New Intelligent Technologies: Are They Making the Workplace Productive?

Abstract

This paper investigates whether intelligent workplace technologies improve firm-level productivity and, if so, under what conditions, with particular attention to their implications for the economic and social sustainability of firms. This investigation occurs in a context where firms increasingly combine automation, artificial intelligence (AI), and work-from-home (WFH) practices to sustain performance under structural shocks such as the COVID-19 pandemic. Despite evidence that firms adopt these technologies jointly and reorganize work accordingly, existing research typically examines them in isolation. We develop a micro-founded, task-based production model in which firms allocate tasks between on-site and remote labor and automated capital in an optimal manner. This model allows both automation technologies and remote work collaboration tools to affect productivity and coordination costs that are central to long-term organizational sustainability. Using firm-level survey data from nearly 4000 large firms across industries and countries (2018–2021), we show that working from home (WFH) exhibits diminishing productivity returns when scaled in isolation, reflecting rising coordination frictions. In contrast, firms that combine WFH with automation and digital collaboration tools experience significantly higher labor productivity growth. These integrated technology systems support sustainable productivity by enabling capital deepening, resilient task reallocation, and more efficient use of labor resources over time. Overall, the findings suggest that productivity gains—and by extension sustainable firm performance—stem from integrated workplace technology systems rather than isolated investments, highlighting the importance of coherent technology strategies for organizing work in the post-pandemic economy.

1. Introduction

Over the past decade, companies have invested heavily in new workplace technologies that have fundamentally reshaped how work is organized and executed. Two technology families dominate these investments. The first includes artificial intelligence (AI), such as robotic and agentic automation, machine learning, and natural language processing. These technologies allow firms to redesign and automate task workflows that would otherwise be performed by labor. The second family includes digital collaboration and remote work technologies, such as video conferencing and shared digital workspaces, that enable employees to work effectively outside of traditional office settings.

These technologies bring new forms of intelligence into the workplace [1]. They also offer a contrasting picture of the future of work: remote technology tending to extend labor work in contrast to AI substituting work [2]. Furthermore, while these technologies expanded rapidly during the pandemic—especially when working from home (WFH) became the only scalable way to maintain operations—firms adopted sharply different strategies as the pandemic subsided. Some organizations made remote work and digital collaboration permanent, while others mandated a return to the office and reduced investment in remote work technologies [3]. At the same time, investment in AI and automation, as well as in remote technologies, continued to grow even as spending on traditional software tools plateaued or declined [4].

For executives, this divergence reflects deeper uncertainties about how to integrate these technologies into the future of work and whether intelligent workplace technologies actually increase productivity—and, if so, under what conditions. Evidence remains mixed. Some firms report significant gains from remote work and automation, while others experience coordination failures, declining performance, and disappointing returns. This lack of clarity makes workplace technology strategy one of the most consequential—and controversial—managerial decisions in the post-pandemic economy. Understanding how these technologies jointly shape productivity is therefore not only a question of performance, but also of how firms can build economically and socially sustainable work systems in the post-pandemic economy.

Research on remote work shows highly uneven effects on productivity [5]. When supported by strong digital infrastructure and effective remote tools, WFH can raise individual productivity [6,7]. However, other studies document that technology support may not be sufficient. Organizational failures—such as coordination breakdowns, weaker supervision, and adverse selection into remote work—may limit effective work practices, particularly for teamwork-intensive activities [8,9,10,11]. As a result, estimated WFH effects range from strongly positive to strongly negative [12]. Similarly, research on automation and AI shows that these technologies can raise productivity, but not uniformly. Modern automation typically affects performance by reallocating tasks between humans and machines rather than eliminating jobs outright [2]. Firm-level studies confirm productivity gains from automation and AI, but emphasize that these gains depend heavily on complementary investments in workflows, skills, and organizational design [13,14,15,16].

Interestingly, a common limitation of both literatures is that the impact of technology investment and productivity has often been examined in isolation. WFH is treated as an arrangement, automation as a technology, and AI as an investment in innovation. In practice, however, firms adopt these technologies together, and their productivity effects depend on how they are combined. Therefore, this article argues that intelligent workplace technologies should be understood as a system rather than as independent tools. Remote work, automation, and digital collaboration jointly determine how tasks are allocated among on-site employees, remote employees, and machines. Thus, productivity outcomes depend less on whether a firm adopts a single technology and more on how technologies are combined to reorganize work. In fact, our sample demonstrates that firms rarely invest in only one type of technology; for example, they invest in both remote work and automation technologies. This finding underscores the importance of jointly modeling both types of technology, mitigating the risk of omitting correlated technology investments and reducing the typical bias in estimating the returns to technological innovation.

To make strong predictions, we develop a task-based framework grounded in the economics of task allocation [2] and test its implications using firm-level data from nearly 4000 large firms across industries and countries. Although technology adoption is often analyzed with conceptual frameworks such as UTAUT (Unified Theory of Acceptance and Use of Technology) and TOE (Technology–Organization–Environment) [17], we argue that these frameworks lack the micro-foundations necessary to specify the conditions under which firms and workers would favor or resist the diffusion of these technologies (refs. [2,18]). We resort to analyses tasks as either performed by workers (on-site or remotely) or by physical capital, while, additionally, both remote work and capital can be enhanced by intelligent technologies such as online collaboration tools and software automation, if firms have invested in those sets of technologies. In such an enriched framework, tasks are allocated based on preferences and the relative performance (price/productivity) of available technologies, with automation and remote work technologies typically acting as substitutes at the task level (refs. [19,20]). In turn, task-level arbitrage influences aggregate labor demand at the firm level, shaping workers’ preferences and wages. This creates an additional feedback loop for the relative attractiveness of technological choices, resulting in more robust predictions of technology adoption and use than those derived solely from the intrinsic usability of technologies, as in the UTAUT or TOE frameworks. In other words, the task framework focuses on optimal firm decisions that directly impact productivity changes.

The joint use of theoretical foundations and empirical analysis provides three clear insights. Firstly, we demonstrate that the productivity effects of working from home (WFH) cannot be interpreted independently of firms’ broader workplace technology portfolios. Unlike the prevalent reduced-form literature, which estimates an average WFH effect, our task-based framework demonstrates that observed WFH coefficients incorporate endogenous task reallocation between labor and capital. Consequently, standard estimates of WFH productivity effects are biased downward when complementary automation and remote support technologies are omitted.

Secondly, we reconcile seemingly conflicting findings in the literature by distinguishing task-level substitution from firm-level complementarity. Although automation can replace remote labor in routine tasks, we demonstrate that firms that adopt integrated bundles of work-from-home (WFH) policies, automation, and digital coordination tools experience higher overall productivity growth through capital deepening and reduced coordination frictions. This distinction explains why crowding out and productivity gains can coexist empirically. Thirdly, we provide quantitative evidence on the range over which WFH affects productivity. Rather than asking whether WFH “works” on average, we identify a non-linear productivity profile: moderate levels of WFH increase total factor productivity (TFP) and labor productivity. However, excessive remote work intensity generates diminishing or negative returns unless supported by substantial workplace technology investment.

The rest of the paper is organized as follows. Firstly, we provide background on the impact of technological innovation on firm adoption and productivity (Section 2). Then, we present a microeconomic task production model that includes the opportunity to use artificial and remote intelligence technologies. We then test the theoretical prediction that labor productivity is influenced by the optimal mix of remote and artificial intelligence technologies, in addition to traditional on-premises work practices. Section 5 discusses the findings and offers conclusions.

2. Workplace Technology and Productivity

2.1. Remote Work: Work from Home (WFH) as an Operating Model

From a business perspective, modern remote work, and essentially work from home (WFH), is not simply a location choice; it is an operating model enabled by digital collaboration technologies. Without reliable communication platforms, shared digital workflows, and real-time coordination tools, remote work quickly becomes inefficient. In effect, remote work initially emerged as a response to commuting constraints, notably during the oil crises of the 1970s, but early forms were severely constrained by limited technological support. The advent of widespread internet access marked a decisive shift by enabling work from nearly any location [21,22]. Garrote et al. [23] and Bai et al. [24] emphasized the critical role of reliable connectivity, while subsequent case studies [6,7] demonstrated that new digital tools, such as video conference tools and meeting recordings, can enhance remote work productivity.

As an operating model, WFH is not necessarily usable for all types of tasks: Dingel and Neiman [25] and Boeri et al. [26] show that only about 40% of tasks are truly tele workable, with many jobs requiring physical presence or direct supervision. Effectiveness also depends on workers’ technological skills and organizational readiness (refs. [5,6,7,21,22,23,24,25,26,27]). Not surprisingly, WFH studies report highly conflicting productivity effects, ranging from substantial productivity gains (+10% to +30% for routine or individually focused tasks; refs. [6,7]) to near-zero effects when coordination frictions offset those gains [8], and even significant declines (–10% to –40%) for teamwork-intensive, creative, or mentoring-related activities (refs. [9,10]).

Importantly, WFH studies have maintained a mono-technology focus, treating WFH as a binary or intensity variable while paying limited attention to the technological infrastructure that underpins remote work. Consequently, productivity differences often reflect unequal access to enabling technologies, such as stable broadband, video conferencing platforms, collaboration suites, and digital workflow tools, rather than remote work itself. Poor connectivity can mimic WFH productivity losses, while mature digital ecosystems can mitigate coordination frictions. Since these technologies are rarely measured or controlled for, the literature often conflates the pure effects of WFH with variation in the quality of remote work technologies.

2.2. Automation, AI, and Workflow Redesign

The history of automation has long sparked controversy. In the early 19th century, English textile workers—later known as Luddites—destroyed machinery they believed threatened their livelihoods. Later on, automation in manufacturing gained prominence in manufacturing with Henry Ford’s assembly line, followed by the introduction of industrial robots in the 1950s—especially in Asian automotive firms [28]. By the 1960s, information technology entered production environments. Antz et al. [13], then ref. [2] demonstrated that industrial robots significantly increased labor productivity while exerting only modest effects on total employment. With the rise in digital technologies and AI, automation has expanded into services such as healthcare, finance, and retail. Robotic Process Automation (RPA), often combined with AI, can execute routine tasks faster and more accurately than humans [29]. Studies report improvements in task accuracy of 10–15% and time reductions of around 20% [30]. RPA is increasingly considered a general-purpose technology though its benefits depend on process standardization and organizational readiness. Early estimates suggested that 50–70% of tasks could be automated, with this share increasing as AI capabilities improve—especially with generative and agentic AI [31].

As much as remote technologies favor new labor practices, automation and AI influence productivity primarily by redesigning workflows. RPA and AI systems automate routine tasks, free human time for higher-value activities, and standardize processes. However, their benefits materialize only when workflows are redesigned and supported by appropriate digital infrastructure (refs. [15,16]). Crucially, automation interacts with remote work. Automated workflows can reduce coordination burdens in remote settings, while poorly integrated automation can exacerbate fragmentation and rework. New recent research highlights that workplace technologies may substitute for each other at the task level while complementing each other at the organizational level [1]. This insight motivates our integrated approach. Productivity depends on how firms combine work-from-home (WFH) policies, automation, and digital tools, rather than on adoption decisions taken in isolation.

As previously mentioned, most existing research relies on reduced-form analyses and focuses on individual technologies, despite the fact that firms usually adopt groups of technologies whose interactions influence performance. The pandemic is a clear example of this. During the pandemic, firms simultaneously expanded their investments in remote work technologies (labor-supporting) and automation tools (capital-supporting). Motivated by these gaps, we propose a structural, micro-founded framework for analyzing workplace technologies. This framework clarifies the mechanisms governing their use and impact and advances beyond fragmented scholarly works.

3. A Model of Optimal Workplace Technology Investments

3.1. Framework Selection

We envision a scenario in which firms have three options for executing tasks: on-site labor, remote labor, and automated capital. Remote work reduces the disutility of workers but increases coordination costs. Automation increases capital productivity but requires scale and process standardization. Lower labor demand adjusts wages but makes labor more competitive with technology. Digital collaboration tools and AI shift the relative efficiency of these options. Thus, the model needs to formalize how firms can best rebalance tasks across these three modes and how productivity depends on combining them rather than adopting them in isolation. Our approach selects a microeconomic, task-based production perspective that extends seminal contributions [2,32,33]. The economic intuition here is that the firm is not choosing “WFH” or “AI” as discrete policies. It is choosing how each task is executed. Workplace technologies matter because they change the relative cost and effectiveness of labor versus capital at the task level.

These models explicitly recognize that technology operates at the task level first. Our framework adds that firms’ executives optimize task allocation between labor and physical capital based on the substitutability and complementarity of new intelligent technologies associated with labor and physical capital. Additionally, workers have preferences regarding work and location, which implies that managers selecting labor tasks must also consider the workforce’s reaction, as well as the supply of work and wages.

Intuitively, technology adoption, wages, and task allocation are jointly determined. This distinguishes the framework from acceptance-based models (e.g., UTAUT, TOE), in which technology use is driven by attitudes rather than equilibrium incentives.

Finally, the framework is simplified so that one can focus on the core dynamics of technology choice. Firstly, we consider a monopoly firm, abstracting from competitive interaction between rival firms, which does not affect technology decisions. On the other hand, workers select where and how much to work. Wages are determined when the demand and supply of labor are balanced. Firms perform many different tasks. Some tasks are easy to automate, some can be performed remotely, and others require an on-site presence. Workplace technologies affect the cost and effectiveness of performing these tasks.

3.2. Working Assumptions

3.2.1. Sales

We consider a monopolist firm selling all of its produced output $Y$, at price $P$, facing a downward-sloping demand curve with constant price elasticity, k > 0:

$$P=Y^{-\kappa }$$

(1)

This means that as output expands, prices fall in a constant proportion; this assumption is standard in modern productivity models. Likewise, market power matters because productivity gains do not translate one-for-one into revenue per worker. The elasticity κ scales how strongly technical efficiency shows up in observed productivity.

3.2.2. Production

Following task-based models of production [2], output is produced through a continuum of tasks that can be allocated to on-site labor, remote labor, or capital. Tasks differ in their suitability for automation and for remote execution. Remote work, while valued by workers, may generate coordination and communication friction such as delays, miscommunication, reduced informal knowledge spillovers, or weaker supervision—especially in team-based or creative activities. These frictions reduce effective labor productivity and are therefore modeled as part of the production technology rather than as worker disutility.

In our framework, coordination frictions manifest as diminishing returns on the intensity of remote work. At low to moderate levels, remote work may increase productivity; however, at higher levels, coordination costs offset these gains. Crucially, workplace technologies mitigate these frictions. Investments in digital collaboration tools, workflow standardization, monitoring systems, and automation reduce the effective coordination costs of remote work by improving information flow, task modularization, and supervision.

Formal analysis shows that labor-augmenting remote technologies and capital-augmenting automation technologies increase effective productivity and influence the task cutoff between labor and capital. Consequently, automation can replace certain remote tasks, while remaining complementary to remote work at the firm level through capital deepening and organizational restructuring.

It is important to distinguish between task-level substitution and firm-level complementarity. At the task level, improvements in automation efficiency shift tasks toward capital and may reduce the optimal share of remote labor. At the firm level, however, complementarities arise because workplace technologies jointly reduce coordination costs, improve task allocation, and raise total factor productivity. This distinction reconciles the presence of substitution effects in specific tasks with the empirical finding that firms adopting integrated bundles of workplace technologies experience higher overall productivity growth.

Finally, the concave relationship between remote work intensity and productivity reflects the interaction between worker preferences and production-side coordination costs. In routine, codifiable tasks, automation may dominate and crowd out remote labor. In non-routine, high-skill tasks, remote technologies primarily reorganize work rather than replace it. Our analysis hence captures this heterogeneity through non-linear effects and interaction terms, rather than through explicit task-level data.

Expressed in mathematical terms, the above is formalized as an output made up by a distribution of tasks that are executed by two types of inputs: physical capital $K$ (e.g., the workplace, plant, or office) and labor $L$. Both inputs are, respectively, paid the competitive rental rate $r$ and wage $w$. While the capital supply is fixed [2], we assume that workers derive utility from consumption and leisure and may value remote work due to reduced commuter time, autonomy, and flexibility. Formally, the labor supply is derived from worker utility:

$$U={\left[(\tau wL{)}^{-\rho }+\gamma (1-L{)}^{-\rho }\right]}^{-1/\rho }$$

(2)

where 0 < γ < 1 parametrizes worker preference for leisure [34,35,36], and $\rho$ governs curvature.

Labor is split between on-site work $L_p$ and remote work $L_h$. Define the WFH share as $\epsilon =L_h/L$. Remote work affects the disutility of labor through

$$\tau ={\tau }^{\prime }{\displaystyle \frac{\epsilon }{1-\epsilon }},{\tau }^{\prime }>0$$

(3)

And WFH changes how costly work feels to employees. As remote work expands, effective labor disutility falls, reflecting lower commuting costs and greater flexibility for the workers. For tractability, we also set $\rho =-1$ (quasi-linear utility), which directly implies that labor supply responds positively to wages and the effectiveness of remote work arrangements, holding production efficiency constant. In fact, the optimum labor supply condition given by dU/dL = 0 from (2) implies the following:

$$w={\displaystyle \frac{\gamma (1-\epsilon )}{{\tau }^{\prime }\epsilon }}.$$

(4)

And higher remote intensity $\epsilon$ reduces the equilibrium wage by lowering the marginal disutility of work (thus, ${\tau }^{\prime }$ = 1 − ε (respectively, increases to 1 − ε/ε) implies that workers only care about home work (respectively, care about total labor). Thus, a decrease in τ measures a bias in favor of remote work). Said differently, as remote intensity rises, workers accept lower wages for the same utility. This is a general equilibrium cost effect, not a behavioral assumption, and it directly affects task allocation between labor and capital.

Each task, x, is distributed in [0, 1] and is ranked from a low to high comparative advantage to capital versus labor as in [2]. Workers are thus assumed homogeneous within task classes, but in aggregate, automation primarily replaces low-skill, routine tasks, while remote work technologies reorganize high-skill, non-routine tasks.

The respective marginal productivities are given by (with $b>$ $a>0$ in order to ensure a meaningful equilibrium where both types of inputs are used by the firm) the following:

$$Y_K\left(x\right)=b/(1+x),Y_L\left(x\right)=a.$$

(5)

with $b>$ $a>0$, and coordination costs are reflected in the diminishing returns to WFH.

In practice, firms may invest upfront in intelligent technologies, C, D, to support each type of input, and one notes, respectively, the following:

$$a=c\left(1-\epsilon \right)+\epsilon \left(c.z{C}^m\right)=c(1+\epsilon \left(z.u-1\right))$$

(6)

$$b=f\cdot D^d=f\cdot v$$

(7)

where z/f are the premium of productivity attached, respectively, to remote technologies and to AI automation for the task x, while m, d are valued in (0,1) and capture technology complement effectiveness and returns to scale. For ease of notation, $C^m=$ u, $D^d=$v. Thus, WFH does not automatically raise labor productivity. It does so only when supported by effective digital tools ($z$). Similarly, automation productivity depends on AI/RPA maturity ($f$), not on adoption alone.

3.3. Equilibrium

3.3.1. Mix Capital/Labor

At the task level, automation and remote work act as substitutes, even if at the firm level, both are complementary through general equilibrium effects on task reallocation, and capital deepening/each task is allocated to the input with the lowest unit cost per effective task:

$${\displaystyle \frac{Y}{p\left(x\right)}}=\mathrm{m}\mathrm{i}\mathrm{n}\left({\displaystyle \frac{1+x}{f.v}},{\displaystyle \frac{\gamma }{c(1+e \left(zu-1\right))\tau }}\right)$$

(8)

This yields a task cutoff of $0<\alpha <1$ such that tasks below/above α are produced by capital/labor:

$$K={\int }_0^{\alpha } k\left(x\right)dx={\displaystyle \frac{Y}{r}}\alpha ; r={\displaystyle \frac{\alpha Y}{K}}$$

(9)

$$L={\int }_{\alpha }^1 l\left(x\right)dx={\displaystyle \frac{Y}{w}}\left(1-\alpha \right); w={\displaystyle \frac{(1-\alpha )Y}{L}}$$

(10)

And at the optimum, the cost of automation adjusted by productivity is the same as productivity-adjusted wages, that is

$$\begin{array}{c}{\displaystyle \frac{1+\alpha }{b}}={\displaystyle \frac{w}{a}}={\displaystyle \frac{\gamma }{a\tau }},\\ \alpha ={\displaystyle \frac{\left(1-\epsilon \right)\gamma fz}{c(1+e \left(zu-1\right))\epsilon {\tau }^{\prime }}}-1\end{array}$$

(11)

where $\alpha$ is the share of tasks performed by capital. It increases with capital automation capability $v$ and decreases with labor attractiveness as pushed by remote technologies’ effectiveness, z. Thus, technologies supporting WFH and automation interact through task reallocation.

3.3.2. Optimal Output/Price

In turn, the aggregate product price charged by the firm is the sum by part of the marginal cost of output, produced by labor or capital. This is written as follows:

$$\mathrm{l}\mathrm{n} P={\int }_0^{\alpha }\mathrm{l}\mathrm{n}\left({\displaystyle \frac{r(1+x)}{b}}\right)dx+{\int }_{\alpha }^1\mathrm{l}\mathrm{n}\left({\displaystyle \frac{w}{a}}\right)dx$$

(12)

After manipulations:

$$\ln P=\alpha \mathrm{l}\mathrm{n}\left({\displaystyle \frac{r}{fv}}\right)+(1+\alpha )\mathrm{l}\mathrm{n}(1+\alpha )-\alpha +(1-\alpha )\mathrm{l}\mathrm{n}\left({\displaystyle \frac{w}{c(1+e \left(zu-1\right))}}\right)$$

(13)

Injecting K and L from (9) and (10), into (13)

$$\begin{array}{c}\ln \left({\displaystyle \frac{r}{b}}\right)=\ln r-\ln f-lnv=\ln \left({\displaystyle \frac{\alpha Y}{K}}\right)-\ln b \\ =\ln Y+\ln \alpha -\ln K-\mathrm{l}\mathrm{n}(f.v)\end{array}$$

(14)

$$\ln \left({\displaystyle \frac{w}{a}}\right)=\ln Y+\ln \left(1-\alpha \right)-\ln L-\ln (c.(1+\epsilon \left(zu-1\right)))$$

(15)

Using (1) transformed in logarithm ($\mathrm{l}\mathrm{n}P=-k\mathrm{l}\mathrm{n}Y)$, then the optimal output becomes

$$\ln Y={\displaystyle \frac{\left[\begin{array}{c}\alpha \ln K+\left(1-\alpha \right)\ln L-\alpha \ln \alpha +\alpha \ln \left(fv\right)-\left(1-\alpha \right)\ln \left(1-\alpha \right)+\\ \left(1-\alpha \right)\ln (c.(1+\mathit{\epsilon }\left(\mathit{zu}-1\right))-\left(1+\alpha \right)\ln \left(1+\alpha \right)+\alpha \end{array}\right]}{1+k}}$$

(16)

Taking the exponential, Y becomes a Cobb–Douglas form, Y = Y(K,L):

$$Y=\mathrm{A}.{\left({\displaystyle \frac{fvK}{\alpha }}\right)}^{1-\mathrm{B}}{\left({\displaystyle \frac{(c.(1+\epsilon \left(zu-1\right)))L}{1-\alpha }}\right)}^{\mathrm{B}}$$

(17)

And A = $exp\left({\displaystyle \frac{\alpha +\left(1+\alpha \right)\ln \left(1+\alpha \right)}{1+k}}\right)$ and B = $({\displaystyle \frac{\left(1-\epsilon \right)\gamma fz}{c\left(1+e \left(zu-1\right)\right)\epsilon {\tau }^{\prime }}})/(1+k)$. Thus, the economy behaves like a Cobb–Douglas system with endogenous shares. Technology affects productivity by simultaneously shifting capital intensity, labor effectiveness, and the task cutoff. Furthermore, for α = 0 (labor-only firm), A = 1 and Y = $c.(1+\epsilon \left(zu-1\right))$ depends on the portion, $\epsilon$, of WFH, selected jointly by firm and workers while the labor elasticity, B, depends on market power ($\kappa$) and on task allocation ($1-\alpha$), which in turn depend on the relative performance of automation and WFH support technologies supporting, respectively, capital and labor.

3.3.3. Optimal Technology Mix

Considering profit maximization yields firm gross profit, $\pi =\left(1-k\right)Y,$ with Y defined above in (17), the optimal technologies adoption satisfies the following:

$$Y(1-k)(1-\alpha )\left({\displaystyle \frac{z\epsilon }{(1+\epsilon \left(zu-1\right))}}\right) =\mathrm{r}$$

(18)

$$Y(1-k){\displaystyle \frac{\alpha }{v}}=r$$

(19)

Equalizing marginal productivities determines the cutoff intensity:

$$\alpha ={\displaystyle \frac{z^2 ϵ}{1+z^2 ϵ+ ϵ(v+z-1)}}={\displaystyle \frac{z^2 ϵ}{1+z^2 ϵ+ ϵ\mathrm{D}}}\mathrm{w}\mathrm{i}\mathrm{t}\mathrm{h} D =v+z-1>1$$

(20)

And thus, better WFH tools increase capital intensity by lowering labor costs and enabling efficient automation. Equation (20), in particular, shows that the fraction of tasks allocated to capital is moderated by the effectiveness of capital-augmenting technologies ($v$) but increases in the effectiveness of WFH-supporting technologies ($z$) and of remote work intensity ($\epsilon$), as their effectiveness limits the demand for labor. Equalizing (20) to (11) leads to an equation of second degree in ε, for which the positive root is given by

$${\epsilon }^{*}=-{\displaystyle \frac{\gamma fv(u-aD)+a{\tau }^{\prime }u+\sqrt{\left[\gamma fv\right(u-aD)+a{\tau }^{\prime }u{]}^2+4\gamma ua[D(\gamma fv+a{\tau }^{\prime })+a{\tau }^{\prime }z]}}{2a\left[D\right(\gamma fv+a{\tau }^{\prime })+a{\tau }^{\prime }z]}}$$

(21)

And for $\Delta =(\alpha *,\epsilon *$), the following comparative statics hold:

• $𝜕\Delta /𝜕u>0$; ${\displaystyle \frac{𝜕\Delta }{𝜕\gamma }}>0$: Better remote collaboration tools and stronger worker preference for remote work raise both WFH intensity and capital intensity;
• $𝜕\Delta /𝜕a<0$ and $𝜕\Delta /𝜕{\tau }^{\prime }<0$: Higher on-site labor productivity or stronger on-site bias reduce both WFH as well as capital intensity;
• $𝜕{\alpha }^{*}/𝜕v>0$ but $𝜕{\epsilon }^{*}/𝜕v<0$: AI improvements in capital shift tasks toward capital, while reducing the optimal WFH share.

Thus, the economic intuition is that in a task-based production structure, improvements in AI capital-augmenting technologies ($v$) reduce reliance on remote labor. By contrast, improvements in WFH-supporting technologies ($z,u$) and stronger worker preferences for WFH ($\gamma$) increase both remote work and automation, as firms reorganize tasks away from purely on-site execution. This highlights that WFH technologies and AI automation are interdependent choices, or

Proposition 1.

WFH–automation linkage. WFH and automation are tightly linked. Improvements in AI technology increase the share of capital tasks and reduce the optimal WFH share. Meanwhile, improvements in WFH-supporting technologies increase both the WFH share and the extent of capital.

3.4. Technology Mix and Labor Productivity

The consequence of this interdependence is that the productivity effect of one workplace technology, e.g., AI, must control for the other tech investments, e.g., remote tools, if one wants to measure the true separate technology effect. As an example, Equation (22) computes the total elasticity of labor productivity with respect to remote support technology C, which decomposes the productivity effect of remote support technology into (i) a direct efficiency channel and (ii) an indirect task reallocation channel operating through automation versus remote technology effectiveness:

$${\displaystyle \frac{\mathrm{dln}\left(\frac{Y}{L}\right)}{dlnC}}=\left[{\displaystyle \frac{1-\alpha }{1+k}}{\displaystyle \frac{z\epsilon }{1+\epsilon \left(zu-1\right)}}+{\displaystyle \frac{1}{1+k}}\left(\ln \left({\displaystyle \frac{K}{L}}\right)-\ln \left(1+\alpha \right)\right){\displaystyle \frac{z^2}{\left[1+\epsilon \left(z^2+v+z-1\right){]}^2\right.}}\right]{\displaystyle \frac{d\epsilon . }{dlu }} . u.m$$

(22)

where from above, $\alpha >0; \mathrm{a}\mathrm{n}\mathrm{d} {\displaystyle \frac{d\epsilon }{du}}>$ 0. In particular, the first term ${\displaystyle \frac{1-\alpha }{1+k}}{\displaystyle \frac{z\epsilon }{1+\epsilon \left(zu-1\right)}}$ capturing the direct effect is scaled down by labor intensity, (1 − α)/(1+), while the second term that includes the indirect task reallocation ${\displaystyle \frac{z^2}{\left[1+\epsilon \left(z^2+v+z-1\right){]}^2\right.}}$ effect is scaled down by the ratio of K/L (in log). The natural interpretation of Equation (22) is that technology improving WFH effectiveness raises productivity both directly and indirectly by amplifying capital augmentation incentives, which is largely overlooked as an indirect channel in the literature. All the above then imply the following:

Proposition 2.

Optimal Technology Mix. Improvements in technologies that support working from home (WFH) raise labor productivity through direct efficiency gains and endogenous reallocation toward automation. In contrast, improvements in capital-scale efficiency raise productivity primarily through capital deepening and partially crowd out remote work. Consequently, reduced-form estimates that fail to account for the choice of technologies jointly misattribute productivity gains across workplace technologies.

4. Empirical Analysis

This section empirically tests the predictions of the task-based framework developed in Section 2 and Section 3, and especially the core implication of the model that work from home (WFH), automation, and complementary workplace technologies are jointly determined, and that estimating the productivity impact of WFH in isolation leads to biased inference. We therefore estimate the effect of WFH on labor productivity conditional on automation, capital intensity, and supporting digital technologies, explicitly allowing for non-linearities and complementarities predicted by the theory.

Two empirical choices also follow directly from the model. Firstly, because output is not comparable across industries, we use revenue per worker as the dependent variable. Revenue reflects both technical efficiency and pricing behavior. As shown in Section 3, market power scales the elasticity of output with respect to inputs, which we control explicitly. Secondly, because WFH and capital intensity are endogenous firm choices, we rely on an instrumental-variable strategy to recover causal effects. Finally, we have already mentioned that the analysis uses survey data from 4015 large firms, of which 3905 remain after validation and cleaning. The sample is remarkably large and covers both COVID and post-COVID time, and is able to capture the range of technology investment types linked to workplace. To the best of our knowledge, this is the first type of study to consider such a range of workplace technologies, and allows integrating the effect of technology on the task reallocation and capital/labor intensity of firms.

4.1. Specification and Hypotheses

4.1.1. Empirical Model

Recall that firms face a downward-sloping demand curve $P=Y^{-\kappa }$, implying revenue $R=PY=Y^{1-\kappa }$. Dividing by employment yields a labor revenue productivity:

$$\mathrm{l}\mathrm{n}(R/L)=(1-\kappa )\mathrm{l}\mathrm{n} A(\epsilon )+{\displaystyle \frac{1-\kappa }{1+\kappa }}\alpha (\epsilon )\mathrm{l}\mathrm{n}(K/L)$$

(23)

where revenue per worker depends on three channels: direct WFH productivity effects, captured by $A\left(\epsilon \right)$; arbitrage responses between technologies supporting WFH and automation, captured by $\alpha \left(\epsilon \right)$; and capital–labor intensity, $K/L$, whose productivity elasticity is scaled by market power. Because the theory also implies that $A$ depends on remote work technology quality $z$, and $\alpha$ depends on both remote and automation technology $v$, the empirical specification must jointly include WFH intensity, automation, capital intensity, and technology interactions. Guided by this structure, we estimate the following:

$$\begin{array}{cc}\mathrm{l}\mathrm{n}(R/L)=\hspace{1em}\hspace{1em}& c+{\beta }_1\mathrm{l}\mathrm{n}\epsilon +{\beta }_2(\mathrm{l}\mathrm{n}\epsilon {)}^2+{\beta }_3\mathrm{l}\mathrm{n}{v}^{\prime }+{\beta }_4\mathrm{l}\mathrm{n}{z}^{\prime }\\ & +{\beta }_5\mathrm{l}\mathrm{n}{z}^{\prime }\cdot \mathrm{l}\mathrm{n}\epsilon +{\beta }_6\mathrm{l}\mathrm{n}(K/L)+{\beta }_7\mathrm{l}\mathrm{n}(K/L)\cdot \mathrm{l}\mathrm{n}\epsilon \\ & +{\beta }_8\mathrm{l}\mathrm{n}(K/L)\cdot \mathrm{l}\mathrm{n}\epsilon \cdot \mathrm{l}\mathrm{n}{v}^{\prime }+{\beta }_{9}M+X^{\prime }\delta +w\end{array}$$

(24)

where $w$ is an error term and $X$ is a vector of controls described below.

4.1.2. Hypotheses

Each term in (24) also corresponds directly to a mechanism described by the theoretical model, as summarized by

Table 1. Model hypotheses.

Parameter	Expected Sign	Hypothesis/Interpretation	Theoretical Rationale
β1	>0	Baseline WFH elasticity: moderate increases in WFH raise revenue per worker, controlling for market power, capital intensity, and technology quality.	Task model: low–moderate WFH improves task execution and reduces frictions.
β2	<0	Concavity/diminishing returns of WFH: the marginal productivity effect of WFH decreases at high intensity.	Task allocation model predicts non-linearity: coordination costs dominate when WFH becomes excessive.
β3	>0	Effect of remote work technology: better remote support technologies raise revenue per worker.	Remote tech increases the productivity of WFH tasks, shifts cutoff α, and complements labor.
β4	>0	Effect of automation/AI technologies: automation intensity contributes positively to labor productivity.	Automation augments or substitutes tasks efficiently, boosting output per worker.
β5	>0	Interaction between WFH and automation: productive WFH complements capital deepening from AI/automation.	Joint adoption creates complementarities predicted by the task-based model.
β6	>0	Standard capital intensity effect: higher K/L raises revenue per worker.	Cobb–Douglas benchmark; capital deepening increases output per worker.
β7	>0	Additional capital effects: remote/automation technologies act as extra capital, enhancing labor productivity.	Model extension: technology deepening complements traditional capital.
β8	>0	Higher automation intensity improves productivity at optimal WFH levels.	Capital–labor substitution and augmentation effects in the task model.
β9	<0	Market power reduces revenue per worker: higher markups lower measured revenue per employee.	Pricing and demand structure: revenue per worker decreases as market power increases.

.

See Equation (24).

Said succinctly, non-linear WFH effects, ${\beta }_1>0$, ${\beta }_2<0,$ capture the concavity in $A\left(\epsilon \right)$. WFH–capital complementarity: ${\beta }_7>0$ captures ${\alpha }^{\prime }\left(\epsilon \right)>0$. Meanwhile, technology augmentation ${\beta }_3,{\beta }_4,{\beta }_8>0$ captures the role of automation and remote support technologies. Finally, market power ${\beta }_9<0$ reflects the scaling effect of markups on revenue productivity. Table 1 summarizes the hypotheses and their theoretical justification.

4.2. Endogeneity and Estimation Strategy

As shown in Propositions 1 and 2, work-from-home (WFH) intensity and capital intensity are jointly defined with productivity. Therefore, ordinary least squares would confound cause and effect. We address this issue by using two-stage least squares (2SLS), in which the current WFH intensity/capital–labor intensity is instrumented by its 2018 lagged value. Clearly, these lagged variables are best interpreted as predetermined rather than strictly exogenous. Identification relies on the presence of organizational inertia, adjustment costs, and regulatory frictions that cause workplace configurations to be slow-moving relative to short-run productivity shocks. Importantly, the pandemic represents a major structural break, weakening the link between pre-pandemic workplace choices and post-pandemic productivity realizations.

To further mitigate concerns related to time-invariant unobservables, we performed regressions with country–industry fixed effects, which account for persistent institutional, technological, and organizational characteristics jointly determined by national context and sectoral task composition. Additionally, we have implemented robustness checks based on a control function approach and placebo regressions using pre-pandemic productivity growth, which demonstrate robustness (see later). The results are not reported due to space limitations, but they exhibit the robustness of the base case against various sensitivities. Taken together, these results substantially reduce the risk that our estimates are driven by residual endogeneity or reverse causality.

4.3. Data and Validation

4.3.1. Data Origination

The estimation uses survey data collected at the end of 2020 by a consulting firm interested in the analysis of the COVID impact on firm resilience and recovery. The survey has been used elsewhere [4]. The survey covers both the post-COVID period (2021) and a retrospective pre-COVID benchmark (2018) and covers large enterprises headquartered in ten zones/countries—Europe (France, Germany, UK, Italy, and Spain), North America (US and Canada), and APAC (Singapore/China and Japan)—and covers 16 NACE-2 aggregate sectors with more than 1500 respondents (see Appendix A,

Table A1. Country–industry sample.

Sectors	%	Countries	%	Respondents	%
Banking + Capital Markets	12%	United States	34%	Chief Executive Officer	15%
Media/Entertainment	8%	Australia	9%	CS Officer/CIO	15%
Health	7%	Japan	9%	CTO	15%
Industrial Goods	7%	Germany	7%	COO	13%
Insurance	7%	Spain	7%	CMO	13%
Pharma/Life Sciences	7%	France	7%	CHRO	13%
Software/Platforms	7%	UK	7%	CDO/CAO	3%
Airline/Transport	6%	Italy	7%
Consumer Goods	6%	Canada	4%
Retail	5%	China/Singapore	7%
Automotive	4%
Capital Markets	4%
Chemicals	4%
Energy	4%
Natural Resources	4%
Utilities	4%
Aerospace	3%

and

Table A2. Evolution revenue, firm distribution.

Revenue Change	+2 Years Before COVID-19	Expected 2022
More than 10% decline	0%	1%
Between 5% and 10% decline	1%	4%
Between 0% and 5% decline	4%	27%
No change	5%	13%
Between 0% and 5% growth	46%	24%
Between 5% and 10% growth	37%	22%
More than 10% growth	6%	9%

).

Two-thirds of the companies generated revenues between USD 1 billion and USD 10 billion by the end of 2020. Large companies were selected because they tend to adopt and deploy new technologies faster than others and are therefore good predictors of broader technology adoption patterns (refs. [37,38,39]). On average, firms increased their revenue by 9% over the three-year period (2018–2021), despite experiencing a decline of 7% in 2020 due to the acute phase of the COVID-19 pandemic. Average labor productivity growth amounted to 4.9% over the same period. The average company achieved a profit margin of 15% and maintained a return on total capital employed (or equity) of 6% (approximately 12% for equity-based returns).

4.3.2. Data Qualification

The agency designed the survey and then had several market experts validate it. It was administered to a pre-selected list of company directors to test its readability and ensure it could be completed in under 20 min. The final questionnaire was sent to a panel of over 15,000 trained respondents, who were instructed to complete one to three surveys per year. Panel members were given the option to opt out, and 4015 companies responded—yielding an approximate response rate of 30%. After cleaning and validation, the final sample consisted of 3905 firms.

To ensure the reliability of the collected data, several validation steps were implemented, despite the limitations of each step. Regarding external validation, we collected accounting data for a random subsample of 100 firms and matched it with the survey responses. We tested for statistical differences between early and late respondents as a test of non-response bias. Finally, we performed a common-method variance test to examine systematic answer bias. Regarding matching, we found 92% consistency with the survey responses. Comparing the first 400 responses (approximately 10% of the sample) with the last 400 responses based on submission time revealed no statistically significant differences across key variables. The CMV test, which used factor analysis, revealed that the first factor captured 28% of the total variance, indicating a sufficiently low effect (ref. [40]). As an additional assurance of data integrity, respondents were reminded of the confidentiality of their responses. These steps provide strong confidence in the data quality, even if one may still acknowledge the risk of self-report bias, as CMV reduces but does not entirely eliminate risk.

4.4. Descriptive Statistics and Technology Diffusion

4.4.1. WFH Adoption and Persistence

Our measure of WFH refers to individuals spending at least one day per week working from their remote location of choice. Thus, the definition of WFH encompasses hybrid work rather than task-level remote intensity. As such, estimated coefficients should thus be interpreted as lower bounds of the true productivity effect of well-supported remote work.

As shown in

Table 2. WFH density evolution, % of employees working from home.

	Aerospace and Defense	Airline, Travel, Transport	Automotive	Banking (Retail or Investment)
WFH Pre-COVID-19	19.9	19.0	21.3	23.8
WFH Post-COVID-19	31.8	29.4	31.7	33.7
	Health	High Technology	Industrial Goods and Equipment	Insurance
WFH Pre-COVID-19	16.4	18.3	19.6	20.0
WFH Post-COVID-19	23.1	27.9	28.1	27.5
	Chemicals	TMT	Consumer Goods	Energy
WFH Pre-COVID-19	21.5	20.9	18.9	16.8
WFH Post-COVID-19	28.9	33.6	24.2	27.7
	Pharmaceutical, Bio Tech, Life Sciences	Retail	Software and Platforms	Utilities
WFH Pre-COVID-19	6.9	4.3	6.9	5.7
WFH Post-COVID-19	27.9	24.9	29.4	28.3

, the proportion of WFH in the sample was 16.5% in 2018, rising to 24.6% post-COVID, in line with other official statistics.

4.4.2. Workplace Technologies

Turning to workplace technologies, our data capture the extent to which digital remote collaboration tools have been deployed within the firm (driving $v^{\prime }$), as well as the adoption of AI technologies (driving $z^{\prime }$). The list of AI technologies is consistent with other academic surveys of AI adoption (Rammer et al., 2022 [41]) and includes machine learning, natural language processing, and computer vision, in addition to RPA.

Table 3. Workplace tech diffusion.

	AI	RPA	WFH
Stage of tech diffusion	Outside RPA		Tech tools
Not started	11%	9%	6%
Pilot	24%	42%	22%
Scaling in the organization	20%	32%	42%
Scaled up whole enterprise	44%	17%	30%

shows that by 2021, 72% of companies had adopted remote collaboration tools, and 59% had adopted RPA. However, only 30% of firms deployed remote work technologies across their entire workforce. Similarly, only 17% had implemented RPA across their entire enterprise. These figures confirm findings from other studies that workplace technology coverage is often incomplete, even among large firms [9,12].

Table 4. Deployment priorities among ten technologies, post-COVID. (List: cloud, AI, blockchain, machine learning, data management, cybersecurity, edge computing, RPA, remote work, data lake.)

Priority Spent	Collaborative Remote Work	RPA
Most important	6.5%	6.9%
Second most important	7.8%	0.1%
Third most important	7.5%	6.4%
Fourth most important	7.1%	6.6%
Fifth most important	6.5%	6.7%
Not in top 5	64.5%	66.7%

provides further insights into technology spending priorities during the pandemic across ten digital technology categories. Two-thirds of firms did not rank workplace technologies among their top five priorities, and fewer than 7% identified workplace technologies as their primary technology focus.

4.4.3. Control Variables

As discussed earlier, it is necessary to control for additional variables that affect the pattern of labor productivity growth, such as other firm asset variables beyond physical capital, labor hoarding, regulation, and broad economic productivity spillovers [42].

Regarding the first type of controls (assets), we include measures of digitalization (denoted “DIG”) and research and development (“RD”), as both types of capital are typically associated with higher labor productivity (refs. [43,44,45]). In terms of digitalization, we construct an index that captures the extent to which companies use five digital technologies—cloud computing, mobile technologies, broadband access, edge computing, and the Internet of Things (IoT)—outside of AI. We use research and development (R&D) expenditures as a measure of firms’ innovation and intangible capital. In our sample, the average firm allocates approximately 4.5% of its revenue to R&D, slightly higher than the typical share for large firms. Secondly, we include firm characteristics as control variables. Beyond accounting for the firm’s business scope (business-to-consumer, business-to-business, or business-to-business and business-to-consumer), we also include firm size and whether the firm is a multinational enterprise, as these variables are known to be associated with higher labor productivity [46].

In our sample, 42% of firms operate as MNEs. Additionally, 42% of firms are active in B2C markets, 41% in B2B, and the remainder in B2B2C. Thirdly, the market context plays a significant role in determining productivity. We therefore control industry and country using dummy variables, as these capture differences in competition, production processes, and spillovers within industries. Country dummies additionally account for broader national spillovers, regulatory frameworks, and labor policies, including those related to training, employment protection, and labor market flexibility [47,48]. Fourthly, the period under study includes the COVID-19 pandemic, which may have significantly disrupted productivity patterns, including the use of inputs at their optimal capacity. Our dataset also measures the impact on profit growth during the first six months of the outbreak, as well as in the subsequent six- and twelve-month periods. The difference in profit recovery in the six months following the pandemic compared to the first six months of lockdown and the acute endemic phase provides a strong indicator of firm resilience (“R”) and thus of effective factor utilization (refs. [49,50,51]).

5. Results and Discussions

5.1. Diagnosis

We estimate Equation (24) using least squares with predetermined lagged variables as instruments (IV). Using one lagged variable for each endogenous regressor (lagged WFH for current WFH and lagged capital intensity for current K/L) implies that the model is exactly identified. We report first-stage (Kleibergen–Paap) F-statistics to document the relevance of the instruments. In all specifications, the F-statistics are well above the conventional instrumental variables (IV) threshold of 10. Secondly, we performed Durbin–Wu–Hausman tests to compare OLS and IV estimates. The DWH tests systematically reject the null hypothesis of exogeneity of WFH and K/L. This confirms that our lagged instruments capture meaningful endogeneity rather than sampling noise (see the bottom of

Table 5. Labor productivity regression results, IV, 2021.

Effects	Parameter	RPA + WFH Coefficients	Standard Error	AI + WFH Coefficients	Standard Error
	β1	3.56 **	1.45	3.01 *	1.62
	β2	−0.57 ***	0.21	−0.32 **	0.17
	β3	0.91	0.78	0.76 ***	0.17
	β4	1.2 ***	0.38	1.91 **	0.88
	β5	0.31 **	0.14	0.28 *	0.16
	β6	28.1 ***	8.21	29.4 ***	9.12
	β7	0.31	0.22	0.34 *	0.19
	β8	0.8 *	0.45	0.61	0.45
	β9	−8.2 ***	2.1	−6.9 *	3.7
Controls	B2B	0.45 *	0.22	0.49	0.32
	B2C	−0.73	0.64	−0.88	0.67
	MNE	1.31 ***	0.43	1.26 **	0.61
	Size	−0.34	−0.26	−0.24	−0.17
	Resilience	1.58 **	0.67	1.34 ***	0.47
	R&D	2.12 **	0.97	1.98 *	1.11
	Digital assets	1.42 **	0.68	1.72 **	0.78
Fixed effects inclusion
	Country	YES		YES
	Industry	YES		YES
Statistics	R-square	0.423		0.404
	Prob (F) sign	<0.01%		<0.01%
Instrument	K-P F	16.95 *		21.2 *
	DWH	6.12 **		7.08 ***

Yearly revenue productivity: 3.66%. B2B2C, non-resilience, and local firm is default, digital assets: percentage of use; R&D % spent in portion of revenue. *** p < 1%; ** p < 5%; * p < 10%; IV for K/L and WFH are lagged variables (2018).

).

Since lagged variables are more predetermined than causal, we remind the reader to interpret the estimates as quasi-causal rather than fully structural. Equation (24) was also estimated by introducing country–industry fixed effects, which absorb time-invariant unobservables that could otherwise bias the lagged IV strategy. The results, available from the author, remain stable. As an additional robustness check, we implemented a control function approach. Residuals from the first-stage regressions of WFH and capital intensity are included in the productivity equation to control for the remaining correlation between the regressors and unobserved productivity shocks. We find that the main coefficients remain stable, confirming that our results are not driven by any large form of residual endogeneity.

To operationalize Z and V, we treat workplace technology effectiveness as a latent deployment construct rather than a binary adoption decision. The survey provides two complementary indicators for each technology class: (1) the breadth of organizational deployment (pilot, partial, or enterprise-wide) and (2) the strategic priority of investment. Using simple adoption dummies would overlook substantial heterogeneity in deployment depth and organizational integration, as well as exacerbate measurement error. Since these variables are correlated, we aggregate the indicators using principal component analysis (PCA). The first principal component captures the common variation underlying deployment breadth and investment priority, explaining more than 70% of the total variance across workplace technology categories. We use this component as our empirical proxy for technology intensity because it aligns closely with the model’s interpretation of z and v, the continuous effectiveness parameters, and it reduces attenuation bias relative to dummy-based measures (

Table 6. PCA of deployment and priority.

Variable	Corr (Deployment, Priority)	λ1	λ2	PC1 Variance Explained
RPA	0.45	1.45	0.55	72.5%
Remote Tech	0.67	1.67	0.33	83.5%
AI	0.62	1.62	0.38	81.0%

).

5.2. Results and Hypothesis Testing

The results are displayed in Table 5. The first set of columns uses RPA only, while the second set of columns are based on broader AI technologies, and act as a reality check. In general, the pattern is rather similar for both columns, with the exception that β₃, β₇ are statistically significant when using more general AI than only RPA technologies. We also highlight goodness of fit in the range of 40%, and the F-statistic of global significance is always positive at less than 1%. Thus, the model seems to be well specified.

We first concentrate on control variables, then focus on capital types, and then on WFH and on the role of supportive workplace technologies.

Table 7. Productivity growth split, 2021, average of both estimations.

	Revenue Productivity Type
	Capital Deepening			TPF Growth
Parameters	Workplace Tech	Digital Tech	Physical	WFH	Workplace Tech	R&D
β1				0.55
β2				−0.15
β3	0.22
β4	0.35
β5					0.12
β6			0.96
β7				0.06
β8					0.30
β9
R&D						0.10
Digital assets		0.78
Total	0.58	0.78	0.96	0.46	0.42	0.10
Total of total			2.32			0.98

aggregates coefficients from Table 5 into economically meaningful productivity contributions.

5.2.1. Controls

There is no statistically significant difference in control variables between the market segments covered by companies (business-to-business versus other types, such as business-to-consumer or business-to-business-to-consumer). Multinationals contribute approximately 0.5 percentage points to annual labor productivity growth compared to domestic firms, which aligns with previous findings. Resilient firms (i.e., those that recovered from the pandemic) seem to contribute 0.6 percentage points to annual labor productivity growth, compared to other firms. This effect is relatively large but smaller than other estimates in the literature. Digital assets (excluding AI) contribute 0.8% to yearly labor productivity gains, which is at the high end of typical estimates. Meanwhile, the effect of R&D intensity on labor productivity growth is relatively small at 0.1% per year.

5.2.2. Capital Deepening

There are three types of capital deepening. Beyond digital, non-AI assets, there is traditional physical capital deepening, as well as new remote and artificial intelligence workplace capital investments. Regarding traditional capital, the latter exerts a positive and significant impact on labor productivity (β₆ > 0), albeit in the lower range of elasticities (0.26 to 0.3) observed in the labor productivity literature (typical figures are often evaluated at about 0.3–0.4 [52,53]). This elasticity level is an order of magnitude higher than that of workplace technology, given the significant difference in investment allocation. Given the significant growth in the budget allocated to workplace technologies between 2018 and 2021, workplace technologies contributed about 50 basis points to yearly productivity growth, roughly two-thirds of the contribution of physical deepening. In total, all types of capital deepening contributed 2.3 points of revenue productivity growth per year between 2018 and 2021.

5.2.3. Standalone WFH Labor Productivity Effects

An important insight from the model is that working from home (WFH) can have a differentiated impact through total factor productivity and, thus, indirectly affect labor productivity shifts. We find that β₁ > 0, which provides strong evidence supporting a positive relationship between WFH and changes in labor productivity, as in [54]. Despite controlling for downward bias in the typical estimated effect of WFH, the estimate in this research is nevertheless lower than in rare studies demonstrating a positive effect of WFH (e.g., [6,7,27]). However, those positive effects were found in specific firms and activities, such as call centers and online firms. In a more general model of optimal task production arbitrage, the complete effect of WFH (outside tech support) must also account for the hypotheses that β₂ < 0 and β₇ > 0. Both hypotheses are herewith confirmed (for β₇; however, only for AI, and not for RPA metric), highlighting that productivity impact decreases with a higher WFH intensity, and that WFH and capital are strategic complements with respect to labor premises as a result of task optimization models.

Taking all effects together, the total productivity effect of WFH before tech support is measured by ${\beta }_1+{2\beta }_2+{\beta }_{6}ln\left({\displaystyle \frac{K}{L}}\right)$= 0.45 points of yearly labor productivity growth.

5.2.4. Effects of Workplace Technologies

Turning to the hypothesis that workplace technologies should affect the trajectory of labor productivity growth, outside pure capital deepening, we find this to be the case ${\beta }_4, {\beta }_8$, > 0, with, however, a larger effect from RPA/AI automation than from remote work support technology. In total, remote technology support adds 0.4 points of growth, on top of 0.6 of capital deepening, or a total of 1 point of revenue productivity growth.

5.2.5. Range of Productivity Effects Associated with Work from Home

Our estimates suggest that the impact of WFH on total and labor productivity varies substantially with its intensity and the presence of complementary workplace technologies. At low to moderate levels of remote work—corresponding to hybrid arrangements of approximately one to two days per week—WFH is associated with positive productivity gains, reflecting improved task execution, reduced commuting friction, and better task work mode matching.

Quantitatively,

Table 8. Estimated impact annual labor productivity (bps).

WFH (%)	Low Tech	High Tech
10	25	34
20	43	64
30	55	90
40	61	112
50	60	130
60	53	144
70	39	154
80	18	160
90	−11	162
100	−45	158

shows that, at the current average WFH density, labor productivity increases by approximately 0.3 to 0.6 percentage points per year, holding other factors constant. This magnitude is economically meaningful and comparable to the lower bound of traditional capital-deepening effects observed in firm-level productivity studies.

However, as WFH intensity rises further, productivity gains exhibit clear diminishing returns. Beyond a threshold corresponding to high remote work intensity, coordination and communication frictions dominate, leading to a decline in productivity. Furthermore, in the absence of complementary workplace technologies, high WFH intensity may reduce labor productivity.

Importantly, investments in remote collaboration tools and automation shift this productivity profile outward (see the second column in Table 8). Firms that deploy these technologies experience a wider range over which WFH remains productivity-enhancing, with positive labor productivity effects persisting at higher levels of remote intensity. In these firms, the combined effect of WFH and workplace technologies can account for up to one percentage point of annual labor productivity growth, primarily through total factor productivity (TFP) improvements rather than pure labor substitution.

Existing studies typically report positive or negative average productivity effects of WFH, which often depend on context or occupation. Our results suggest that these divergent findings are not contradictory but reflect different positions along a non-linear productivity curve. Studies focusing on voluntary hybrid arrangements capture the upward-sloping segment of this curve, while studies examining extensive or poorly supported remote work capture its downward-sloping segment. By explicitly modeling technology complementarities, our framework explains why similar levels of WFH can yield very different productivity outcomes across firms.

5.3. Discussions

Using the results, we can discuss a few important issues:

5.3.1. Does WFH Matter?

In line with our theoretical framework, we find that work from home (WFH) indeed affects labor productivity growth, both directly as part of TPF growth, and indirectly through the addition of complementary workplace technologies. When one talks about the WFH effect, however, we find that the technology support effect is relatively large, more than one point of productivity growth on, for a total of about 1.5 points, and in a major way, due to capital deepening.

The model also confirms β₂ < 0, or the diminishing marginal productivity of WFH as well as β₇ > 0 (complementarity between WFH and physical capital) in alignment with the task allocation mechanism whereby WFH initially increases efficiency but becomes concave at higher intensities, and productivity gains are amplified in capital-augmented production, limiting labor, shifting to WFH-based workers’ preferences.

5.3.2. How Important Is the Technology Dimension as Part of Workplace Productivity Improvements?

The fact that workplace technologies (RPA/AI and remote digital tools) generate 1 point of labor productivity growth nominally, or about 0.6, the outside price effect is rather comparable to other estimates of new technology adoption on labor productivity. Acemoglu, et al. [55] document an impact of automation on labor productivity gains of 4 points over 5 years in the French manufacturing sector (period 2010–2015). Ref. [13] show that changes in robot density have contributed just less than 0.4 points to annual labor productivity growth, in the period 1993–2007. Our period of analysis is 2018–2021, and at the time of a major push (capital deepening) linked to COVID. This capital-deepening effect has reduced significantly post-COVID, however.

5.3.3. How Important Is the Broad AI Dimension as Part of Productivity Improvement?

The productivity effects of workspace technology deployment are not solely confined to the subset of work technologies. The impact is like other firm studies on the use of AI on revenue growth, such as in [56,57] (see Gao, X., & Feng [58] for a synthesis of AI impact on productivity.) Regarding digital (outside AI) diffusion, we find that its effect on labor productivity growth is smaller than that of AI. Furthermore, the impact is lower than a recent OECD study that concluded that digital technologies’ diffusion has pushed firm labor productivity by about 5% per year [44].

5.3.4. What Are the Managerial Implications?

The findings offer several actionable insights for firms and managers responsible for workplace transformation. Firstly, working from home (WFH) can improve productivity, but only when supported by robust digital collaboration tools and redesigned workflows. Firms that treat WFH as a mere location policy without investing in coordination technologies are unlikely to sustain productivity gains. Operationally, this means that managers should not ask whether to “allow WFH,” but rather how to redesign task execution under hybrid conditions. Tasks that are modular, standardized, and executable by one person are more likely to benefit from remote work. In contrast, highly interdependent or creative tasks require stronger coordination mechanisms or a partial on-site presence. Thus, optimal WFH policies are task-specific rather than firm-wide and should be designed at the workflow level rather than based on headcount ratios.

Secondly, automation and AI technologies yield larger and more substantial productivity gains than work-from-home (WFH) tools alone. However, the extent of their impact depends critically on organizational integration. Firms that deploy automation without redesigning workflows or aligning remote work practices may experience fragmentation rather than efficiency gains. Therefore, managers should view automation as a coordination technology that reshapes how tasks are sequenced, monitored, and combined across locations, rather than as a labor-saving device per se. Thirdly, the complementarities between workplace technologies matter. Firms that deploy WFH tools, automation, and digital coordination tools together can exploit profitable arbitrage between labor and capital. This leads to higher total factor productivity and labor productivity growth. This implies that technology investment decisions should be evaluated as portfolios rather than as isolated projects. From a capital allocation perspective, underinvestment in one component (e.g., remote collaboration tools) can also sharply reduce the returns on others (e.g., automation).

Importantly, the results also caution managers against excessive intensity of remote work. Beyond moderate hybrid arrangements, productivity gains diminish and may become negative if coordination costs become dominant. Therefore, firms should closely monitor performance metrics and treat WFH intensity as a continuous decision variable subject to optimization rather than as a binary choice. Finally, evidence suggests that only a minority of firms have fully exploited the potential of integrated workplace technologies. Many firms remain stuck in the partial deployment or pilot stages, which limits productivity gains. This highlights a managerial opportunity: productivity improvements from workplace technologies are possible, but require deliberate organizational redesign, skill development, and leadership commitment.

6. Conclusions

The impact of technological innovation on corporate performance has long been recognized. The novelty of this article lies in its focus on a specific class of intelligent workplace technologies—automation, AI, and work-from-home technologies—that directly reshape how work is organized, rather than merely how it is executed. Complementarities will be essential for unlocking the next wave of productivity growth.

Following Baldwin and Okubo (2024) [1], we zoom in on corporate adoption of digital workplace technological innovations—such as software automation, which alters the mix of tasks performed by workers, and remote work technologies that shift the focus of work outside the office (“remote intelligence”). The importance of these technologies within enterprises should not be understated: they proved highly effective during the COVID-19 lockdown, and they have since become mainstream, with the majority of large corporations in our sample having scaled or fully deployed these workplace technologies by the end of 2021.

Our framework is based on the predictions of a task-based model, which assumes that firms will invest in automation and work-from-home (WFH) technologies to optimize work organization. This approach considers the optimal mix of inputs and supporting technologies, taking into account trade-offs between cost and performance. We empirically test the impact of workplace technology adoption on labor productivity trajectories of thousands of large firms. We calibrate our estimation to the specific impact of work-from-home technologies to benchmark against the recent literature on the effects of WFH on productivity.

By integrating theory and firm-level evidence from nearly 4000 large firms across countries and industries, our paper makes three core contributions. Firstly, it demonstrates that estimated WFH productivity effects cannot be interpreted independently of firms’ broader workplace technology portfolios. Reduced-form estimates that omit automation and digital coordination tools embed endogenous task reallocation and misattribute productivity gains or losses. Secondly, the paper reconciles conflicting findings in the literature by distinguishing between task-level substitution and firm-level complementarity. Although automation may displace remote labor in routine tasks, firms that adopt integrated bundles of work-from-home (WFH) technology and digital tools experience higher overall productivity growth through capital deepening and reduced coordination frictions. This distinction clarifies why empirical studies report both crowding-out effects and productivity gains depending on task composition and technological support. Thirdly, the analysis provides quantitative evidence on the range and non-linearity of WFH productivity effects. Rather than asking whether working from home (WFH) “works” on average, the paper shows that productivity gains arise at moderate levels of remote work and diminish at high intensity unless supported by substantial workplace technology investment. This finding shifts the debate from binary assessments of remote work toward optimization problems centered on task allocation and technology mix.

Beyond its immediate findings, the paper has broader implications for research and policy. For researchers, the results underscore the importance of modeling technology adoption as a system and accounting for endogenous task reallocation when estimating productivity effects. Future work could extend this framework using task- or worker-level data to study skill heterogeneity, wage dispersion, and employment dynamics. For policymakers, the findings underscore that gains in productivity from WFH and AI depend on investments in complementary digital infrastructure, training, and organizational capabilities. Regulations that address remote work or automation in isolation risk overlooking these interactions.

In sum, intelligent workplace technologies neither guarantee a productivity boost nor pose an inevitable threat to labor. Their impact depends on how firms design tasks, deploy technologies, and manage coordination under hybrid conditions.

As firms continue to experiment with new work models in the post-pandemic economy, understanding these complementarities will be essential for unlocking the next wave of productivity growth. Beyond short-term productivity effects, our findings have direct implications for the sustainability of firms and work systems. Integrated workplace technologies contribute to economic sustainability by supporting durable productivity growth through capital deepening and improved task allocation rather than one-off efficiency gains. They also contribute to social sustainability, as hybrid work arrangements supported by digital coordination tools allow firms to reconcile worker preferences for flexibility with organizational performance constraints. Finally, by reducing commuting intensity and enabling more efficient use of physical infrastructure, hybrid work models may indirectly support environmental sustainability, even though environmental outcomes are not directly measured in this study. Taken together, the results suggest that sustainable workplace transformation depends less on the adoption of individual technologies and more on the coherent design of work systems that jointly optimize productivity, resilience, and long-term resource use.

This study has several notable limitations. Firstly, the data are survey-based, and many company respondents cannot be fully identified, preventing us from matching survey responses with external datasets. Secondly, we lack direct information on worker skills and abilities, although skill distributions affect both the feasibility of WFH and the scope for automation; we partially control for this through industry and firm fixed effects. Thirdly, the post-COVID period we examine corresponds to the immediate aftermath of the lockdown, when firms had not yet fully recovered from the pandemic shock. Differences in resilience versus underlying performance may therefore influence the results, even though we attempt to control for firm-level resilience. Nevertheless, these technologies are becoming increasingly powerful and are already contributing to labor productivity growth among firms that choose to adopt them—much like the capital-deepening effects of traditional physical technologies in previous decades.

Several extensions to this research naturally arise. Firstly, future work could refine the estimates by incorporating more detailed task information and explicit measures of worker skills. Secondly, a broader perspective integrating both the worker and firm sides would help capture potential resistance to technological change, which may shape productivity outcomes. Finally, the effects should also be examined in relation to employment dynamics—particularly the potential decline in traditional on-site jobs, as both WFH and automation technologies reshape their scope.