The model declined the risk. The regulator asks why that was fair.
Underwriting is the decision to take a risk on, to load it, or to walk away. A model can rank applicants faster and more finely than any underwriter, but a decline is an adverse decision, and a fair-treatment regulator will ask the insurer to show that it rested on something it is allowed to use — and that the applicant was treated fairly in the making of it.
The underwriting model is one of the easiest wins to demonstrate and one of the hardest to govern. Fed an applicant's history, the data the insurer holds, and whatever third-party signals it can lawfully reach, a model will score the risk and recommend accept, decline, or a loading, in a fraction of the time a manual file takes. A European supervisory survey found half of non-life carriers and a quarter of life insurers already running such models in production, and African insurers under pressure to grow penetration and contain loss ratios are adopting them quickly. The capability is real. The exposure is in what follows the score.
A decline is not a neutral output; it is an adverse decision about a person, and the conduct regimes now treat it as such. South Africa's Treating Customers Fairly framework expects insurers to demonstrate fair outcomes across the whole product lifecycle, from design and promotion through to the decision to offer cover at all. Nigeria's reform act and its move to risk-based capital put underwriting discipline at the centre of supervision. Kenya's regulator, having tightened the rules on declined claims, is plainly attentive to the decisions that precede them. An insurer that can rank risk brilliantly but cannot explain why it turned a particular applicant away has automated its way into a fairness problem.
The global direction of travel removes any doubt about where this ends. The European Union's artificial-intelligence regime classes risk assessment and pricing in life and health insurance as high-risk, requiring documented risk management, bias testing, and a record for each decision, with penalties reaching tens of millions of euros. No African regulator has adopted that text, but the principle it encodes — that an automated adverse decision about a person must be explainable and tested for unfair discrimination — is exactly the principle the fair-treatment regimes already assert. The insurer that builds for explainability now is building for the supervision that is coming, not only the supervision that is here.
The African market sharpens the problem in a particular way. Penetration is low and growth depends on reaching applicants with thin files — the informal-sector worker, the first-time buyer, the microinsurance customer — for whom the insurer holds little conventional history and the model leans hardest on alternative and inferred data. That is exactly the data most likely to act as a proxy for income, location, or community, and exactly the population a fair-treatment regulator watches most closely. The drive to widen access and the duty to decide fairly meet directly inside the underwriting model, and an insurer that cannot show its basis is exposed on both counts at once.
The operational difficulty is that the underwriting model is usually a black box bolted to the workbench. It returns a score and a recommendation, and the underwriter either accepts it or overrides it, but in neither case is the basis preserved in a form a regulator could later read. When a declined applicant complains — and in a low-trust, low-penetration market, declined applicants complain to the regulator more readily than in mature ones — the insurer is left reconstructing a rationale after the fact, which is the weakest possible position to defend a decision from.
There is a further, quieter risk in the data the model reaches for. Underwriting, especially in life and health, draws on medical history and other sensitive personal data, and the more signal the model ingests, the better it scores and the larger the data-protection exposure grows. The underwriter wants the richest possible picture; the data protection officer has to be able to say what sensitive data the model saw and on what basis. That tension, left unresolved, becomes a finding under the data-protection regime running alongside the insurance one.
The insurers that come through this are not the ones with the sharpest model. They are the ones who can attach, to every decline and every loading, a basis that rests on permitted factors, that was tested for unfair discrimination, and that an underwriter can show to a regulator without reconstructing it from memory. The score is a commodity. The defensible adverse decision is not.
A decline the underwriter cannot justify is not risk selection. It is a fair-treatment finding the model wrote on the insurer's behalf.
Where each sits.
Akki governs which data fields enter the underwriting model and logs the basis on which each is processed, so the insurer can state precisely what the model saw — including which sensitive medical fields, on what lawful basis — and reproduce the decision exactly. The walk-back from a disputed decline to the data and factors behind it becomes a query rather than a forensic reconstruction.
Solva structures the underwriting reasoning through its five stages and produces the basis underneath each accept, decline, or loading — the factors that drove it, their permitted grounding, and the confidence the evidence warranted. Where the basis would rest on something the insurer cannot justify, or where the signal is too thin to support a decline, Solva surfaces that rather than ratifying the score. The underwriter who declines a risk can show why; the regulator who queries it sees a reasoned, tested decision rather than a black-box output.
In life and health underwriting, SyniSense earns its place: medical history and other sensitive data are anonymised at the perimeter before the scoring model reasons over the risk pattern, and re-identified inside only to issue the decision. The model assesses the risk without holding the identifiable medical record off-platform. In purely commercial or motor underwriting its role is lighter, and the weight sits with Akki and Solva.
For the underwriter, the decline carries its own justification. The decision is made against a reasoned, tested basis rather than a bare score, so the underwriter is no longer ratifying an output they cannot explain and would have to defend alone if it were challenged.
For the data protection officer, the sensitive-data question has an answer. What medical and personal data the model reasoned over, and on what basis, is logged per decision, and in life and health the identifiable record never leaves the perimeter — which is what the data-protection regime sitting alongside the insurance one requires the insurer to be able to demonstrate.
For the applicant, the protection is structural. An adverse decision rests on permitted factors and has been tested for unfair discrimination, which is the fair-treatment outcome the conduct regulators are reaching for and the trust the market most lacks.
For the board and the regulator, the underwriting book stops being a latent conduct liability. The insurer can show, across the portfolio, that its automated decisions are explainable, grounded, and tested — the posture the fair-treatment regimes demand today and the high-risk regimes will demand tomorrow.