The rapid diffusion of generative AI has revived fears that a new wave of automation will inevitably shift income away from workers. This column uses data covering nearly the universe of French firms between 1994 and 2019 to test the relationship between automation and the labour share of income at the firm level. While firms clearly automated extensively over this period, the evidence shows that it can profoundly reshape jobs and production without necessarily shifting income from labour to capital.
The rapid spread of artificial intelligence has revived an old concern: if machines take over a growing share of production, do workers receive a smaller share of income? This concern builds on a large literature arguing that automation has contributed to the decline in labour’s share of income across advanced economies (Karabarbounis and Neiman 2014, Acemoglu et al. 2020, Acemoglu and Restrepo 2022, Hubmer and Restrepo 2026).
The logic is straightforward. If machines replace workers, firms spend less on labour while producing the same – or more – output. Labour’s share of value added should therefore fall.
But does this actually happen inside firms?
Surprisingly, we have little direct evidence. Most studies relate automation to changes in labour shares across industries or the economy as a whole (Autor and Salomons 2018, vom Lehn 2018). Yet automation decisions are made by firms. If automation truly lowers labour’s share, we should first observe that relationship within firms, where automation actually occurs.
In recent work (Bárány et al. 2026), we provide a direct firm-level test of this mechanism.
A simple prediction
Economists largely agree on one consequence of automation: it disproportionately displaces workers performing routine tasks. Since the seminal work of Autor et al. (2003), routine occupations – occupations involving repetitive, codifiable tasks – have been viewed as the most susceptible to automation. Their decline has been documented across many countries and is widely regarded as one of the defining features of labour market polarisation (Autor and Dorn 2013, Goos et al. 2014, Bárány and Siegel 2018).
This insight provides a practical way to measure automation at the firm level. Rather than tracking every robot, software package, or AI application adopted by firms, we ask whether firms reduce their reliance on routine workers. A falling routine employment share provides a broad measure of automation because, regardless of the technology involved, automation tends to replace routine work first.
Our task-based model formalises this intuition and yields a simple prediction: among otherwise similar firms, those experiencing larger declines in routine employment should also experience larger declines in their labour share.
France offers an unusually rich testing ground
Testing this prediction requires two types of information that are rarely available together: detailed data on the occupational composition of each firm’s workforce, allowing us to measure routine employment; and firm-level data to measure labour’s share of value added.
French administrative data provide exactly this combination. We combine matched employer–employee records with firm accounting information covering nearly the universe of French firms between 1994 and 2019. During this period, the aggregate routine employment share fell by 12.5 percentage points, while the aggregate labour share declined by just 2.6 percentage points.
Figure 1 Routine employment share and labour share distribution across French firms
Source: Figure 1 in Bárány et al. (2026).
At the firm level, the data reveal widespread routinisation. The median firm’s routine employment share fell from roughly 60% to below 40% over the sample period, pointing to widespread automation across firms. Yet labour shares hardly moved (Figure 1). This contrast is striking.
But comparing cross-sectional distributions is not enough. To test the theory, we need to compare firms that started from similar positions – in the same industries, with similar labour shares and similar reliance on routine work – and ask whether those that reduced their routine employment share more also experienced larger declines in their labour share.
Figure 2 shows the results by firms’ initial routine employment share decile. If the standard theory were correct, every estimate would lie above zero. Instead, most are close to zero or negative, directly contradicting the theoretical prediction.
Figure 2 Routine share change and labour shares change
Source: Figure 2 in Bárány et al. (2026).
The empirical pattern differs sharply from the theoretical prediction. While the model predicts a positive relationship, we find the opposite. There is no evidence that firms experiencing larger declines in routine employment also experience larger declines in their labour share. If anything, the relationship is often negative and statistically significant.
The result is remarkably robust. It holds across broad sectors, including manufacturing – the sector most commonly associated with automation – and remains unchanged when we isolate plausibly exogenous variation in automation using an instrumental variables strategy.
This is the central finding of our study. While firms clearly automated extensively over this period, our evidence shows that this did not necessarily shift income away from labour. The benchmark view that automation mechanically reduces labour’s share within firms is therefore not supported by the data.
Why might labour’s share remain stable?
Our findings do not imply that automation has no effect on firms. Rather, they suggest that its effects are more complex than the benchmark task model assumes. Several mechanisms could keep a firm’s labour share stable even as routine employment falls.
One possibility is that while automation eliminates routine jobs, it also creates demand for other jobs, such as engineers, software specialists, technicians and managers who design, operate and maintain increasingly complex production systems. Recent research has also emphasised that automation may create entirely new tasks rather than simply replacing existing ones. Both mechanisms imply that the labour share needs not to fall even as routine employment declines.
Another possibility is that automation raises the value generated by completing non-automated tasks. This can either directly or indirectly – via an increase in bargaining power – increase the wages of the workers performing these tasks.
Our evidence cannot distinguish between these mechanisms. It does show, however, that any successful explanation of labour share dynamics must account for a simple empirical fact: routine employment often falls without labour’s share falling alongside it.
Looking beyond firms
While our empirical results speak to within firm changes in the labour share, they have implications for the role of automation in the decline in the aggregate labour share as well. As shown by previous research for the US (Autor et al. 2020), it is true for France as well that most of the aggregate decline in the labour share reflects a reallocation of economic activity towards lower labour share firms, rather than a general fall in the labour share within individual firms. It can be further shown that the dominant force behind the aggregate decline – as in the US – is a cross-dynamics effect: firms whose labour share declines are also those expanding their value-added share.
Our firm-level evidence adds an important piece to this puzzle. Because automation does not systematically lower labour’s share within firms, it also cannot be the force driving the cross-dynamics term. Taken together, our findings suggest that automation is not responsible for the decline in the labour share – neither within firms nor in the aggregate economy.
Implications for the AI debate
The rapid diffusion of generative AI has revived fears that a new wave of automation will inevitably shift income away from workers. Our findings suggest a more nuanced picture.
Our evidence comes from the previous wave of automation, covering French firms between 1994 and 2019. During this period, firms that automated more did not experience larger declines in their labour shares, despite substantial reductions in routine employment.
This does not imply that AI will have no distributional consequences. AI affects a broader range of tasks than earlier automation technologies and may reshape markets, organisations, and bargaining power in new ways. Whether it will ultimately reduce labour’s share of income remains an empirical question.
The broader lesson is that automation should not automatically be equated with a declining labour share. It can profoundly reshape jobs and production without necessarily shifting income from labour to capital.
Source : VOXeu
