The decision becomes the product.
For a generation, South African financial institutions competed on products. The next contest is over decisions: who gets credit, cover, an account or a flag, how fast, and whether the institution can say why. The institutions that pull ahead will be the ones that can show, for every automated decision, who owns it, what it earns and how the customer can answer it.
This is the first edition of The Possible: a free future view of one industry at a time, published by Digital Edge once a fortnight. We start with financial services because it is where South Africa's institutions already make the most decisions about people, at the highest volume, under the most law. What happens here will set the pattern for every sector that follows.
A future view is not a forecast. We do not tell you what the market will be worth in 2031 or which bank will win. We describe what is now possible, what the evidence says is standing in the way, and what a well-designed institution could look like if it chose to build for it. Every fact carries a named source at the foot of this page. Every opinion is ours to defend.
Our argument in one line: in financial services, artificial intelligence is turning the decision itself into the product. Credit, cover, onboarding and fraud control are all, underneath, decisions about a person. The institutions that treat each of those decisions as something designed, owned, measured and explainable will capture the value. The rest will run the old institution faster and wonder where the return went.
1. The value gap is an operating-model gap
The global evidence on AI value is now consistent, and it is sobering. Boston Consulting Group's 2025 study, The Widening AI Value Gap, surveyed more than 1,250 companies. Sixty per cent report little or no material value from AI. Thirty-five per cent are scaling and starting to see some. Only 5 per cent, the group BCG calls “future-built”, achieve value at scale, and they show 1.7 times the revenue growth and 1.6 times the EBIT margin of the laggards.
Everyone can rent the same models. The gap between the 60 per cent and the 5 per cent is a gap in how the work is designed.
BCG adds the finding that matters most for a financial institution: 70 per cent of AI's potential value sits in core business processes, not in support functions. In a bank, an insurer or a lender, the core processes are the decisions. Underwriting. Claims. Onboarding. Collections. Advice. Fraud. That is where the value is, and it is also where the risk, the regulation and the customer's dignity sit.
Most institutions began somewhere else: summarising meetings, drafting emails, searching policy documents. Those pilots were sensible places to learn. They are not where the money is. Closer to home, the ITWeb Brainstorm CIO Survey 2026 found that 49 per cent of organisations have generative AI in production to some degree, and 26 per cent of CIOs name operationalising AI as their single biggest pressure point. Adoption is no longer the question. Moving the core is.
What becomes possible. Imagine an institution where every AI use case is registered against two things before it ships: the core process it changes, and the line in the profit and loss account it is meant to move. A use case that cannot name both does not go live. Every quarter, the executive committee reads a single page that shows what each one has actually done to that line. The 60 per cent is not a fate. It is what happens when nobody is required to write that page.
2. Automated decisions will have to explain themselves
South Africa has a law that already anticipates most of this. Section 71 of the Protection of Personal Information Act (POPIA) says, in summary, that a person may not be subject to a decision that has legal consequences for them, or affects them to a substantial degree, if it is based solely on automated processing of personal information intended to profile them. The section names creditworthiness, reliability, performance at work, location, health, personal preferences and conduct among the things such a profile might assess.
There are exceptions. A solely automated decision is allowed in connection with a contract, where the person's request was met or appropriate measures protect their legitimate interests, or where a law or code of conduct provides for it with such measures. Those measures must give the person an opportunity to make representations about the decision, and require the institution to give them sufficient information about the underlying logic of the automated processing so that they can.
Read that list again with a financial institution in mind. Creditworthiness is named in the Act. Reliability, conduct and location describe most fraud and affordability models. Section 71 is not a niche clause for a future AI regime. It is a description of the decisions financial institutions make every minute of the day.
Two forces are about to put it under pressure. The first is volume: as more of the core runs on models, more decisions cross the line from “assisted” to “solely automated”, often without anyone choosing that they should. The second is agents. Gartner predicts that 40 per cent of enterprise applications will include task-specific AI agents by the end of 2026, up from under 5 per cent in 2025. It also predicts that more than 40 per cent of agentic AI projects will be cancelled by the end of 2027, because of cost, unclear value or weak risk controls. An agent that declines, prices, flags or freezes is making a decision about a person. Somebody has to own it.
What becomes possible. Imagine every automated decision about a customer arriving with what we call a Decision Receipt: what was decided, the main reasons, what could change the outcome, how to respond, and a named person with the authority to look again. Built once into the decision pipeline, the same receipt serves section 71, the complaints process, the internal auditor and the customer. The institution stops treating explanation as a legal cost and starts treating it as part of the product. A declined applicant who understands why, and knows what to change, is far more likely to come back than one who receives a form letter.
3. Identity becomes shared infrastructure
Every decision in financial services starts with a prior one: is this person who they say they are? That question is about to be asked at a scale and a price South Africa has not seen before.
The Financial Intelligence Centre's annual report for 2025/26 records 60,115 institutions registered with it at 31 March 2026. They include 37 banks, 206 life insurers, 2,697 credit providers, 257 money or value transfer providers, 362 crypto asset service providers and 219 foreign exchange dealers. Schedule 1 of the FIC Act was widened by Act 22 of 2022 to bring in crypto asset service providers, money or value transfer providers, trust and company service providers and high-value goods dealers. Each of these institutions must know its customer.
The rules are changing shape at the same time. The FIC's revised Guidance Note 7A, published on 1 September 2025, is risk-based and method-neutral: it tells institutions what they must achieve, not which tool to use. Guidance Note 7B, issued in June 2026, addresses biometric attributes as a means of verification. Under POPIA section 26, biometric information is special personal information, and processing it needs a basis under section 27. South Africa was placed on the Financial Action Task Force's grey list in February 2023 and removed on 24 October 2025. Staying off it depends on institutions continuing to show that customer due diligence works in practice, not only on paper.
And the economics have shifted. On 1 July 2025 the Department of Home Affairs raised the price of a real-time check against the National Population Register from R0.15 to R10, with batch checks at R1 per field. A check that was once a rounding error is now a line item. Meanwhile the threat grows: the South African Banking Risk Information Centre (SABRIC) reported 97,975 digital banking fraud incidents in 2024, up 86 per cent, with gross losses of R1.888 billion, up 74 per cent. MarketsandMarkets, a commercial research firm, estimates the South African identity verification market at USD 171.3 million in 2025 and forecasts USD 385.7 million by 2030. Treat that as a vendor estimate, but the direction is not in doubt.
What becomes possible. Imagine identity treated as infrastructure rather than as a form at the front of every product. Verification designed once, to the risk-based standard Guidance Note 7A asks for, with biometric processing on a clear section 27 basis, priced per check so a small credit provider pays for what it uses rather than for a platform it cannot afford. And imagine that layer designed with section 71 in mind too, because a failed liveness check that locks a genuine customer out of their money is itself an automated decision about a person. The institutions that get identity right will onboard faster, lose less to fraud and turn away fewer real customers. Those that bolt it on will pay R10 a time to learn the same lessons slowly.
4. The operating model: decision rights before technology
None of the three shifts above is mainly a technology problem. Each one asks the same institutional question: who owns this decision, under which rules, and how would we know if it were going wrong?
Most financial institutions cannot yet answer that quickly. Decision rights for credit, model risk, compliance, product and technology sit in separate committees with separate registers. A model can be approved by one, deployed by another and complained about to a third. When a regulator, an auditor or a customer asks why a decision rule changed, the answer is rebuilt from minutes and memory. We argued in our essay on the institution as an operating system that technology runs on top of these arrangements and inherits every flaw. In financial services, AI makes the flaws run faster.
The same logic applies to Islamic finance. As Shariah-compliant books grow and automate, Shariah compliance can no longer be an audit that happens after the fact. It has to be a property of the workflow, checked at the moment of decision, in the way we sketched in our concept Shariah by Design.
What becomes possible. Imagine a single decision register, kept the way engineers keep version control. Every material decision about a customer, human or machine, has a named owner, a current rule version, a value case, the controls that apply to it and the customer's route to answer it. Change a rule and the register records what changed, when, why and on whose authority. The board reads the institution's operating system directly. We call this the Institutional Change Log. It is not glamorous. It is the thing that lets an institution move fast without losing track of what it has become.
Five years out: one institution, imagined
Picture a mid-sized South African lender in 2031. It is not a real firm and not a client; it is a sketch of what the evidence above makes possible.
- Its AI portfolio is short. Every use case sits in a core process, credit, collections, onboarding or fraud, and each one is reported against the profit line it was built to move. Pilots that could not name that line were stopped early and cheaply.
- Every automated decision about a customer leaves a receipt. Declined applicants can see the main reasons, what would change the result and whom to talk to. Complaints fall, not because fewer people are declined, but because fewer are left guessing.
- Identity is a layer, not a form. Customers are verified once to a risk-based standard, biometric data is processed on a documented basis, and the cost of each check is known and managed.
- Its agents act within limits a named person set, and that person can stop them. When one goes wrong, the institution knows within the hour, not at the next audit.
- Its decision register is the first document the board reads. When the Information Regulator, the Prudential Authority or a customer asks why, the answer takes minutes.
Nothing on that list needs a technology that does not exist today. All of it needs decisions about ownership, design and measurement that most institutions have not yet made.
Five questions for your executive committee
- Can we list every AI use case in production, and name the core process and the profit line each one is meant to move?
- Which of our decisions about customers are made solely by automated means, and could we give a declined customer sufficient information about the underlying logic tomorrow?
- Does every automated decision have a named owner: a person, not a committee?
- What does verifying one customer cost us end to end, including the R10 Home Affairs check, the retries and the genuine customers we lose at onboarding?
- If an AI agent acted on our behalf tomorrow, who could stop it, and how quickly?
What this edition does not claim
It is not a forecast, and we have produced no market sizing of our own. Where we cite a commercial estimate we say so. It is not legal advice: we summarise POPIA section 71 and the FIC guidance for discussion, and your legal and compliance functions should advise on your specific obligations. The institution in the final section is imagined.