The Lazy Loop - The Quiet Failure of Human-in-the-Loop

Last month we explored AI use from the perspective of using it to research.

This month we are exploring AI use from the perspective of the process owner/decision maker.

What’s the obligation for Human-in-the-Loop in AI?

Here’s a statement that may surprise you: there is no overarching requirement in Australian law for a human to approve every AI-assisted decision.

I suspect there are a few faces being pulled at that opening line. Isn’t the decision point exactly what Human-in-the-Loop is designed to protect?

Not necessarily. It depends on the legislation governing each decision. Section 6A of the Social Security (Administration) Act 1999 (Cth), for example, computer programs make social security decisions on the Secretary’s behalf, without switching off the requirements of lawful authority and procedural fairness.

In Pintarich v Deputy Commissioner of Taxation (2018) 262 FCR 41, by contrast, a computer-generated ATO letter was held not to be a legally effective decision because no officer had reached a conclusion through a mental process. The lesson isn’t that every decision needs a human. It’s that where automation runs without proper authority or design, you may not be able to say whether a decision was made at all.

Private businesses face no universal rule either, but existing obligations make supervision necessary in practice:

  • Privacy Act 1988 (Cth): reasonable steps for compliance, accuracy and security (APPs 1.2, 10 and 11), and uploading client data into a third-party AI tool may be a disclosure needing justification (APP 6).

  • Competition and Consumer Act 2010 (Cth), Sch 2, s 18: claiming AI advice is reliable, or “reviewed by a person” when that review is a rubber stamp, risks misleading or deceptive conduct.

  • Corporations Act 2001 (Cth), ss 180 and 181: reliance on AI goes directly to whether directors have exercised care, diligence and good faith.

Timing matters

From 10 December 2026, new APPs 1.7–1.9 require privacy policies to disclose where a computer program is used to make or do something substantially and directly related to making decisions that could significantly affect individuals. Rules-based software is caught too. For APP entities, “where does software make or shape our decisions?” is about to become a question you must answer publicly.

What’s the problem?

The law is interested in whether the organisation designed, controlled and can stand behind the process. Yet the common mistake is treating a person at the end of the line as a control: automate, add someone to click “approve”, call it oversight. Set and forget. That is the Lazy Loop.

The reviewer sees the output, not the data, rules or prompts behind it, so most errors are baked in before they reach the human. Volume does the rest: someone approving hundreds of items a day isn’t exercising judgement, they’re keeping up.

Meanwhile models update, data drifts and the law changes, so a process sound at go-live may not be sound a year later.

The result is a human carrying accountability without the visibility, time or authority to exercise it. That isn’t oversight. It’s a liability sponge.

Automation does not remove governance or good business practice. It requires a deliberate layer on top, one that lifts people out of the grunt work and puts their judgement where it counts.

What should organisations be doing?

Engineering the loop properly means placing people across the whole life of the process, not just at the final click:

  • Before: IT, Legal and operations design the process together, identifying the decisions that matter, the legal duties and the risks.

  • During: automation handles the volume, while clear triggers escalate exceptions and high-impact items to a person with the authority to intervene. Each item must still be assessed on its merits.

  • After: results are sampled and reviewed against complaints, feedback and outcomes, and the process adjusted accordingly.

That is the uplift: the person who used to key in data now needs to test the system, handles cases needing judgement and improves the process. Less grunt work and a shift to governance focus.

To get there, organisations should be asking:

  • Are IT, Legal and operations all involved in engineering the process?

  • What does the law require of the decision maker to discharge their duty?

  • Where does software shape significant decisions about individuals, and does our privacy policy say so?

  • Who owns the outcome, and can they act on it?

  • Does the process address the individual merits of each item?

  • Does our review cover all use cases, feedback and complaints?

  • How often do we check that results are accurate, repeatable and defensible? 

Where do you start?

You don’t need to rebuild everything at once. Map where automation makes or shapes your decisions, then work outwards.

Automation doesn’t take responsibility off your plate. Done well, it takes the grunt work off your people so they can focus on the high value work that responsibility demands and assists accuracy, reduce re-work and complaints and improve service.

Who can help?

Our experienced team can support your organisation to adopt AI, develop workflows, maintain compliance and allow it to stay focused on its core business. We keep it simple, effective, and help you develop a clear strategy that moves you forward.

If you want to do ensure you get it right from the outset, contact us for a conversation.

Previous
Previous

Lawyers are using AI. But for what?

Next
Next

The AI Agents That Wouldn’t Take No for an Answer