AI is everywhere in the news right now. Today it went all the way to the top, with the Prime Minister using a speech at the University of Sydney to announce Australian Standards for AI, a new Office of AI, and a clear ambition: Australia designing and building this technology, not just adopting it.
Yet most people's daily contact with AI remains the most visible and easily understood systems: generative AI, the large language models and image generators. Frequently trained on copyright-infringing inputs and deployed as primary tools for text and creative output to mixed results, generative AI seems to fall short of the potential promised.
But generative AI is not the only show in town.
Industrial AI is a very different beast. It closes the loop from seeing and thinking to acting in an industrial environment, with physical AI and embodied AI working together to make it happen. And unlike much of the AI conversation, this one comes with measurable results.
What it Looks Like in Practice
A nine-month research program led by the University of Technology Sydney, Turning AI into Productivity went looking for industrial AI in the wild, visiting fourteen innovation ecosystems across Europe and the Nordics. What the team found was specific, task-level, and already paying off. The key for early adopters was simple: good data, clear objectives, and in-team champions focused on outcomes, not tools.
In Dortmund, an AI-assisted screwdriving system guides fastening sequence and quality assurance on assembly lines, cutting rework. At Siemens in Munich, AI helps generate work instructions and programming support for industrial robots, compressing tasks that once occupied specialist programmers for days. And in Eindhoven, ASML treats AI as an industrial method, embedded in lithography, precision manufacturing, and systems engineering across the Brainport region.
Note where the Siemens system pulls its knowledge from: the company's own engineering framework, approved documentation, machine manuals, and validated code libraries, with governance built in. Australia has a direct local parallel. AMCA, the peak body for the nation's HVAC and mechanical contractors, worked with ARM Hub to build AI tools that draw exclusively from AMCA-approved safety content, cutting the time to produce a compliant Safe Work Method Statement from four hours to under 15 minutes, with every answer traceable to an approved document.
The pattern is consistent: a general-purpose engine, the company's own knowledge as the keel, and skilled workers at the helm. This is what turns raw AI capability into usable, useful industrial AI.
So Why isn't it Everywhere?
In 1987, economist Robert Solow famously quipped that the computer age was visible everywhere except in the productivity statistics: the productivity paradox. AI is repeating the pattern: enormous global investment, racing capability, patchy measured gains.
The UTS research points to why. AI's returns depend on what it combines with: skills, data governance, management capability, and the institutions around a firm. Gains from general-purpose technologies arrive only after firms make these complementary investments, what Brynjolfsson, Rock and Syverson describe as the productivity J-curve.
The research team named the specific point where industrial AI stalls: the Management Chasm. Firms run successful pilots, then fail to cross into production, where AI has to be integrated with real workflows, workforces, and processes. The binding constraint is management capability, which as Bloom and Van Reenen's research shows, varies enormously between firms and correlates strongly with productivity.
The ecosystems that cross the chasm share some furniture: translation institutes like Germany's Fraunhofer network and Kaiserslautern's DFKI, whose engineers carry AI into ordinary firms through long-trusted relationships, and testbeds like SmartFactory KL, a working demonstration factory where manufacturers can trial AI in production conditions and fail cheaply.
What to Actually do
Closer to home, the ARM Hub AI Adopt Centre, the same outfit behind the AMCA tools, has spent more than 18 months working directly with 300-plus Australian SMEs, from Echuca to Gladstone. Its findings match the international evidence: AI creates real value when it reduces search time, rework, and dependency on key individuals. It fails when it generates summaries disconnected from action, lives outside existing workflows, or solves a problem no one on the floor owns. ARM Hub's recommendations give SMEs a concrete starting point:
- Get the data right first. The big ERP uplift can wait. It's the wrong opening move, and the capability it promises gets built along the way instead. Structured, well-organized, trustworthy data is the foundation everything else relies on.
- Complete 5-10 small, targeted automations over 12 months. Incremental wins compound into real capability. This is how firms cross the Management Chasm in practice: step by step, inside daily business, without a disruptive transformation program.
- Solve a real problem your people face today. Start where the pain is, with someone on the floor who owns it.
Why does this matter? Because scaling Australia's manufacturing capability will take every tool we have. And no one else is slowing down.
Researchers Warn AI Talking Toys May Misread Emotions And Need Stronger Safety Rules