A provision arrives on the balance sheet looking like a fact. One line. One figure, often carried to the last rand. It sits beside the loans and advances with the same authority as the cash balance, as though it too were simply counted. It was not counted. It was decided.

Behind every expected credit loss is a chain of judgements that the single number is engineered to conceal. What counts as default — ninety days, or the first missed signal before it? When has credit risk increased significantly enough to move a loan from a twelve-month view to a lifetime one, and who set that threshold, on what evidence? Which macroeconomic futures were considered, and what probability was assigned to the one where things go badly? Where the model could not see a risk, what overlay was laid on top, and by whose hand? Each of these is a judgement made by a person. The figure is the last thing that happens in that chain, not the first. It is a conclusion wearing the costume of an observation.

This is worth saying plainly, because the accounting standard itself is honest about it. IFRS 9 did not replace judgement with a formula. It replaced the old discipline of waiting for a loss to occur with the harder discipline of estimating one before it arrives — expected loss in place of incurred loss, forecast in place of hindsight. That was the right correction; too little, too late was the failure of the last crisis. But the correction moved the weight of the accounts onto forecasts, scenarios and expert overlays. It made provisioning a forward-looking act, and forward-looking acts are acts of judgement whether or not we admit it.

Which is where the number becomes dangerous. A figure carried to the rand looks certain. The judgement beneath it is anything but. False precision is not a technical flaw; it is a form of concealment. It invites the reader — the board, the auditor, the regulator, the analyst — to trust the decimal places and forget the assumptions. This is the exact inversion of what a financial statement is for. The account exists to unconceal: to make visible what the business knows, and to be honest about what it does not.

Regulators have understood this, and have made unconcealment a rule. Post-model adjustments and management overlays — the places where human judgement is applied directly on top of the model — are permitted, but they must carry written justification and may not become a lever for smoothing earnings from one period to the next. The IFRS Foundation has flagged the governance of these overlays as a live concern, precisely because their subjectivity is so easily hidden. South Africa’s Prudential Authority has been explicit in its own directives on the regulatory treatment of accounting provisions. The message across all of it is the same: you may exercise judgement — you must exercise judgement — but you must be able to show your reasons. A provision that cannot explain itself is not prudent. It is opaque.

Nowhere does this matter more than at the edge of the formal economy, and that is where South African modelling is tested most severely. Consider the borrower with a thin file, or no file at all: the informal trader, the household paid in cash, the spaza owner whose entire business runs through a till the credit bureau has never seen. To a model trained on the formal economy, this person is not low-risk or high-risk. This person is noise. And a model handles noise by falling back on assumptions — assumptions built from populations that look nothing like them.

The consequences run in both directions, and both are borne by the person, not the spreadsheet. Under-provision, and the lender is eventually blindsided by losses it did not see coming — after which it does the rational, defensible, devastating thing: it withdraws from the segment entirely. Whole communities are rationed out of formal credit, left to the stokvel and the loan shark. Over-provision, and capital is tied up against losses that were never coming; credit is priced beyond reach, and inclusion is strangled by caution dressed as prudence. Either way, the judgement embedded in a provisioning model has decided who gets served and on what terms — and it has done so invisibly, inside a number nobody thought to interrogate.

This is the case for judgement stated at its sharpest. When the data is rich, a model can carry much of the load, and judgement can sit lightly on top. When the data is thin — which is to say, exactly when the stakes are human and the borrower is poor — judgement is not a supplement to the model. It is the discipline. The temptation is to let the model’s silence stand in for an answer, to provision by default and call it objectivity. That is not objectivity. It is abdication with good posture.

So provisioning is a judgement expressed as a number, and the discipline is not in producing the number. It is in refusing to let the number lie about where it came from. A provision is a claim about a future that has not yet happened, made by people who cannot fully know it, on behalf of people who will live inside its consequences. To provision well is to hold that honestly: to state what is known, to name what is assumed, to own the judgement rather than hide behind the arithmetic. The figure is only the signature at the bottom. The judgement is the whole document.

That is the standard. It does not flatter the model. It illuminates it.