DELOITTE AUSTRALIA
Rob points to research from METR that tracks how well Gen AI handles extended tasks on its own.
The latest frontier models can tackle work that would take a human over two and a half hours, but with a success rate of about 50 %. For 80 % reliability, this time drops to under 30 minutes.
“ If we think about the quality and consistency we expect of people in our organisations – would we accept having to check in on them every two and a half hours and they still get it wrong half the time? That is where today’ s most advanced general-purpose AI models are starting from, which is why the right checks and balances are so important,” he says.
What makes this interesting is that capability doubles every seven months. For Rob and his teams, this creates constant tension between what’ s technically possible and what makes business sense right now. Yet the evolution of Gen AI tools is constantly changing that equation, letting teams achieve more with the same resources.
Behind the shift from coding to managing AI Few professions face bigger disruption from AI than software engineering.
Rob quotes Kent Beck, inventor of the JUnit testing framework:“ The value of 90 % of my skills dropped to zero – and the leverage for the remaining 10 % went up 1,000 times,” he says.
That shift runs deeper than individual productivity. At Deloitte, technical skills are spreading across the organisation.
“ The thing with the productive potential of advanced agentic software engineering tools is that even seeing is not quite believing”
Robert Valk, Engineering CTO, Deloitte Australia
Teams from financial advisory and banking recently worked on blockchain payment technologies – domain experts doing deeply technical work.
“ The thing with the productive potential of advanced agentic software engineering tools is that even seeing is not quite believing,” Rob says.
“ You have to actually kick the tyres; you have to use these tools and deliver real outcomes to truly appreciate the productivity gains that are on the table.”
But AI can’ t work miracles on its own.“ AI amplifies what you know and what you can do, but if you don’ t know anything and you can’ t do anything, then there’ s nothing to amplify,” he says.
This changes which skills matter most. Management capabilities – clear communication, effective delegation, structured work – are becoming increasingly critical as engineers evolve from writing code to managing AI agents that write code.
6 January 2026