Australian business leaders split on AI accountability

ABBYY has published research showing Australian business leaders are divided over who should be responsible when AI systems go wrong. The findings point to a governance gap despite broad AI use in larger organisations.

The study, conducted by Opinium, surveyed 1,200 senior managers at companies with more than 100 employees across six countries, including Australia. Within the Australian sample, views on accountability were spread across organisations, vendors, end users and regulators, suggesting there is no settled view on where liability should sit when AI produces harmful or incorrect outputs.

Australian respondents were split almost evenly between two positions. Some 27% said the organisation using AI should bear responsibility, while another 27% said it should be shared between the organisation and the vendor supplying the technology.

A further 17% said the AI vendor alone should be accountable. Another 15% pointed to the individual or end user, while 14% said policymakers and regulators should carry ultimate responsibility.

Trust and risk

The survey also found a gap between confidence in AI accuracy and confidence in its safe operation. Australian leaders recorded net trust of 68% that AI produces accurate outputs, but only 17% said they completely trust AI systems to operate without introducing unacceptable risks.

Net trust on the risk question stood at 60%, lower than the result for accuracy. The figures suggest that while many businesses see AI as useful, a sizeable group remains cautious about errors, security failures and weak governance.

Concerns about output reliability were reflected in the risks cited by respondents. More than a quarter of Australian organisations, or 28%, said hallucinations leading to incorrect business decisions were an AI-related security risk.

Data exposure ranked even higher. Some 45% of respondents identified confidential data leakage as their top security concern, making it the most commonly cited risk in the Australian findings.

Human review

Australia recorded the highest level of human involvement among the markets surveyed. According to the research, 61% of business leaders said humans review all important AI decisions before action is taken.

That puts Australia ahead of the other countries in the study, suggesting organisations here are more likely to keep people in the loop for consequential uses of AI. Even so, the results also show notable gaps in oversight.

Nearly a quarter, or 23%, said human review takes place only after AI decisions have already been implemented. Another 13% said their organisation has no formal approach at all.

Taken together, that means 26% of organisations using AI are operating without what the research describes as meaningful human oversight. For companies deploying AI in live processes, that raises questions about internal controls, escalation routes and responsibility when a system causes harm.

Governance effect

The findings suggest governance frameworks are linked to stronger results from AI projects. More than three-quarters of Australian organisations, or 76%, said they have a formal AI governance framework in place.

Among those surveyed, 75% said governance had made their AI initiatives more successful. Just 5% said it had made them less successful.

Organisations with formal frameworks also reported better execution. Some 45% said their AI initiatives exceeded expectations, while 64% said governance made it easier to scale AI beyond the pilot stage.

The figures are notable because many companies have struggled to move AI work from experimentation into regular operations. The Australian data suggests that setting rules around oversight, data use and responsibility may help organisations expand projects more confidently.

Even with those frameworks in place, respondents said barriers to stronger returns remain. Data quality was cited as the biggest obstacle to improving return on investment from AI, with 22% of Australian leaders naming it as the main constraint.

That points to a practical problem facing many companies: systems may be easier to govern on paper than to run effectively when the data feeding them is inconsistent, incomplete or hard to verify. It also ties to broader concerns about provenance, storage and the permitted use of information inside AI systems.

Roman Kilun, Chief Compliance Officer at ABBYY, said the findings showed businesses were still wrestling with basic questions around oversight and value. "Businesses are embracing AI, but our research shows that many are still struggling with key questions around accountability, governance and ROI," Kilun said.

He added that internal communication and data controls remained central issues for management teams. "Senior management need to do a better job of communicating what their responsible AI policies are companywide. In addition, AI requires organizations to know where data comes from, whether it can be trusted, how it can be used, and where it is stored and processed. As data sovereignty grows in importance, a shift toward machine-readable controls that govern data throughout its lifecycle is required to ensure consistent compliance," Kilun said.

Kilun said many organisations were now trying to move AI work into day-to-day use. "Organisations we work with are moving beyond pilots and proofs of concept into execution, and clear frameworks and a shared understanding of responsibility are critical to building confidence in AI. At ABBYY, we're committed to being transparent about our own governance approach and supporting customers on that journey," he said.