Robotics as the bridge between AI code and industrial results

In his article ‘The AI productivity take-off is finally visible’ in the Financial Times on 15 February, Stanford economist Erik Brynjolfsson sees the long-discussed AI productivity surge at a turning point: His analysis suggests that US productivity rose by around 2.7 per cent in 2025, which is ‘almost twice’ the decade average.

However, this macro interpretation for industry will only be truly reliable once it is reflected in the hard operational figures for automation – and this is precisely where robotics becomes the decisive transmission stage between AI in code and AI in results.

Brynjolfsson describes this development as the typical J-curve of general-purpose technologies – a pattern in which a change or investment initially produces measurably poorer results and only subsequently leads to above-average improvements. In short, it is as simple as it is indisputable: investments and conversion costs come first, output gains later.

From the investment phase to the harvest phase

In this sense, he describes companies that ‘move from this investment phase to a harvest phase’ and where the earlier conversion work becomes visible ‘as measurable output’. In this context, it is crucial for robotics to determine whether AI actually addresses the classic touchstones of automation. The key indicators for the reliability and availability of AI robots are primarily the integration effort, the variance in the range of parts, the commissioning time, and the robustness and uptime/MTBF in 24/7 operation.

Consistently working through the J-curve to industrial maturity

The current humanoid discourse provides a rare and clear reality check. Michael Tam, Chief Brand Officer of the Chinese robotics company UBTech, says in the Financial Times article ‘Robots only half as efficient as humans’ from 25 January about the humanoid robot ‘Walker S2’ that it only achieves 30 to 50 per cent of human productivity and that ‘only for specific tasks such as stacking boxes and quality control’. Tam also explains the current competitive pressure facing companies: ‘If Tesla has the advantage of bringing its own humanoid robots into the production line, then BYD and others may be left behind.’ Translated into productivity logic, this probably means that humanoids are not primarily being purchased today because they are already economically superior, but because manufacturers want to secure learning curve options and first-mover advantages. This fits with Brynjolfsson's thesis: macro productivity is not generated by spectacular demos, but by consistently working through the J-curve to industrial maturity.

AI productivity in robotics is a scaling race

At the same time, classic industrial robotics shows how quickly scaling becomes a location factor when technology and ecosystem are ‘ready’. The International Federation of Robotics (IFR) quotes its president, Takayuki Ito, in the context of the latest World Robotics Report: ‘The new World Robotics statistics show that 2024 will see the second-highest annual installation figure,’ and estimates the global operational stock at ‘4,664,000 units in 2024.’ Unsurprisingly, according to the IFR, China is decisive for competitive dynamics, accounting for 54 per cent of global deployments in 2024. For European industry, this is a direct signal: AI productivity in robotics is not just a tool upgrade, but a scaling race in which data, integration standards, safety certification and service networks determine cost curves.

Adaptation costs in robotics programmes are particularly high

Current findings on the ‘industrial AI J-curve’ illustrate very clearly why this often appears with a delay in productivity data. Last summer, the Massachusetts Institute of Technology (MIT) School of Economics and Management summarised the results from the US Census Survey data as follows:

‘AI adoption tends to slow productivity in the short term,’ and even after controls, there is ‘a 1.33 percentage point decline in productivity.’ And Kristina McElheran, Professor of Strategic Management at the University of Toronto, emphasised in this context: ‘This initial decline – the downward side of the J-curve – is very real.’ These adjustment costs are particularly high in robotics programmes because every AI function has to be translated into hardware realities: sensor drift, gripper tolerances, process windows, cycle synchronisation, safety distances, hygiene, spare parts concepts and dealing with ‘long-tail’ variance. If these complementary investments are lacking, AI is more likely to create new bottlenecks than deliver higher overall equipment effectiveness.

Where will the AI productivity leap first become visible in the context of robotics?

Where can the productivity leap claimed by Brynjolfsson nevertheless first become visible in the context of robotics? Most likely where AI reduces non-value-added time around robotics – through faster layout and cycle time design via simulation, more automated parameterisation of vision/inspection, more robust anomaly detection for predictive maintenance, and fewer ‘engineering hours per cell’ through reusable software stacks. Economists also see signs of this effect already emerging in some areas from a macro perspective. For example, in a commentary at the beginning of February, Stephen Brown of Capital Economics concludes from output/employment decoupling in the ICT environment: ‘All of this suggests that AI is making a major contribution to productivity growth.’ For robotics, this could mean that the leap in productivity will not come as a sudden humanoid revolution, but as a cumulative effect of better perception, more reliable autonomy in narrowly defined tasks and drastically reduced integration and operating costs – in other words, precisely those levers that transform pilot cells into a scalable factory architecture.