On Tuesday, Sam Altman told CommBank CEO Matt Comyn that he was struggling to explain why the latest OpenAI models should not simply be called AGI. An internal OpenAI reasoning model had recently disproved an 80-year-old geometry conjecture with a 150-page summary of its proof. And yet the one task Altman says AI cannot do for him is respond to his messages.
"Like many people, when I think about my email queue, there is nothing I'd less rather have to do than respond to those messages," he said. He tested it – delegating email, Slack and text messaging to ChatGPT. The experience surprised him. "I found it surprisingly dehumanising to watch, even when I had it reply to messages saying this is Sam's AI."
Although ChatGPT is completely capable of writing replies, the results felt "icky". "We really do care about our interactions with people," he said, "and this thing, which is a huge amount of my time, is not something that I can imagine myself outsourcing to an AI anytime soon."
Despite using AI extensively, even the CEO of OpenAI feels the gap constantly. "I still know I should be using so much more AI than I am, and I use a lot." It was an admission that everyone using AI knows well. Hearing it from Altman suggests it is structural, not personal.
The observation has revised his thinking on jobs. The fear that AI would eliminate entry-level white-collar work faster than societies could adjust has not played out as he expected. His inbox experiment is part of the reason he now thinks it won't.
Closer than comfortable
Comyn interviewed Altman, appearing via video link from the US on a public holiday, for close to an hour. The conversation ranged across some of the most pressing questions facing technology leaders: how to run a business when the planning cycle can no longer keep pace with the technology, how enterprises can absorb AI capability fast enough to stay competitive, and whether the productivity gains organisations are reporting are actually showing up in revenue.
Underlying all of it was a technology that, by Altman's own account, is now crossing thresholds he had not expected to see so soon. "The thing that I used to say that AI couldn't do yet was come up with important new knowledge," he said. "And now it can."
His prediction for the next shift follows from this. Rather than a system you prompt and wait on, Altman expects AI to become persistent – always running, aware of your goals, acting without being asked. "You'll just be talking to a system about your goals," he said, "providing editorial feedback, and it'll just constantly be working to try to help you."
For enterprise technology leaders the implication is less philosophical than operational. A system that acts continuously without prompting is a different governance problem to one that waits for instructions.
The planning problem nobody has solved
The question Altman says CEOs are asking him most often is not about which AI tools to buy or how to source compute. It is more fundamental.
"How can I run a company on an annual or quarterly planning cycle when the whole world is changing every month or every two months or less?"
He cited Satya Nadella at Microsoft and Mark Zuckerberg at Meta as examples of leaders visibly reinventing their operating models in real time. Nobody, he said, has discovered the answer yet.
Comyn put the speed problem in concrete terms. Many organisations have moved from annual to quarterly planning cycles, he noted – but the gap between quarterly and the cadence AI now demands is not incremental. It is exponential.
The security function illustrates what happens when organisations run out of time to deliberate. Altman described a conversation from roughly six months ago with chief security officers who said they would study agentic coding tools in 2026, design an implementation plan in 2027, and consider deployment in 2028.
Competitive pressure collapsed that timeline almost immediately. Engineers told their employers that if they could not use these tools, they would find employers who would let them. "One of the most rapid adoptions of a new technology at a serious enterprise level," Altman said, "that I've ever seen."
He expects the same pattern to repeat. "That's about to have to happen again, and then probably again, and again after that."
Superhuman persistence
The reason AI has become so capable at cybersecurity is not, Altman argued, primarily about intelligence. It is about persistence. "People get tired, people lose their train of thought, people get distracted," he said. "These models pointed at a problem will just keep going."
Finding a security exploit – even one requiring hundreds of chained steps – is a well-defined problem that suits AI precisely. It can run in parallel, hold vast amounts of context, and does not get demotivated. "The world as a whole," Altman said, "is sort of underprepared for what superhuman persistence applied to average human intelligence was capable of."
The use of AI in cybersecurity has led to a growing asymmetry. Models are better at finding vulnerabilities than patching them. But Altman said to expect tools from OpenAI and others within months that can do both – scanning codebases as code is written, before deployment, and running continuously against live systems. The complication is legacy infrastructure: software running in production stacks for two decades, with no surviving source code, will require more than an automated patch. "AI will be pretty good at this," he said. "But there's going to be some real surgery."
The hardest automation target
If the planning cycle problem and the security inflection illustrate how quickly AI is reshaping organisations, Altman's internal goal for OpenAI illustrates where the limits still are. Roughly a year ago, he set a target: by March 2028, OpenAI would have a fully automated AI researcher – a system capable of doing the core intellectual work of the company independently.
The goal is revealing because of what it is automating. Not a support function or a back-office process, but the role that generates OpenAI's primary IP.
Altman said the challenge is not only technical. It is psychological. He wants to make sure people inside the company are "not doing the psychologically convenient thing of pretending that this is in the distant future" – that the work of preparing for it, and examining its implications for safety and alignment, starts now.
The two-year gap also countered Altman's thoughts on AGI earlier in the conversation. Models may be able to disprove 80-year-old geometry conjectures but they cannot yet be trusted to autonomously direct the research agenda of the organisation that builds them. Capability and autonomy, Altman suggested, are not the same thing.
His broader scorecard on OpenAI's predictions is candid. "We have been roughly right on the technological predictions and pretty wrong on the social and economic implications," he said.
The clearest example is job displacement. Altman expected more entry-level white-collar roles to have been eliminated by now than has actually happened. He says he now better understands why it hasn't. And he is, by his own account, "delighted to be wrong" – though he is careful to add that the risk has not disappeared, only been deferred.
On the broader question of transparency, his position is consistent: say what you actually think, surface the fears you genuinely hold, and accept that you will sometimes be wrong publicly.
"I believe that so much of society here is going to be impacted by this, that we are all stakeholders," he said. Better to err toward too much transparency than too little, even if it occasionally means walking back a prediction.
The question nobody wants to ask
Comyn closed the conversation with what Altman described as the most important negative question circulating among enterprises right now. AI spend is rising. People report feeling more productive. They do not want the tools taken away. And yet, for many organisations, the revenue impact remains elusive.
"Where is the revenue?" Altman said. "There's a lot of great things I hear from companies. The negative one I hear is our spending is going up and up, people feel like they're being very productive, people report these amazing things, they don't want us to take the tools away, but where is the revenue?"
Altman hedged his response. It is still early. It will take longer than expected for organisations to actually run more efficiently and build better products using AI. But he drew a line. "If a year from now we're still talking about that same question, I'd be more concerned."
Comyn added a practical dimension that will resonate with finance and technology leaders in equal measure: organisations are only beginning to grapple with how to budget for tokens, how to allocate AI spend across functions, and how to connect that expenditure to business outcomes. The discipline of tokenomics – understanding the unit economics of AI usage at an organisational level – is still being invented.
The implication for boards and executive teams is direct. Reporting productivity sentiment is not the same as reporting productivity. The measurement frameworks that will make AI investment legible to shareholders do not yet exist in most organisations. Building them is the next organisational challenge, and the clock Altman is describing – roughly twelve months before the ROI question becomes genuinely worrying – has already started.
Australia's opening
Other sessions at the conference criticised Australia's pace of AI adoption by the business community. However, Altman gave a relatively upbeat assessment of Australia's position in the AI infrastructure race.
He named three structural advantages: abundant natural resources and clean energy capacity, stable institutions with predictable rule of law, and a strong national security posture with reliable partners. Taken together, Altman said, Australia has the foundations to become a data centre capital of the world – and would rank "in the very top few places" globally if it chose to pursue that.
Altman described demand for sufficiently high quality and low cost AI infrastructure as "effectively uncapped". For a country with the physical and institutional prerequisites already in place, the constraint is not capability, but ambition.
He outlined the strategic choices facing every country engaging seriously with AI. The options are not mutually exclusive, but they require deliberate positioning: become a data centre operator and exporter of AI infrastructure to other nations; secure the ability to import AI intelligence on favourable terms, the way countries have historically negotiated energy contracts; develop base models domestically; or focus national effort on making local enterprises the most AI-efficient in the world. Most countries, Altman suggested, will pursue some combination. The question is which combination, and how quickly the decision gets made.
Australia's research base also drew a mention. Altman flagged it not just in AI but across adjacent scientific fields, and connected it to what he described as one of the most consequential near-term applications of the technology: AI-assisted scientific discovery. A country with strong research institutions and the infrastructure to support large-scale AI compute is well placed to participate in that wave – not just as a recipient of discoveries made elsewhere, but as a contributor.
The same advantages Altman identified at a country level – stable governance, reliable infrastructure, strong institutional frameworks – are the conditions that make enterprise AI deployment tractable. Australia's regulated industries, banking among them, are not behind the curve. They are operating in an environment that, by Altman's account, is better positioned than most to get this right.
