Governance
What human in the loop really means for AI in lending
By Saby Saxena4 min read
In short
What meaningful human oversight of AI looks like for Australian lenders: three levels of oversight, what regulators expect, and review points that work.
Almost every AI vendor in financial services now says their product keeps a human in the loop. The phrase is everywhere, and it’s starting to lose its meaning. For Australian lenders, the useful questions are more specific. Where exactly do people step in? What do they see when they do? And who is accountable when something goes wrong?
Getting this right matters. Regulators are watching how financial firms adopt AI, and borrowers are increasingly aware of it too. Good human oversight isn’t a box to tick. It’s what lets you use AI with confidence.
What human in the loop really means
Human in the loop means a person reviews and approves important actions before they take effect. In lending, that usually covers anything that changes a borrower’s application, account or outcome. AI prepares the information and the recommendation. A person makes the decision and owns it.
That sounds simple, but it falls apart if the review is rushed or the reviewer can’t see why the AI suggested something. A human who clicks approve on 200 items an hour without context isn’t really in the loop. Meaningful oversight needs the right information, enough time and the authority to say no.
Three levels of oversight
It helps to separate three different models, because most lenders will use more than one.
- Human in the loop: a person approves each important action before it happens. Best for credit decisions, hardship outcomes and anything affecting what a borrower pays.
- Human on the loop: AI acts within tight, pre-set limits while people monitor and can step in. Suitable for things like routine reminders or booking a call back.
- Human out of the loop: AI acts entirely on its own. This is rarely appropriate for regulated lending decisions.
The right level depends on the impact and reversibility of each action. Sending a reminder about a missing document is low risk and easy to fix. Changing a repayment arrangement is neither.
What Australian regulators are saying
There isn’t a single AI law for lenders in Australia, but existing obligations already apply. ASIC’s October 2024 report, Beware the gap (REP 798), reviewed how financial services and credit licensees were using AI. It found that some firms’ governance hadn’t kept pace with their AI use, and urged licensees to make sure their risk management and oversight keep up as adoption grows.
For APRA-regulated lenders, CPS 230 Operational Risk Management has applied since 1 July 2025. It expects clear ownership of critical operations and proper management of material service providers, which can include AI providers. CPS 234 sets expectations for information security.
Privacy is changing too. From 10 December 2026, organisations will need to explain in their privacy policies when they use automated decisions that could significantly affect individuals. If AI plays a part in decisions about borrowers, lenders will need to be able to describe it clearly.
Hardship and vulnerability need extra care
Some situations should always involve a person. Under the National Credit Code, borrowers can ask for a hardship variation, and lenders need to handle those requests within set timeframes. Conversations about hardship, family violence, illness or financial abuse need empathy and judgement. AI can help by spotting signs early and routing the borrower quickly to the right team, but it shouldn’t be deciding the outcome.
Designing review points that actually work
- Show the evidence, not just the recommendation. Reviewers should see the documents, data and reasoning behind each suggested action.
- Make it quick to approve, edit or reject, so people aren’t tempted to rubber-stamp.
- Record who approved what, when and why, in a form your audit and compliance teams can use.
- Escalate unusual cases automatically rather than letting the AI guess.
- Review samples of AI activity regularly, not just the items that were flagged.
Read more about our approach on the security and governance page.
Why it’s worth getting right
Strong oversight is sometimes seen as the thing that slows AI down. In practice it’s the opposite. When your team trusts that every important action is reviewed and recorded, it becomes much easier to extend AI into new areas. Weak oversight is what causes projects to stall when risk or audit start asking questions.
Frequently asked questions
Does human in the loop slow everything down?
Not if it’s designed well. When AI prepares complete information and a clear summary, a person’s review is often faster than doing the work from scratch.
Who is accountable when AI is involved in a lending decision?
The lender remains accountable for meeting its obligations. That’s why approvals, records and escalation paths need to be built into the workflow from day one.
Is there an Australian standard for AI oversight?
The Australian Government released a Voluntary AI Safety Standard in 2024 that includes human oversight as one of its guardrails. It’s voluntary, but it’s a useful reference alongside ASIC and APRA expectations.
To see how review points work in onebank, book a demo.
See it on your own scenarios
onebank runs the work in between. Your team approves every decision.