AI now skips human sign-off at 80% of Australian firms

What the survey found

SecurityBrief Australia reported on 4 November 2026 on the Australian results of new Illumio research. Eighty percent of Australian organisations surveyed now let AI make decisions and take actions that needed human approval a year earlier. Only 55% can see and audit every action their AI agents take. A third let AI grant or revoke access permissions with no human review, and 26% allow AI to change code unsupervised.

Problems are already turning up. In the same Australian results, 73% of respondents said they had seen an AI-related security problem, or signs of one, during the past year. Thirty-five percent said AI had accessed data it shouldn't have. Thirty percent said AI tools or agents had connected to systems without security approval, and 28% said AI had produced or triggered an unauthorised action.

The global report, Licensed to Act, released by Illumio on 21 October 2026, comes from a Sapio Research survey of 1,100 IT and security decision-makers in 10 markets. Australia contributed 100 respondents and New Zealand wasn't surveyed. Australia is ahead of the global average on delegation: 70% of organisations worldwide have handed AI the actions that once needed approval, and 49% can audit every agent action. Illumio sells network segmentation and commissioned the research, and an Australian sample of 100 is small. The direction of the findings is still worth acting on, even if the exact percentages are rough.

Why the approval step is the weak point

In most business workflows, an approval step does two jobs. A person checks the action before it happens, and the approval leaves a record of who agreed to the action and when. A supplier bank-detail change waits for a finance manager. A new staff member's system access waits for their line manager. A refund over a set amount waits for a supervisor.

When an AI agent takes over a workflow like these, the approval step tends to get removed because it slows the agent down. The check goes with it, and so does the record unless someone deliberately builds logging into the new process. In the Australian data, 80% of organisations have removed approvals but only 55% can audit every agent action. That leaves roughly a quarter of respondents who gave up human sign-off without putting anything in place to show what the agent did instead.

The Australian Signals Directorate has been warning about exactly this. Christopher Horlyck, ASD's Head of Cyber Security Resilience, told SecurityBrief that "AI agents with excessive access or insufficient controls may take unintended or unauthorised actions." The same article says ASD's agentic AI guidance focuses on restricting access, keeping monitoring and audit logs, and keeping humans in charge of high-impact actions.

Respondents seem to know their existing controls aren't enough. Globally, 82% said AI agents need fundamentally different security controls from traditional software, yet only 25% named an explicit list of permitted actions as one of their two main controls. Most organisations agree the old controls don't fit agents, but few have written down what each agent is allowed to do.

What the results mean for NZ businesses

New Zealand wasn't in the sample, but NZ organisations buy the same cloud platforms, ERP add-ons and service desk tools as their Australian peers. Those products increasingly come with agent features switched on or one setting away. If you're adding AI to purchasing, onboarding, IT service requests or customer refunds, the same pressure to drop approval steps will apply in Auckland as in Sydney.

Any NZ business with Australian operations also has a deadline. From 10 December 2026, Australian Privacy Principle 1.7 requires APP entities to explain in their privacy policies when a computer program makes a decision, or does something substantially and directly related to a decision, that could significantly affect an individual's rights or interests, where personal information is used to run the program. The OAIC published its guidance on 30 September 2026. Under the guidance, the policy must list the kinds of personal information the programs use. It must also list the decisions the programs make on their own, and the decisions where a program carries out a related step. A business that has removed approval steps without logging what replaced them will struggle to write that disclosure accurately, because it can't say which decisions its software is making.

Removing an approval step can be the right call. Many approvals are rubber stamps that add a day's delay and no real scrutiny, and automating low-value approvals is a sensible use of AI. The risk lies in removing approvals without deciding which ones were doing real work. Access changes, payments and changes to master data such as supplier bank details are the actions most likely to end up in an incident report. Those are also the actions where the Australian figures show the most delegation, with a third of respondents letting AI change access unreviewed.

Next steps

• List your automated workflows that used to have a human approval. Include agent features turned on inside SaaS products as well as automations you built. For each one, note who used to approve and what records the process now keeps.
• Sort approvals into ones you keep and ones you drop. Keep a named human approver for access changes, payments, supplier bank details and any decision that affects a customer's or employee's rights. Let the agent handle routine items under clear thresholds.
• Write down what each agent can do. Give each agent its own service identity, list the systems and actions it's allowed to touch, and block everything else. A written list gives you something to test and audit against.
• Log agent actions where the agent can't change the record. Send agent activity to a log the agent has no write access to, then check whether you could reconstruct a week of the agent's actions from that log alone.
• If you trade in Australia, map your automated decisions before 10 December. Compare each workflow with the three APP 1.7 tests and update the privacy policy for the Australian entity.
• Ask vendors how their agent features are controlled. Ask what an agent can do by default, whether its actions can be logged and exported, and whether approval gates can be kept for selected actions.

If you're redesigning workflows to add AI, DataForge's process and automation team can help work out which approvals to keep and how to log the rest. If the bigger question is AI policy and governance across the business, see our AI consultancy services.