AI Underwriting: The Quiet Revolution

Key takeaways

• Document handling and verification are heavily automated already.
• The credit decision remains dominated by interpretable models.
• Explainability obligations are the binding constraint, not technical capability.
• Bias and fairness testing is a live regulatory concern.

Where has automation actually landed?

The pattern is consistent: automation has taken the work where being wrong is cheap and checkable, and stopped where being wrong is expensive and must be justified.

Why has the credit decision resisted?

Because a lender has to be able to say why. A consumer declined for credit may ask for reasons, a regulator may examine the decision, and AFCA may review it in a dispute. A model that produces an accurate answer nobody can explain does not satisfy any of those.

That is why logistic regression scorecards remain widespread in credit decisioning long after more complex methods became available. The scorecard maps directly to points on variables, so the reason for a decline is legible.

There is a second constraint: fairness. A model trained on historical lending decisions learns historical patterns, including any discrimination embedded in them. Because protected attributes can be proxied by permitted variables - postcode standing in for ethnicity, for example - lenders must test that permitted inputs are not doing prohibited work.

What does it mean for you?

Practically, three things.

• Straightforward applications move faster. If you are salaried with standard security, automation is working for you.
• Complex applications may be declined faster, without ever reaching someone who could have understood the context. A quick decline is not a considered decline.
• You can ask for reasons. If you are declined, ask what drove it. You are also entitled to access your credit report and to have errors corrected.

If an automated process has produced an outcome you believe is wrong, escalate rather than reapplying elsewhere immediately - each application leaves an enquiry on your file. Ask for manual review through the lender's internal dispute resolution, and take it to AFCA if that does not resolve it.

Frequently asked questions

Do banks use AI to approve home loans?

Automation is heavily used in document processing, income verification, valuation and fraud detection. The credit decision itself still relies substantially on interpretable models such as scorecards, because lenders must be able to explain decisions.

Can I ask why I was declined?

Yes. Ask the lender for the reasons and request manual review through its internal dispute resolution process. You can also access your own credit report and have any errors corrected.

Is AI-based lending fair?

It depends on how the model is built and tested. A model trained on historical decisions can learn historical bias, and permitted variables can act as proxies for protected attributes. Fairness testing is a recognised obligation and a live regulatory concern.

Will a human ever look at my application?

For straightforward applications at a digital-first lender, possibly not unless something falls outside policy. Complex applications, exceptions and hardship assessments are still handled by people at most lenders.

Related reading

• Credit Scorecards: How Lenders Decide
• Digital-Only Lenders: Speed vs Flexibility
• Open Banking and CDR: How Lenders Now See Your Spending

Sources

• Credit licensing: Responsible lending conduct (RG 209) — ASIC
• Australian Privacy Principles guidelines — OAIC