The Explainability Gap: Belfast's AI Underwriters Need a Human Story

There is a peculiar feature of doing business in Northern Ireland that anyone who has sold anything here will recognise. The market is small enough that reputation propagates faster than advertising. A declined loan discussed at a family dinner in Ballymena, a disputed default shared in a church hall in Derry, an appeal rejected without explanation that becomes a story in a local paper — these things move. For any lender deploying automated credit decisions, that dynamic should sit at the centre of the business case, not in the annex on reputational risk.

This is an opinion piece, and the opinion is straightforward: the binding constraint on AI-driven lending in Northern Ireland is not model performance, capital or engineering talent. It is explainability, and the trust that explainability buys.

Compliance is not the same as comprehension

The regulatory position is not mysterious, even if it is demanding. Under the Consumer Duty, firms must act to deliver good outcomes, avoid foreseeable harm and demonstrate that products offer fair value. That obligation sits awkwardly beside a model whose reasoning no one in the building can articulate to a customer. Consumer credit rules have long required firms to give reasons when declining credit or varying terms, and the wider reform of the Consumer Credit Act is unlikely to soften that expectation. Data protection law adds a further layer, including rights around solely automated decisions with legal or similarly significant effects. Discrimination law — in Northern Ireland, the fair employment framework as well as equality legislation — constrains the use of proxies that correlate with protected characteristics.

The compliance box can be ticked with documentation. Comprehension is harder. A lender can produce a model card, a feature importance chart and a governance minute, and still be unable to answer the only question the customer asked: why me?

Where the models go wrong here

The technical literature on credit model failure is well developed. The local expression of it is distinctive. Thin credit files are common among younger applicants and among people who have never held a mainstream credit product. Self-employment is widespread in the agricultural and construction sectors, producing seasonal and irregular income that standard affordability models handle poorly. Cross-border working lives mean euro-denominated earnings, addresses in the Republic and UK credit histories that capture only half the picture. Rental payment data is patchy.

The most sensitive problem is postcode. In Northern Ireland, geography correlates with community background in ways that make it a hazardous variable. A model that leans on area-level features because they appear predictive is walking towards a discrimination risk that no amount of accuracy uplift can justify. The honest position is that many models do this implicitly, through aggregates the vendor supplied, and that few lenders have examined it properly.

Practical steps that actually move the needle

The work is unglamorous and specific. First, insist on plain-English reason codes that a customer service adviser can read aloud without translating. If the output requires a data scientist to interpret, it is not an explanation. Second, build counterfactual statements — what would have changed the outcome — and test them for plausibility. Telling someone that repaying a £40 mobile bill differently would have secured a mortgage is worse than saying nothing.

Third, sample declines for manual review. Not every decision, but a statistically meaningful share, reviewed by a person with authority to overturn. This serves two purposes: it catches model error, and it produces a feedback record that supervisors find persuasive.

Fourth, document the geography question explicitly. Show the analysis. Demonstrate that postcode features were tested, that their contribution to outcomes was measured, and that the choice to include or exclude them was deliberate rather than inherited from a vendor's default configuration.

Fifth, monitor for drift with the same discipline applied to arrears. Model performance degrades quietly when the customer base shifts, and in a market this small a shift in one sector can move a portfolio's risk profile within a quarter.

Why the human touch is a moat, not a cost

The conventional efficiency argument treats human underwriters as a legacy cost to be engineered away. For a lender operating at scale in a homogeneous mass market, that may hold. In Northern Ireland, it looks like a misreading.

A human reviewer handles precisely the cases where models are weakest: the farmer with three years of variable income, the returning emigrant with no local credit footprint, the applicant whose file reflects a period of illness rather than delinquency. These are not corner cases here; they are a meaningful share of the addressable market. Lenders who can underwrite them well will access borrowers their competitors decline, and will do so at risk-adjusted returns that automated rivals cannot replicate.

There is a second, less calculating argument. A decision reached by a person, explained in a conversation, and open to appeal generates a different quality of relationship than one delivered by an algorithm and a portal. In a market where trust is transmitted socially, that difference compounds.

The bottleneck nobody wants to name

It is tempting to frame this as a technology debate about model architectures and feature stores. It is really a debate about whether lenders are willing to slow down at the exact moment their investors want them to speed up. The institutions that get this right will not be the ones with the most sophisticated models. They will be the ones that can explain a decline to a customer, defend it to a supervisor, and still say yes to the borrowers nobody else will touch.

That is a harder thing to build than a model. It is also considerably harder to copy.