Northwind Bank's Loan Pre-Screening Pilot

Fairness. The 62% vs 78% approval gap across postal-code clusters is an outcome disparity. The fact that overall accuracy beats the human team does not neutralise it: fairness is judged on error rates and outcomes per group, not on aggregate accuracy. Postal code is also a classic proxy variable, so the disparity may be caused by a feature nobody intended to be sensitive.

Transparency. A single 0-100 score with no explanation field and a vendor black box means neither the applicant nor the loan officer can be told why. Transparency requires intelligible, meaningful explanation, not just a number.

Accountability. Nobody owns the outcome: marketing is writing claims, finance is removing reviewers, and the vendor supplies an opaque model. Accountability needs a named human owner for the system's behaviour.

Inclusiveness / human oversight. The unanswered regulator question about appeals shows there is no route for a person harmed by an error to contest it.

A weaker candidate answer is 'privacy'. The evidence is thin unless you argue the postal-code proxy or the ten years of historical decisions encode personal data - worth a sentence, not the headline.

The proposal fails the human-oversight and accountability tests in a specific way: it removes the human at the point where the consequence is worst. The model has a demonstrated group disparity and no explanation field, so an unreviewed automated decline is both unexplained and potentially discriminatory.

Minimum viable alternative within eight weeks:

1. Asymmetric human-in-the-loop. Route every decline above a risk band to a human reviewer; let approvals at the high-confidence end flow automatically. This protects the harmed party while still capturing most of the volume saving - quantify it, because the saving will be smaller than 40% and finance should see the number.
2. Reason codes from a surrogate. You cannot retrain the vendor model, but you can fit a simple interpretable model to its scores and report the top contributing factors per decision, labelled honestly as an approximation.
3. A real appeal path with a named owner and a service-level commitment, answering the regulator.
4. Shadow monitoring of approval rates by group during the pilot, with a defined tripwire that halts automation.

The judgement being tested: you cannot fix fairness in eight weeks, so you contain the harm and instrument for evidence instead of pretending the disparity is solved.

Rewrite: 'Our lending specialists use an AI-assisted score to speed up the initial assessment; a trained underwriter makes every final decision, and every applicant can request the reasons for a decline and appeal.'

Justification: the original claim is a transparency failure and an accountability failure at once. It tells the public that no human is responsible, which is untrue under the recommended design and would be indefensible in a complaint. It also implies a level of autonomy the system does not have, which is the kind of overstatement that turns a technical incident into a regulatory one.

The rewrite does three things: it names the human decision-maker (accountability), it discloses the assistance and the appeal route (transparency and inclusiveness), and it makes no claim about accuracy or fairness that the back-test does not support. Note that this is not cosmetic - if the press release is published as drafted, the bank must actually run the fully automated process it described, which is exactly the process finance wanted and which the disparity data does not support.