Sixty per cent of Australian workers have personally spotted an error made by an AI tool at work. Almost half say the most significant error they saw had consequences beyond minor inconvenience, including financial, compliance or reputational impact.
The findings are from an Australian workplace study commissioned by process automation vendor Appian. Zoho Research surveyed 500 Australian workers in the third quarter of 2026.
More than one in five respondents said employees in their organisation are discouraged from challenging AI recommendations or decisions. Only 14 per cent fully trust the AI-generated outputs used in their organisation.
One in five workers said the most significant AI error they encountered caused a customer service issue. Thirteen per cent reported operational disruption.
The possibility of AI making incorrect decisions was the top concern for 25 per cent of respondents. Data privacy and security risks followed at 22 per cent, then a lack of transparency in how AI reaches decisions at 20 per cent.
"Employees are often the last line of defence before an incorrect output reaches a customer or becomes part of a business decision," said Charlie Hutchinson, SVP Asia-Pacific and Japan at Appian. "If they are discouraged from speaking up, organisations remove one of their most important safeguards."
Workers want human oversight
Ninety-four per cent said AI should not be allowed to make decisions affecting employees or customers without limits or human oversight.
Two-thirds (66 per cent) believe such decisions should require human review and approval, or always be made by people. A further 28 per cent would restrict AI to low-risk decisions.
Senior management remains more trusted than AI. Fifty-two per cent place greater trust in senior leaders, compared with 29 per cent who trust AI more.
Appian also cited global research from Harvard Business Review Analytic Services, which it sponsored. Ninety-two per cent of respondents agreed AI agents need rules-based guardrails to operate safely.
Less than half (48 per cent) said their organisation has defined those guardrails.
"Governance cannot be bolted on after AI is already influencing decisions," Hutchinson said. He said AI needs to sit within governed business processes that "control what it can access, define the rules it must follow, and determine when a decision must be handed to a person".
The findings echo debate over automated decision-making in Australia. Privacy Act amendments passed in 2024 require entities to disclose substantially automated decisions using personal information in their privacy policies. The obligation commences in December 2026.