Banking economics have long relied on the primary relationship. A salary arrives at one bank, and the debit orders, card spend, savings, credit and rewards tend to follow it there. Each product makes the others stickier. Moving everything is inconvenient, so most people don't.
A personal agent changes the arithmetic without asking anyone to switch banks. It can manage several relationships behind one interface and make choices one product, or one transaction, at a time. The customer doesn't have to complete a full bank switch to stop giving one bank all of their activity. The agent does the administration that used to make spreading activity across banks not worth the effort.
This part looks at what that means commercially: how comparison changes, why "best" can't be reduced to a headline rate, what it could do to deposits, and what a bank would have to measure to see it happening.
Kestrel Bank and its customers are fictional, and their products, figures and systems are illustrative. Kestrel is a mid-sized South African retail bank. Its 32-day notice deposit pays a competitive rate. Since the bank started testing how agents see its products, that deposit has been losing comparisons it should be winning.
Relationship bundling becomes easier to unpick
| Today | Agent-mediated possibility |
|---|---|
| One bank is the default for most activity | The default can differ by product and context |
| Comparison is occasional | Comparison can happen continuously |
| Product complexity hides total value | Structured analysis can expose full cost and conditions |
| Switching takes the customer's effort | The agent can do the preparation and administration |
| Marketing influences attention | Machine-readable value influences the shortlist |
| Customer count suggests relationship strength | Product-level balances and flows show economic strength |
In the scenario from Part 1, Naledi, a physiotherapist in Gqeberha, kept her salary at Kestrel while her assistant moved her surplus to a better notice-savings account at another bank each month. Nothing about that looks like churn. She is still a Kestrel client, with an active account and her salary arriving on time. What changed is the share of her financial life that Kestrel earns from.
That is the shift in one picture. The bank still has the relationship, but the agent now decides, product by product, which bank does what.
Price matters, but "best" is multidimensional
The cheapest product can still be the wrong one. A responsible agent comparing savings, credit or payment options should weigh more than one number:
- total price and fees;
- interest or return after conditions;
- access and notice period;
- payment certainty and settlement time;
- fraud protection;
- the quality of dispute handling;
- service availability;
- eligibility and affordability;
- the customer's own preferences and values;
- what data is shared, and with whom.
How the agent ranks those options becomes commercially important. A bank with a slightly lower rate but immediate access, stronger protection and no fees may be the better choice for many customers. It only wins if the agent can see and understand those advantages. Banks will want assurance that agents compare like with like, and don't reduce suitability to the headline rate because it is the easiest field to read.
That assurance cuts both ways. As Part 1 noted, the agent isn't necessarily neutral. An assistant platform that ranks a commercial partner first, or calculates "best" in a way it doesn't disclose, can steer volume as effectively as any bank app ever did. The ranking method is where the next fight over fairness will happen.
Why Kestrel's deposit was losing
Kestrel's product team ran a simple test. They asked several general-purpose assistants, and one built by the bank's own innovation team, to recommend a notice deposit for a client with R80,000 to place and no need for the money within a month.
Kestrel's 32-day notice deposit was rarely recommended, even though its rate was among the best in the comparison. When the team examined the assistants' reasoning, the pattern was consistent:
- The rate was easy to find on the product page.
- The conditions were in a PDF: how interest is calculated, what happens if a notice is withdrawn, the minimum balance, and whether the rate is fixed for the notice period. The assistants either couldn't read it reliably or didn't try.
- Faced with unknown conditions, the assistants did what a careful adviser would do. They treated the unknown as a risk and preferred products whose terms they could read.
The product was competitive. Its information wasn't usable by an agent, and an agent that can't confirm a product's conditions has a good reason not to recommend it. Kestrel's problem wasn't price. It was that its terms were written for a person to read, not for an agent. Part 6 is about how Kestrel fixed it.
Deposit and funding implications
If agents sweep surplus balances towards better value, deposits may become less sticky. The possible effects include:
- faster response to rate differences;
- more volatile savings balances;
- pressure on funding costs;
- more precise deposit pricing;
- conditions and liquidity value becoming more important;
- products designed so that their conditions are easy for an agent to read.
This is a scenario, not an established market outcome. Adoption, regulation, transfer friction, tax, customer behaviour and trust will decide how large it becomes, and South African evidence on agent-mediated savings behaviour doesn't yet exist. A treasury team should model it as a sensitivity rather than a forecast. Ask how the bank's funding mix would behave if a defined share of surplus retail balances began moving monthly towards the best visible offer, and which products would hold their balances because of the conditions they offer.
Kestrel's treasury ran exactly that sensitivity after the Head of Digital Channels raised Naledi's pattern as a modelled scenario. The answer was not alarming, but it was specific. The balances most at risk were the ones that stayed only because moving them was inconvenient, not because the product was better.
Agent-mediated requests for offers
Andrew Baker's article, discussed in Part 1, imagines agents inviting banks to bid. He presents this as a forward-looking scenario, and it should be read that way. It is still worth designing for, because the shape of the request is easy to predict:
Ask eligible banks for a 12-month offer on R250,000. I need capital protection, no monthly fees and access with 32 days' notice. Present the best three and do not apply without my approval.
To respond, a bank would need to return a structured, time-bound offer that contains:
- the rate and how it is calculated;
- the term and access conditions;
- fees and penalties;
- eligibility;
- when the offer expires;
- protection and regulatory disclosures;
- what an application requires;
- the provider's identity, with evidence for it.
The protocols in Part 2 could carry the request and the responses. An agent card can advertise that a bank answers offer requests, and a bank-side service can return structured terms. A fair, regulated marketplace needs much more than message transport:
- participation rules;
- consistent schemas, so offers can actually be compared;
- verified identity for every participant;
- controls against unfair discrimination between customers;
- audit and dispute handling;
- market oversight.
None of these exist in South Africa today for agent-mediated deposit offers. A bank can build the response side now, and test it with its own channels and partners, without waiting for a market to form.
For Kestrel, the first useful step was internal. The team made the 32-day notice deposit able to answer a structured quote request from Kestrel's own app and from one partner assistant in a test environment, with real terms and an expiry. Even without any external market, the exercise forced the product, pricing and legal teams to agree what the offer actually is, in fields rather than paragraphs.
New measures for bank leaders
Customer counts don't show any of this. A bank that wants to see agent-mediated competition happening needs measures like these:
- agent-originated enquiries and conversion;
- how often a product makes the agent's shortlist, by product;
- share of wallet by product, not only by customer;
- retention after an agent comparison;
- response time for structured offers;
- permission activation and revocation;
- fraud and disputes in the agent channel;
- the cost to serve agent-mediated transactions;
- reliability and settlement evidence that agents can see;
- concentration by assistant platform.
Kestrel added two of these first, because they were the cheapest to start measuring and the most revealing. Share of wallet by product showed the thinning of salary-day balances that Naledi's pattern predicted. Shortlisting rate, measured through the bank's own test harness of assistants, showed which products agents couldn't understand. The second turned out to be the better early warning. A product that agents can't read loses before any customer sees it.
The last measure, concentration by assistant platform, matters more over time. If most agent-originated business arrives through one assistant, the bank has moved its dependence from its own app to someone else's interface. That is the platform risk Part 1 warned about, and it should be visible on the same dashboard as the revenue.
Strategic response
The strongest defence is not blocking comparison. A bank that makes itself hard for agents to use doesn't stop agents. It pushes them towards screen-scraping and guesswork, and its products lose comparisons by default.
The better response has two parts:
- Build products that survive comparison. If a product is only held because moving away from it is inconvenient, an agent will find that out. The products worth defending are the ones whose value holds up once fees, conditions, protection and service are counted.
- Make those products easy and safe for agents to use. Their terms should be readable, offers should come as structured quotes, actions should be governed, and there should be evidence of what happened.
The second part is the subject of the next article.
The short version
Agents unpick the bundle without a switch. A customer can stay with a bank and still move the profitable activity elsewhere.
"Best" is many numbers. Price, conditions, access, protection, dispute quality and service all count, and the ranking method is where the fairness questions sit.
Unreadable terms lose. Kestrel's competitive deposit lost because agents couldn't confirm its conditions.
Model deposit mobility as a scenario. Run it as a sensitivity, not a forecast. The balances held by inconvenience are the ones most at risk.
Measure products, not only customers. Share of wallet by product and shortlisting rate show the shift before customer counts do.
Previous: Part 4, a signed mandate is not proof of good judgement. Next: Part 6, make the bank machine-readable. All parts are on the series page.
How CloudNala can help
We help banks and financial-services firms understand how their products look to an agent: testing how assistants compare and shortlist your products, finding the conditions and terms an agent cannot confirm, modelling deposit and wallet-share sensitivities as scenarios rather than forecasts, and designing structured offer and quote services that your product, pricing and legal teams can stand behind.
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