Decision type affects the review approach

Reviews of algorithmic systems look the same at a high level.

Is the model fair? Is it accurate? Are the right attributes being used?

Those are the right questions, but a key factor in how we answer them is the type of decision being made.

This sounds obvious, but some reviews don’t account for the decision type. When this happens, the review either misses important things, or uses a weaker approach than it could. The result can be a less useful review outcome: partial confidence and missed opportunities.

Algorithms for mortgage approvals, insurance pricing decisions, credit limit adjustments, and fraud filters may all use models. But they differ in specific ways: how they affect customers, the harm profile when something goes wrong, regulatory expectations, and the features most likely to introduce problems.

Reversibility

How easily can a decision be undone?

Some are easier. A credit limit can be raised next month. A fraud flag can be lifted when the customer calls in.

Others are harder to undo. A mortgage decline happens at a specific moment: when the customer needs it. A premium that's too high for one group can run for years before anyone notices. A claim that's rejected during a period of financial stress can't be fully remedied after the fact, even if it's eventually overturned.

How reversible a decision is contributes to the focus and the approach. It informs our tolerance for uneven outcomes, and how much weight we give to false negatives versus false positives.

Who bears the cost of an error

In most algorithmic decisions, the cost of an error doesn't fall equally on the institution and the customer.

A wrongly declined mortgage applicant loses the property. The institution loses a loan it would have made. These are not equivalent outcomes.

A wrongly frozen account or an overpriced policy works the same way: the disruption to the customer is higher than the disruption to the organisation.

When the customer bears most of the downside, we may need to be more conservative about model changes, and quicker to investigate early warning signs like spikes in complaints.

On the flip side, organisational risk gets regulatory priority for some types of decisioning systems (e.g. AML). Reviews need to reflect that too.

Regulatory context

Some decision types have decades of regulatory scrutiny behind them.

Mortgage lending, for example, has been a focus of fair lending law for a long time. The expectations are well established.

Others are newer territory. Algorithmic fraud detection is attracting more attention, but the regulatory framework is still developing. That doesn't mean the risk is lower, but it may mean it's less visible.

Hidden or emerging expectations can still result in scrutiny after something has gone wrong. For newer areas, internal standards often need to be ahead of formal rules.

What this means for reviews

When scoping a fairness or accuracy review, we consider the decision type. What harm does an error cause, and to whom? What are the regulatory expectations? Which features are most likely to act as proxies for protected attributes in this specific context?

The answers will be different depending on what the model is doing, and the review should reflect that.

Of course, many review frameworks already tier by materiality and impact. Being explicit about decision type helps sharpen these tiers. The core tests may stay the same, but the emphasis, thresholds, and follow‑up questions might shift.

Disclaimer: The info in this article is not legal advice. It may not be relevant to your circumstances. It was written for specific contexts within banks and insurers, may not apply to other contexts, and may not be relevant to other types of organisations.