Can AI Speed Up Loan Decisions?

The problem, in numbers

Much of the delay in a loan decision is handling, not judgement. When Australia's major banks answered a parliamentary committee's questions on home loan approval times, as reported in July 2021, Commonwealth Bank said branch applications took an average of two days to approve and broker applications an average of 16 days.^{Source 1} These are the banks' own figures, reported by trade press. The reasons they gave were largely about process. NAB said broker applications needed more manual work and less automated processing. Bank of Queensland said broker loans arrived through a gateway and had to be manually rekeyed into its home loan system. CBA said broker applications all went to its operations team because brokers cannot verify certain regulatory obligations on a bank's behalf.^{Source 1}

In our reading, that is the work automated underwriting targets: re-keying, document chasing and verification queues, more than the credit judgement itself. It is also where the law applies. Under Australia's National Credit Act, a lender must make reasonable inquiries about a consumer's requirements, objectives and financial situation, take reasonable steps to verify the financial situation, and assess whether the loan is unsuitable before entering into it.^{Source 4}

What improves, and why

Automation shortens the clerical parts of a loan file: data captured once is not typed in again, and transaction data arrives electronically. ASIC's responsible lending guidance anticipated this: it says open banking and digital data capture services may help licensees streamline their process, potentially letting them complete both inquiries and verification.^{Source 4}

The guidance is technology neutral. ASIC says licensees can meet their obligations online or face to face, though they may need to adjust their processes.^{Source 4} The obligations also scale: the inquiries and verification that are reasonable depend on the consumer's circumstances and the risk of harm, which ASIC describes as "scalability".^{Source 4} In our reading, that leaves room for automation to clear simple, lower-risk files quickly while routing harder ones to people.

What the evidence actually says

The strongest independent evidence on speed comes from US mortgages, and it predates today's machine learning tools. A peer-reviewed study of loan-level US data found that technology-based lenders, which grew from 2% to 8% of the market between 2010 and 2016, processed mortgage applications about 20% faster than other lenders, even after controlling for loan, borrower and location characteristics.^{Source 2} Faster processing did not come with higher defaults.^{Source 2} The authors identified these lenders by an end-to-end online application and centralised underwriting augmented by automation.^{Source 2} In our reading, the study supports automated processing in general, not AI models in particular.

Speed is not the same as a good outcome for the borrower. In the US Federal Reserve's 2023 Small Business Credit Survey, online lenders, which use data and technology to underwrite and price loans, had the lowest satisfaction of any lender type: only 15% of approved online lender applicants were satisfied with their lender, on net.^{Source 3} High interest rates (55%) and unfavourable repayment terms (42%) were the challenges those applicants cited most.^{Source 3} These are self-reported results from a weighted convenience sample of US small firms, and do not isolate AI. But a quick decision can still be a poor deal.

We found no independent, peer-reviewed study measuring how much machine learning models, as distinct from digitised workflows, shorten loan turnaround in Australia. We have not used lenders' own claims about how much automation shortens approval times. The speed evidence is moderate and mostly American, and we grade confidence down for that reason.

Where it falls short

• Explanations are owed. Under EU data protection law, the Court of Justice held in 2025, in a case about a refused mobile phone contract, that a person refused on an automated credit assessment is entitled to an explanation of the procedure and principles actually applied, enough to understand and challenge the decision. Handing over the algorithm is not enough.^{Source 7}

• The score can be the decision. In 2023 the same court held that a credit score is an automated individual decision, prohibited in principle under the GDPR, where lenders give it a determining role in granting credit.^{Source 6}

• Explanation tools are imperfect. BIS researchers say existing explainability techniques have notable limitations, including inaccuracy, instability and susceptibility to misleading explanations, and that complex models would struggle to meet existing model risk rules on documentation, validation and monitoring.^{Source 8}

• Complex files stay slow. Westpac linked slower broker approvals partly to self-employed applicants and construction loans, and BOQ to construction loans.^{Source 1} In our reading, these are the files automation clears least well.

Where governance decides the outcome

ASIC's guidance sets the baseline for automated credit decisions in Australia. Lenders using automated decision systems should test the system, rules and algorithms before implementation and at reasonable regular intervals, make sure it can identify files that need further inquiries or verification and either complete them or refer the application for manual consideration, and make sure it can produce a meaningful record of the assessment.^{Source 4} If a consumer asks, the lender must give a written copy of the assessment free of charge, showing the information relied on and how it was verified, though not commercially sensitive lending criteria.^{Source 4} From 10 December 2026, privacy policies must also describe the kinds of personal information used, and the kinds of decisions made solely by a computer program or where it performs a directly related step, when those decisions could significantly affect an individual. The OAIC lists access to essential banking and credit services as such an interest, and its guidance says a program can be in scope when a loan officer relies heavily on its risk score.^{Source 5} In our reading, a lender should be able to explain an automated refusal in terms a customer can act on before the model goes live.

How to think about adoption

Map where the time goes in your pipeline before buying a model. If the delay is re-keying, document collection and verification queues, digitised capture and straight-through processing for simple files will do much of the work, and that is where the independent evidence points. Set referral rules so the system hands harder files to a credit assessor rather than declining them by default. Generate the written assessment at the time of decision, which ASIC calls good practice even when nobody asks.^{Source 4} Test any model before launch and on a schedule, and watch complaints, early arrears and hardship as signals that fast assessments are missing something, as ASIC suggests.^{Source 4} Treat any vendor's decision-time figure as a claim to check against your own application-to-approval times.