AI, why chat interfaces are bad, and Australia's advantage

Sci-fi already taught us that you can't just jump straight to lightspeed

Except that one time Rian Johnson did it, you don’t get to lightspeed without first engaging the sublight engines.

Sublight engines let starships break through the atmosphere and into orbit. Star Wars always focused the narrative on the final jump to lightspeed, where the stars stretch into white lines as the ship rips free of realspace. But it was never possible without sublight propulsion to get clear of the planet’s gravity field.

That’s where I see Australia with AI.

The hype machine is in overdrive.

Every second post on LinkedIn is about it. News headlines every day. I try to check Hacker News and Reddit every morning just to keep up. It’s my job.

For context, I now work as a Chief Engineer at the Commonwealth Bank of Australia. I am focused on Generative AI, and specifically how we understand, systemise and control this technology so that it can be adopted safely at enterprise scale.

I came to CommBank after five years studying the future of work in my role at Faethm AI. There we deployed language models at scale to extract insights from global job markets. We informed transformation agendas for our clients around the world with tech adoption trends across occupations and industries. Along the way our company joined Pearson Education, and we zoomed in on analysing workforce skills, and enhancing learning strategies within enterprise and at national scale.

All this time, the pattern was clear: we were early adopters of a new fast-moving technology. And we watched the numbers as broad adoption crawled.

Rewind on machine learning

I’ve lived the research first-person through my own career, which started in high-volume data analysis in a management consulting setting. Analysing operational data to find problems, prioritise fixes, and shape strategy.

Enter machine learning. Predictive AI.

As I continued to upskill in my field, data science practices evolved rapidly. We turned that data into predictive capabilities to move for diagnosing the present to anticipating what comes next. It was a lot of work with all the feature engineering, model selection and testing, but with enough data and the right formulation, you could produce a model for use in a novel operational scenario to assist decision-making and facilitate levels of personalisation that were not previously viable.

What struck me with each wave was despite all the creative possibilities real adoption always lagged these capabilities. And in hindsight this should not have been a surprise. Success required the right combination of skills, strategy and support. A team of specialists and domain experts focused deep in a problem space. And the backing of a leader to follow-through with a project from its earliest stages up until it became the way we work, with the trust of customers and regulators alike.

At Faethm around 2020, surrounded by an extremely innovative team of data scientists and AI engineers, I witnessed two early cues hinting at where the world was going:

1. Complex feature engineering was being replaced by embedding text with BERT (Google’s then state-of-the-art language model) and into a classifier

2. Model development was being sidelined in favour of few-shot prompting with GPT-2 (the last OpenAI pretrained model you could download to your computer until gpt-oss 6 years later)

The science of data science was moving steadily from “can we build this model” to “can we prove our approach is reliable for the task at hand”.

Language models, meet chat

In November 2022, gpt-3.5-turbo introduced chat-optimised language models to the world and ChatGPT launched to become the fastest-growing consumer software application in history.

The world went to data science. And the data scientists went to work.

Of course the extremely accessible nature of chat interfaces made it easy to test out new use cases. Summarisation, classification, translation and content generation were the obvious starting points. Entrepreneurs raced to wrap these instant capabilities into new products. Software engineers were experimenting with new side projects. Even me.

The world I saw a hint of in our work at Faethm was here.

A new industry born overnight. Cloud providers scrambled to expand capacity. Teams everywhere built chat interfaces because they were generic and fast to prototype. A few made it into products.

Most didn’t.

In my opinion, chat interfaces plateaued adoption because chat interfaces aren’t where the work actually happens. You can’t run a hospital out of a chat window. Or manage a retail store. Or a supply chain. For most jobs, chat was the wrong abstraction. It unlocked capability, but real productivity needs more.

Productivity in Australia

In Australia, we measure labour productivity as the amount of output produced per hour worked. Output for an industry or sector is usually measured the total value of goods and services produced less those goods and services used in the process. Dry, I know. But it’s the single metric that underpins our entire nation. Wages, living standards, the federal budget.

Labour productivity growth in Australia has been falling since 2015.

I don’t think Australians are lazy, people are working hard. The culprit is that too much of our time is lost to low-output tasks.

• Teachers, including my mother, will tell you about late-night assignment marking.

• Sydney traffic chaos impacts not just logistics drivers but every trade, every specialist moving between job sites.

• Every hour wasted retyping data between systems that don’t talk to each other.

The intuitive temptation is to automate the work. Go full lightspeed.

But at a national scale lightspeed is destabilising. Labour participation drops, consumer sentiment shrinks.

The real opportunity is augmentation. Amplifying our people.

The software powering AI-powered software engineering

The power of Generative and Agentic AI to augment Australia is clear to me. I see it in my own work. As a software engineer, I now debug, prototype, and integrate systems faster than ever.

Until Claude 4 my preferred workflow was mostly moving manually between a chat interface and my development environment, copy-pasting code back and forth.

50% of developers agree this has had positive effect on their productivity.

While Agentic AI promised to let large language models modify the code directly, the previous generation of models weren’t ready for it. The user interfaces continued to evolve and now we have tools like Claude Code, moved AI out of chat windows and directly into developer workflows: the terminal, the IDE, the CI/CD pipeline.

A world of composable add-ons has evolved to integrate new capabilities, providing secure and standardised access to common systems used by engineers in their day-to-day. Like GitHub - the place where code lives. And Jira - where organisations prioritise feature requests and bug reports.

To me, this is a clue about the specialised tools that are needed in all occupations. And here’s the kicker:

AI-powered engineering augments the people who are going to build those tools for everyone.

National AI-Powered practices

Did you know that there are 175,000 software engineers in Australia? I checked the data. As an occupation we represent over 338 million hours of work every year.

But with almost 15 million working Australians we represent only 1.4% of the workforce.

There are eight occupations ahead of software engineering in Australia:

Here’s my view:

AI won’t lift national productivity unless it reaches their workflows. Because productivity isn’t measured in demos and prototypes, it’s output per hour worked.

Just like software engineers are experiencing today, we need the software for AI-powered practices for every occupation. And we know it’s not generic chat interfaces. It needs to live in the workflow itself.

Why Australia could move fast

To some, Australia is a small, distant side-market.

But Australia’s economy is concentrated in a handful of occupation groups: healthcare, education, retail, logistics, trades, clerical work. And that concentration is an advantage when it comes to AI augmentation.

Targeted AI adoption in real work can move national productivity faster than in larger, more fragmented economies.

In fact, 50% of our country’s work hours are done by just 50 occupations, with that second 50% by another 417 occupations.

Another interesting aspect is our concentration of employers:

• Almost 10% of our workforce employed by 63 large organisations (>10k employees).

• A further 15% employed by 868 mid-large organisations (>1k employees).

Half our work hours come from 50 occupations. A quarter our workforce employed by under 1,000 companies. We can move fast. The US needs millions of firms to align before productivity shifts. We don’t. That is our Australian AI advantage.

Now imagine a future where every Australian is equipped with the intelligent, occupation-specific software they need to get the most out of every hour. Right in their existing workflows. Just like we’re seeing with AI-powered engineering as it unfolds today.

AI-powered engineering can be a platform for national productivity

This is Australia’s opportunity.

Remember those 175,000 software engineers that represent 1.4% of our workforce? Here’s the play:

• Augment them with AI-powered engineering tools.

• Incentivise collaboration with domain specialists, regulators and customers across each of our major national occupations.

• Together they build intelligent, occupation-specific systems and guardrails directly into real workflows.

Do that and we get a multiplier effect:

1. Faster software development cycles

2. Better AI-powered tools for teachers, nurses, accountants and everyone

3. Productivity gains that lift the entire economy

Australia’s choice isn’t whether to embrace AI.

It’s whether we stay stuck in the gravity field of demos and prototypes, or engage the sublight engines and lift the whole economy.

I say engage the engines.