The previous seven parts described a possible future: agents as a channel that customers choose, protocols that carry their requests, delegated authority, mandates, competition product by product, machine-readable banks and a wider fraud surface. This last part asks the practical question. What should a South African bank build now, test next and deliberately postpone?
The honest starting point is that South Africa has foundations, not a finished agentic banking market.
Kestrel Bank and its customers are fictional, and their products, figures and systems are illustrative. This series is strategic and architectural analysis, not legal, regulatory, investment or financial advice.
Where South Africa actually stands
There is a lot to build on. South Africa has mature banks, heavily used digital channels and established payment infrastructure. It also has PayShap, the real-time payment service operated by PayInc (formerly BankservAfrica) since March 2023, which added ShapID proxy payments and PayShap Request along the way.
Payments policy is also moving. In February 2026 the South African Reserve Bank published its NPS 2030+ consultation paper on the future of the national payment system, alongside its Payments Ecosystem Modernisation programme. It is a consultation, not final policy. In his article that opened this series, Andrew Baker points to SARB's consultation on the unbundling of payment services as part of the backdrop.
Open banking and open finance are earlier still. There is no binding South African open-banking or open-finance framework. The outputs so far are consultative: a SARB consultation paper in 2020, a position paper from the Intergovernmental Fintech Working Group the same year, and a draft open-finance position paper from the FSCA in 2023. Anyone planning agent-mediated access across banks should treat the regulatory model as still being decided.
POPIA applies throughout. Banks are responsible parties for the client data they hold, and an agent operator processing that data on a client's authority takes on its own obligations. The Information Regulator is the authority to watch on that side.
None of this means universal banking APIs or consumer agents can compare products and transact across every institution today. Protocol maturity, identity, liability, consumer protection, product semantics and commercial rules all remain open. Adoption won't be even either. Baker expects higher-income segments to adopt first, and suggests that for many lower-income clients the first AI channel is likely to be Meta AI on WhatsApp rather than a dedicated finance assistant. That is a point worth planning around. The agent channel may arrive inside a messaging app the bank doesn't control, for customers who never installed a personal finance tool.
The likely sequence
Stage 1: agent-readable information. Product terms and pricing, branch, ATM and support information, educational content, and status and service information. Risk is low because no customer account is accessed and nothing is transacted.
Stage 2: customer-authorised account insight. Balances and transactions, spending and fee analysis, and cash-flow and debit-order forecasts. Consent, privacy, data quality and revocation become central.
Stage 3: recommendation and preparation. Comparing products, generating personalised offers, preparing a payment or account action, and requesting the customer's exact approval. Suitability, ranking, disclosure and liability become material.
Stage 4: bounded execution. Payments to known beneficiaries, transfers between a customer's own accounts, capped recurring instructions and low-risk service actions. Strong authentication, fraud controls, mandate evidence and dispute handling are all required.
Stage 5: agent-mediated competition. Request-for-offer markets, routing across banks, negotiated product terms and delegated actions within explicit policy. This needs ecosystem standards and regulatory confidence beyond what any one bank's architecture can provide.
Over this series, Kestrel worked across the first four stages in a controlled pilot. Part 6 covered the product passport (Stage 1), and Part 3 the read-only access and preparation (Stages 2 and 3). The capped own-account transfer from Part 4 is the one small piece of Stage 4. Stage 5 is not something a single bank can do alone, and Kestrel hasn't tried.
What banks should build now
These eight capabilities are useful whatever happens with protocols and regulation, because each improves the bank's existing channels as well:
- A governed capability inventory. Identify read, analyse, prepare and execute APIs, with owners and risk classes.
- Machine-readable product data. Establish authoritative semantics, conditions, timestamps and versioning.
- Modern delegated identity. Separate customer, agent, operator and workload identity, with scoped, short-lived authority.
- An agent policy gateway. Enforce consent, operation, account, amount, beneficiary and timing restrictions outside the model.
- An evidence layer. Correlate intent, recommendation, mandate, authorisation, execution and outcome.
- Agent-channel fraud controls. Add signals, limits, step-up authentication and incident response.
- A multi-assistant strategy. Avoid depending exclusively on any one AI platform.
- A controlled test environment. Give approved teams and partners realistic schemas and synthetic data.
Kestrel's experience suggests an order. The capability inventory and product data came first, because they were cheap and exposed problems quickly; the passport work in Part 6 surfaced disagreements inside the bank about its own products. The policy gateway and the evidence layer came next, because nothing beyond read-only access was safe without them. The evidence layer here is the one described in The Agent Evidence Layer series, applied to banking.
What to pilot next
Good early pilots have clear customer value, bounded authority and outcomes you can observe.
| Pilot | Why it is useful | Authority limit |
|---|---|---|
| Fee explainer | Tests product semantics and retrieval | Read-only |
| Multi-account cash-flow view | Tests consent and data integration | Read-only |
| Savings comparison | Tests machine-readable product conditions | Recommendation only |
| Payment preparation | Tests payee, amount and evidence flow | The customer executes after exact approval |
| Own-account sweep | Tests bounded delegated action | Cap, reserve threshold and revocation |
| Dispute assistant | Tests evidence packaging and servicing | No automatic liability decision |
The own-account sweep is where Kestrel learned the most. It was the pilot whose test run went wrong in Part 4, when the assistant counted money as spare without allowing for a debit order. Kestrel now pairs every sweep with a reserve threshold that the customer sets and the bank enforces, so the agent's view of "spare" can never take the account below it.
What to postpone
- Broad autonomous movement of money.
- Creation of new beneficiaries without strong confirmation in a bank-owned channel.
- Binding credit or investment decisions led only by a general-purpose agent.
- Agent marketplaces whose ranking incentives aren't disclosed.
- One consent granting many unrelated banking powers.
- Relying on browser automation for material transactions where governed APIs are feasible.
- Production use before responsibilities for disputes and recovery are agreed.
The second item earned its place in Part 7, where the beneficiary rule and in-app confirmation stopped an injected recommendation. Postponing isn't the same as ignoring. Each of these deserves design work and regulatory conversation now, so that the bank is ready when the evidence and the rules support it.
An illustrative 24-month roadmap
The timing is illustrative. Regulatory engagement and each bank's risk appetite should decide the real dates. The shape matters more than the months. Foundations start first, pilots overlap with them rather than waiting for them to finish, and delegated execution comes last, after interoperability with partners has been proven.
The architecture views to keep
The views in architecture layers and purposeful views apply here, trimmed to the questions agentic banking raises:
| View | Question answered | Audience |
|---|---|---|
| Market and ecosystem context | Who owns the interface, the service and the rails? | Board and executives |
| Customer intent journey | What does the customer ask, see and approve? | Product, conduct and design |
| Capability responsibility | What do the agent, the bank and the human each do? | Business, risk and architecture |
| Identity and mandate flow | Who authorises whom to do what? | IAM, payments, risk and legal |
| Fraud and threat view | How can judgement or execution be manipulated? | Fraud and security |
| Evidence and dispute view | What proves the path from intent to outcome? | Operations, legal and audit |
| Transition roadmap | How does the bank move from read access to bounded action? | Programme and architecture |
Kestrel's pilot used four of them: the customer intent journey, capability responsibility, identity and mandate flow, and the evidence and dispute view. The others are added when there is a decision that needs them.
Board-level questions
- Which part of our customer economics depends on controlling the interface, or on inertia?
- Which of our products would win a transparent machine comparison?
- Can we expose capabilities without exposing uncontrolled authority?
- Who owns agent-channel strategy across digital, payments, data, fraud and product?
- How will customers understand and revoke delegated access?
- What evidence would we need to resolve a disputed agent transaction?
- What concentration risk comes from the largest assistant platforms?
- Which open standards should we test, or help shape?
- How will we serve customers who don't use advanced assistants?
- What is the smallest safe pilot that produces useful learning this year?
The short version
Foundations, not a finished market. South Africa has real payment infrastructure and active consultation, but no binding open-finance framework yet.
Five stages, in order. Readable information, authorised insight, recommendation and preparation, bounded execution, then competition, which no one bank can deliver alone.
Build what helps anyway. A capability inventory, product data, delegated identity, a policy gateway and an evidence layer improve today's channels too.
Pilot small and bounded. Read-only first; one capped own-account action; the customer sets reserves and the bank enforces them.
Postpone deliberately. Broad autonomy, new beneficiaries and binding credit decisions wait for evidence and rules.
The first-mover advantage will not belong to the bank that publishes the first MCP server. It will belong to institutions that learn how to provide dependable, competitive and governable financial capabilities through interfaces customers choose.
That completes the series. It began with Andrew Baker's great agentic unbundling, discussed in Part 1. The provocation holds up: when the agent becomes the interface, banks will compete harder for each product and transaction. The parts in between added what it takes to compete safely: protocols matched to their jobs, permissions that distinguish read from execute, mandates backed by evidence of judgement, products agents can read, and controls that assume the agent may be wrong. Whatever pace adoption takes in South Africa, those are capabilities worth having.
Start again from Part 1, or see every part on the series page.
How CloudNala can help
We help South African banks and financial-services firms plan agent-mediated banking at a pace their risk appetite supports: building the capability inventory and product data first, designing delegated identity, the policy gateway and the evidence layer, choosing pilots with bounded authority and clear stop criteria, and keeping the architecture views your board, risk and regulatory conversations need.
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