Australian AI training gap becomes cost, productivity burden

Australian businesses are struggling to extract value from artificial intelligence after acquiring AI tools without adequate training or a clear understanding of how the technology should be used.

That lack of efficiency becomes a cost issue, as well as a strain on productivity.

Many organisations, particularly in the mid-tier and upper-mid-tier market, are not making deliberate, board-level decisions about AI investment. Instead, AI capabilities are increasingly arriving through existing software licensing agreements.

Microsoft's Copilot is one example, with organisations suddenly finding themselves with large numbers of licences but precious little understanding of how to deploy them effectively and at scale, said Advisor at IBRS, Joe Sweeney.

"Companies have these tools, and then they realise they don't know what to do with it, and that's a training issue," Sweeney said.

AI adoption is also often occurring at a departmental level, with individual managers deciding to deploy tools within their teams rather than organisations developing a coordinated strategy.

While tools can have wide-ranging benefits across different departments, such as sales, people management and marketing, in some organisations only teams whose managers are proactive are reaping the benefits.

This lack of unification across a business can lead to dissatisfaction among the employee base, potentially leading to increased use of shadow AI, which brings its own risk potential for governance, data leakage and decreased observability.

Citing research conducted with over 16,000 Australian workers, IBRS found that formal training was a much stronger indicator of successful technology adoption than characteristics such as age, occupation or gender.

And the most successful AI users are not necessarily the ones using the technology to automate as much work as possible.

Instead, research suggests effective users approach artificial intelligence with a 'scientific mindset', using the technology to challenge their own thinking, test assumptions and identify gaps.

"They're not using AI to do the job for them. They're using AI to question themselves, to prove themselves, to look for what they're missing," Sweeney said.

As a result, the value of AI can be difficult to capture via traditional productivity metrics. 

Organisations should instead consider improvements in the quality of services and deliverables, as well as the quality of decision-making and thinking.

He also warned that businesses need to invest in training around AI verification and validation, rather than focusing solely on traditional governance rules such as not entering sensitive customer information into AI systems.

Pointing to an example where an AI-generated recommendation could have resulted in the deletion of an entire CRM database if it had been followed, Sweeney stressed the importance of validating what is and is not permissible for tools.

Beyond generative AI tools accessed through prompts, deeper workplace process redesign has been identified as a more significant but often overlooked area of investment.

Businesses need to consider not simply the cost of an AI tool, but the cost of rethinking how an existing process should work when AI is introduced.

Building new tools from scratch must also be done with clear customer outcomes in mind, Sweeney cautioned. Companies should ensure adequate testing is done before making these tools available.

"Test, test, test. Always try to be wrong and then correct the mistake," he said.

AI costs set to become less onerous as token model shifts

Sweeney also predicted that concerns around the cost of generative AI processing will lessen as businesses take a more sophisticated approach to the use of AI models.

The emergence of AI systems designed to iteratively work out what a user wants has created significant processing costs, particularly when organisations allow models to repeatedly reason, search for information, generate software and check their own work.

A more efficient approach is to use orchestration, with humans determining the overall task and then structuring calls to different AI models for specific jobs.

This would significantly reduce the amount of processing required, while falling model costs and improving technology are likely to push AI costs down further.

"It's already dropping. We're just using it faster than it's plummeting," Sweeney said.

He predicts AI will become increasingly embedded into business processes and eventually cease being a distinct technology conversation.

"Generative AI of many different types and models will be so obvious in how we use it, and where it should be used and at what cost, that we just won't talk about it anymore," he said.

"It'll become invisible."