with her assistant, rather than inside any one bank's app.
In this scenario, the assistant notices that Bank B pays more on notice savings than Kestrel, and proposes moving surplus each month after her salary lands. She agrees. Nothing dramatic happens at Kestrel. Naledi is still a client, her salary still arrives, and her debit orders still run. But the balance that used to sit in her Kestrel account for most of the month now leaves within days.
Three layers of a banking relationship
It helps to separate the relationship into three layers, because they do not move together.
| Layer | What the client experiences | What can change |
|---|---|---|
| Interface | Conversation, app, website, branch or call centre | Can move to an independent assistant |
| Decision and orchestration | Comparison, advice, routing and workflow | Can be shared with, or led by, an agent |
| Regulated capability | Deposits, credit, payments, identity, fraud and settlement | Remains with authorised institutions and rails |
The interface becoming optional does not make the bank optional. It changes how the bank's capabilities are discovered and selected. A bank that is excellent at the bottom layer but has relied on the top layer to keep clients' activity in-house will feel the change first.
What happens to "customer ownership"?
Banks do not own customers. They earn relationships through trust, service, value and convenience. Much of what looks like loyalty is also friction: the effort of comparing, applying and moving money. An agent that lowers that effort makes it easier for a client to use several providers at once, and so makes the difference between loyalty and inertia visible.
That is why the measure that matters can shift from registered customers to:
- share of payments routed through the bank;
- share of deposits and stable funding;
- share of credit use;
- preference at the level of individual products;
- service reliability as an agent sees it;
- retention after comparison;
- the cost to win and serve each transaction.
At Kestrel, the Head of Digital Channels modelled Naledi's pattern across a segment of salaried clients. In the model, the number of registered clients barely moves. Balances held between salary days thin noticeably. It was presented as a scenario to plan for, not a crisis, but the lesson was uncomfortable: a bank can keep the account and lose the profitable activity. The relationship stays legally open while becoming economically thinner.
The agent is not necessarily neutral
Moving the interface away from the bank does not remove gatekeeping. It can move it somewhere else. An assistant platform may influence:
- which institutions can integrate at all;
- which products appear in a comparison;
- how "best" is calculated;
- whether commercial partnerships affect ranking;
- which data stays portable if the client changes assistant;
- what fee or revenue share is charged to providers;
- which actions require the platform's own identity or wallet.
Baker makes the same point: platforms that control discovery could become the new gatekeepers, which is one reason he argues banks should back open protocols. The risk for a bank is swapping dependence on its own app for dependence on one external assistant whose ranking it cannot see and whose terms it cannot negotiate. The risk for a client is a comparison that looks independent but isn't.
Neither risk is a reason to avoid the channel. Both are reasons to insist on open, governed access, transparent ranking criteria and data portability, and to avoid exclusive arrangements with any single platform.
The app still matters
None of this makes the bank's own channels redundant. Kestrel's app remains the right place for:
- identity enrolment and device binding;
- high-risk approvals, such as a new beneficiary;
- disputes, complaints and servicing;
- rich financial planning and product education;
- regulatory notices and account management;
- recovery when an external assistant or platform fails;
- clients who simply prefer direct control.
In the scenario above, when Naledi's assistant prepares something unusual, Kestrel asks her to confirm it in the Kestrel app, on the device the bank has already bound to her identity. Part 3 explains why that step matters.
The strategic mistake is treating the app as the only legitimate channel. A better model treats the bank's capabilities as secure services, available through both bank-owned interfaces and the interfaces customers choose for themselves. Part 2 looks at the protocols that might carry those services, and at what each of them does and does not guarantee.
Where South Africa stands
South Africa has real foundations for this: mature banks, widely used digital channels, real-time payment infrastructure such as PayShap, and active work by the South African Reserve Bank on modernising the national payment system, including a consultation in 2026 on its NPS 2030+ vision. What it does not have is a binding open banking or open finance framework. Policy on that has been consulted on several times, but nothing equivalent to a mandated data-sharing regime is in force.
So the scenario in this part is exactly that: a scenario. No South African client can today point an independent assistant at every bank and have it compare and transact across all of them. Baker himself expects higher-income clients to adopt first, and suggests that for many lower-income clients the first AI channel may be an assistant inside a messaging app they already use. Both points matter for anyone planning in rand rather than in headlines. Part 8 sets out a staged roadmap that starts with what is possible now.
Executive questions
Kestrel's executive committee used six questions to test its exposure. They work for any bank:
- Which of our profitable client activities rely mainly on switching friction?
- Could an external agent understand the full cost and conditions of our products?
- Would our strongest products win if they were evaluated independently every month?
- Which of our services could be safely exposed as read, recommend, prepare and execute capabilities?
- Are we prepared to serve a client whose primary interface belongs to someone else?
- How would we avoid dependence on a single assistant platform?
The third question produced the longest discussion. Kestrel's 32-day notice deposit pays a competitive rate, but its conditions sit in a PDF. Part 5 picks up what happened when an agent tried to compare it.
The short version
The interface can move without the relationship ending. An assistant can become where the client expresses intent while each bank keeps the regulated capability.
Separate the three layers. Interface, decision and regulated capability move at different speeds; the interface moves first.
Measure activity, not accounts. A bank can keep the client and lose the profitable flows.
The agent is not automatically neutral. Gatekeeping can move to the assistant platform; back open, governed access.
The app keeps its hard jobs. Identity, high-risk approval, disputes and recovery stay with the bank's own channels.
In South Africa this is a scenario, not a market. The foundations exist; the rules, standards and liability arrangements are still forming.
Next: Part 2, MCP, A2A and AP2 are not the same thing. All parts are on the series page.
How CloudNala can help
We help banks and financial-services firms work out what agent-mediated channels mean for them before the market decides it for them: modelling which client activities depend on interface control, testing whether products would survive independent comparison, mapping which capabilities could be exposed safely, and shaping a channel strategy that avoids dependence on any single assistant platform.
Work with CloudNala
CloudNala helps organisations move from technology ambition to practical execution across cloud, AI, data, platform engineering and digital services.
Whether you are exploring AI, modernising your cloud environment, building a public-sector digital service, or turning an idea into a working MVP, we can help you shape the roadmap and deliver the next step.
Book an Agent-Ready Banking Review or write to us at consult@cloudnala.co.za