AI and Immunity to the COVID-19 Pandemic

Abstract

We evaluate economic immunity to the COVID-19 pandemic accorded by a sector’s exposure to artificial intelligence (AI). Using an event study design, we causally estimate the impact of AI industry exposure (AIIE) on shifts in economic activity following the first pandemic-induced lockdown in India. We use payments data from one of India’s largest payment gateways and find that a one-standard-deviation increase in AIIE arrests the post-pandemic decline in industry payments by a third. High AIIE of neighboring sectors further improves a sector’s economic resilience, emphasizing the importance of supply chain partners in technology adoption and change. Further analyses speak to the mechanisms underlying the documented impact of AI on economic activity. Notably, we present evidence suggesting that the observed effects of AI are distinct from those attributable to IT investments or digitization, that AI investments are associated with superior information capabilities prior to the pandemic that arguably enabled agility and responsiveness, and that AI-enabled adaptation occurs in a manner that complements labor rather than displacing it. We discuss the implications of our findings in terms of the assessment of the business value of AI and management of disruptions associated with crises events.

Keywords

AI Economic Activity Remote Work Business Agility Resilience

“Machines don’t fall ill, they don’t need to isolate to protect peers, they don’t need to take time off work.” Daniel Susskind¹

1. Introduction

Artificial intelligence (AI) is widely touted as the “most important general-purpose technology of our era” (Brynjolfsson and Mcafee, 2017). A rich body of research in technology and operations management emphasizes how AI, big data, and machine learning can be used by firms to make better decisions, including better demand forecasting (Cui et al., 2022; Chang et al., 2021; Lau et al., 2018; Zhu et al., 2021), improved pricing, procurement, inventory, and service operations decisions (Ban and Rudin, 2019; Cui et al., 2022; Cohen, 2018; Ellis et al., 2018; Geva and Saar-Tsechansky, 2021; Swaminathan, 2018; Yang et al., 2022); and enhanced customer management and care (Gursoy et al., 2019; Nenova and Shang, 2022; Queenan et al., 2019; Yoganarasimhan, 2020). However, there is little systematic documentation of empirical evidence of the effects of firm or sectoral exposure to AI on fundamentals and economic performance (see Felten et al., 2021 and Raj and Seamans, 2019 for an articulation of this gap in the literature). This study addresses this important research gap by providing an empirical assessment of the economic impact of exposure to AI across a broad cross-section of industries in India during the COVID-19 pandemic. Disruptions associated with the pandemic resulted in an exogenous and widespread increase in the need to reorganize business operations, and our study, therefore, constitutes a test of claims in prior work that the most significant effects of AI in the future would result from its enablement of reorganization of business practices (Brynjolfsson and Mcafee, 2017; O’Shaughnessy et al., 2023).

The COVID-19 pandemic was an unprecedented crisis that posed exceptional risks and challenges to individuals, businesses, institutions, and societies (Ding et al., 2021; Reinhart, 2020). Policy responses to the pandemic, including social distancing and isolation or lockdowns resulted in frequent, abrupt, and substantive cratering of demand and disruptions in business operations.² Emergent studies (Chen and Biswas, 2021; Dohale et al., 2022) and industry research reports (McKendrick, 2021) suggest that AI facilitated greater business continuity during this crisis via reorganization to meet the challenges brought upon by the pandemic. Greater exposure to AI renders processes and work more amenable to reorganization with learned systems³ at similar performance-cost ratios that was necessary for business operations after the pandemic-induced disruption. Thus, AI-enabled reorganization potentially aided both in the pre-emptive and reactive stages of decision-making as a response to the lockdown (Dohmen et al., 2023; Pournader et al., 2020).

Consistent with our thesis, a report published by PwC India (PwC, 2020) notes a marked increase in the integration of AI with business processes within the Indian and the global economy in the wake of the pandemic. Based on extensive interviews and surveys conducted with senior executives and decision-makers, the report concludes that 45% of Indian organizations increased AI use as a result of the pandemic, and attributes the increased adoption to the role AI technologies played in facilitating a range of business functions, including contactless customer interactions, remote work-related productivity enhancements, and dealing with novel decision-making contexts, wherein prior knowledge and expertise were potentially of limited use. The report further notes a qualitative shift post COVID-19 in AI-related challenges from being primarily technical (e.g., unavailability of data) to being more related to business performance (e.g., measuring the business value of AI), indicating greater integration of AI technologies with business operations. In line with the findings of the report, a systematic review of media reports that we describe in Section 4.1 shows that adoption of AI post the pandemic within the Indian context spanned industries, business domains, and various forms of AI.

Existing research emphasizes that AI technologies differ in important ways from traditional IT. Notably, contemporary AI technologies demonstrate the ability to learn and improve in novel ways through data and experience, which in turn allows for complex forms of automation that obviate human intervention in decision-making, and at the same time, makes these technologies more opaque to humans. For example, Berente et al. (2021) describe in detail the corresponding attributes of autonomy, learning, and inscrutability of AI technologies and their implications for effective management. Prior work has also linked these very characteristics of AI to its vast potential to transform human–machine collaboration, and business processes and functions in ways that are qualitatively different from those enabled by conventional IT (Brynjolfsson and Mitchell, 2017; McAfee and Brynjolfsson, 2008). The clear conceptual distinction between AI technologies and traditional IT in terms of their nature, implementation, and opportunities with regard to transformation of business processes and functions makes it important to assess AI’s economic impact specifically.

The COVID-19 outbreak in India and the consequent nationwide lockdown (shutdown) imposed by the Indian government present a unique and ideal empirical setting to causally identify the impact of AI exposure on economic activity. First, the timing, severity, and scope of the lockdown in India were unanticipated, constituting an unexpected shock to firms. Figure EC1.1 in the E-Companion shows the timeline of COVID-19 cases in India, where the shaded area corresponds to the duration of the first lockdown. The latter was imposed very early in the outbreak’s trajectory when India had registered fewer than 500 active COVID-19 cases. It came into force on March 24, 2020, was initially enforced until April 14, 2020, and later extended until May 31, 2020. Thereafter, the government announced a series of unlock phases that lasted through September 2020. Second, the lockdown was enforced just 4 hours after it was announced. The unexpected and exogenous nature of the shock and its speed suggest that firms’ ability to respond in a timely or systematic fashion to the unfolding crisis was very limited. Thus, it was firms’ pre-existing conditions (including AI exposure) that engendered resilience to and recovery from the crisis. Pre-existing AI exposure, together with the exogenously imposed need to reorganize business operations during the pandemic, allows for empirical identification of the business impact of AI. Finally, unlike in the United States and other advanced economies, where lockdown intensity varied across regions within countries, the first phase of lockdown in India was homogeneous in intensity across regions, sectors, and firms, pervasive in its impact, and extremely stringent.⁴ Other than a few essential services, all commercial, industrial, travel, religious, and cultural activity were shut down during this period.⁵ These aspects of the shock create the opportunity for a study that uses a narrow time window to test the causal relationship between exposure to AI and economic performance and in turn, immunity to crises.

For our primary analyses, we leverage high frequency payments data at the $State \times Sector$ level as a direct measure of real economic activity. The payments data comprises official data on daily value and volume of payments from the payments gateway, unified payments jnterface (UPI), managed by the National Payments Corporation of India (NPCI).⁶ The granular nature of these data allows for the estimation of the impact of AI exposure on business continuity over a short time window, thereby mitigating identification challenges associated with variation in other extraneous factors that can impact economic activity over a longer time horizon. Our primary measure of exposure to AI extends the suitability for machine learning (SML) index developed by Brynjolfsson et al. (2018) to India. Survey responses from about 3,100 Indian workers to the task evaluation rubric in Brynjolfsson et al. (2018) are used to score occupations for their SML. The score is a pre-pandemic measure, as the survey was administered between June and December 2019, well before the first incidence of COVID-19 in India and implementation of a nationwide lockdown. Following Felten et al. (2021), the occupation level scores are aggregated to estimate AI industrial exposure (AIIE) to achieve concordance with the payments data.

We implement a difference-in-differences strategy to assess differential changes in payment activity across sectors during the 21-day nationwide lockdown in India. We find that sectors with higher AIIE demonstrate more resilient economic activity. On average, the lockdown led to an average decline of 26% in the total value of payments and 25% in the count of payments received across sectors. However, sectors with high AIIE witness a significantly lower decline in payments received. A one-standard-deviation increase in AIIE stems the decline in payments received by the sector by 9 percentage points during the lockdown. Further, we find that the benefits of AIIE to a sector are moderated by the AIIE of downstream sectors. Specifically, we use the input–output production matrix (two-digit NIC sectors) to build AIIE scores for each sector’s input (upstream) as well as output (downstream) production linkages. We find that a sector with high output AIIE—that is, high AI exposure of downstream sectors—suffers a lower decline in payments received post-lockdown, relative to the baseline case discussed above. That is, when the technology and business logic are not aligned with customers’ capabilities, the resultant adjustment costs for customers ultimately reduce the net benefits to adopters of the technological innovation within the supply chain (Afuah, 2004; McElheran, 2015).⁷

The aforementioned analyses, which use high frequency payments data to causally estimate the impact of AI exposure on business continuity over a short time period, reflect the response of the median firm within a sector in the aftermath of the lockdown. We complement these analyses with those based on alternative measures of both business continuity and AI exposure. Specifically, we demonstrate the robustness of our findings to firm-level indicators of business continuity obtained from financial information in the Prowess database. Our findings also remain robust to the use of alternative measures of AI exposure estimated in prior work (Brynjolfsson et al., 2018; Felten et al., 2021), serving as a validity check for the survey methodology used to estimate our AI scores. Finally, we report analyses that document a positive impact of AI exposure on a range of worker outcomes during the pandemic. Many of these worker outcomes arguably also serve as indicators of business continuity (e.g., continued hiring, continued monetary compensation, provision of non-pecuniary benefits, job security) and therefore, the corresponding analyses, which use measures of firm and occupational-level AI exposure, lend additional support to our primary findings.

We additionally report analyses that shed light on the mechanisms underlying the observed economic resilience engendered by AI during the lockdown. First, we use a systematic review of media reports following the lockdown in India to document the diverse domains to which AI technologies were applied and the multitude of ways in which they allowed for reorganization of business processes and functions during the pandemic. We also empirically distinguish the demonstrated effects of AI on economic performance during the lockdown from those of IT or digitization. Second, we present findings from a pre-pandemic survey that suggest that firms’ AI exposure relates to enabling conditions for reorganization of business processes. Notably, we find that higher AI scores relate to superior linkages to suppliers/customers, greater business process integration, and improved information sharing internally within the organization, all of which were potentially valuable in allowing firms to adapt and reorganize processes swiftly in the wake of the pandemic. Finally, in addition to complementing our primary findings, results pertaining to the effects of AI exposure on worker outcomes described above allow us to speak to the central debate within the AI literature regarding whether AI-driven reorganization of businesses is likely to complement or substitute labor (Brynjolfsson and Mitchell, 2017). These analyses suggest that exposure to AI was associated with both more favorable objective (hiring shifts, compensation shifts, and non-pecuniary benefits) and subjective (job security, work satisfaction, and overall evaluation of the employer) outcomes and in turn, indicate that at least over the period we study, exposure to AI enabled business continuity in a manner that complemented rather than substituted labor.

Our findings yield important insights for the literature on the business value of AI. A limited number of studies that have examined the performance implications of AI innovations have produced conflicting results, with some documenting positive effects of AI on business outcomes (Babina et al., 2024; Rock, 2019) and others finding negative or mixed effects (Li et al., 2021; Lui et al., 2022). The varying conclusions of these studies emphasize the need for further research to adjudicate these differences. Recent work in the domain of operations management (Choi et al., 2022; Mithas et al., 2022) further attests to this need. Emergent research (Brynjolfsson and Mcafee, 2017; O’Shaughnessy et al., 2023) suggests that the most substantive effects of AI in the future would manifest in the redesign of business processes and transformation of the nature of work. Therefore, one explanation for these mixed findings is that the need or intent to reorganize business processes was heterogeneous across samples of firms studied in prior work, which, in turn, resulted in variation in observed effects of AI on business value. Moreover, when endogenously determined by firms, reorganization plans are likely to be long term and not designed to bring immediate productivity gains (Bughin et al., 2018), thereby, making it challenging to assess the relationship between AI-enabled reorganization and business value. The pandemic resulted in an exogenously imposed, unanticipated, and widespread need to reorganize business processes to meet challenges created by associated disruptions, thereby allowing for empirical identification of the business impact of AI. Our study leverages granular, high-frequency payments data and an event study design to assess whether AI exposure resulted in greater resilience immediately after the pandemic-induced disruptions and the associated need to reorganize business operations kicked in.

Our findings also have important implications for prior work on operational disruptions and risk management. A growing stream of research in this domain evaluates how firms can mitigate risks from natural and man-made hazards (e.g., Bakshi and Kleindorfer, 2009; Kleindorfer and Saad, 2005; Mehrotra and Schmidt, 2021; Schmidt and Raman, 2021). Such unanticipated events can result in substantive disruption of firms’ production management and supply chain infrastructure, leading to large-scale economic losses (Kleindorfer et al., 2003). Accordingly, the literature on operational risk management has investigated the various firm actions required to mitigate risks and sustain operations during disruptions (e.g., Dong and Tomlin, 2012; Tomlin and Wang, 2011; Tomlin, 2006). Prior research has argued that firms’ prior investments in digital infrastructure, such as control systems (Schmidt and Raman, 2021) and information technology (Kleindorfer and Saad, 2005; Park et al., 2023) determine the extent to which their operations remain resilient during such catastrophic episodes. Relatedly, existing research pertaining to disruptive technologies discusses the potentially substantial impact of integrating AI with firms’ production management and supply chain operations (Choi et al., 2022; Sharifi et al., 2021). However, there is little research that systematically examines the economic impact of AI integration on firms’ operations during crises. An assessment of the role of AI in the specific context of the COVID-19 induced crisis is particularly important because of both the severity of the impact of the pandemic on business operations and the widespread adoption of AI during this time.

The remainder of this paper is organized as follows. We describe our empirical context and data in the next section. Section 3 discusses our primary results. Section 4 details the mechanisms. Section 5 discusses and concludes the study.

2. Data

We use multiple sources of data, as summarized in Table EC2.1 in the E-Companion. The first dataset relates to the measurement of AI. Acemoglu et al. (2022) articulate three measures of AI that capture the amenability of occupations to AI capabilities. We use two of these measures in our assessment.⁸ Our primary measure of AI draws on responses to the Suitability to Machine Learning (SML) survey, a 23-item rubric from Brynjolfsson et al. (2018) that evaluates SML of the primary task of each respondent’s occupation along a 5-point scale. In contrast to traditional IT investments, AI and ML technologies are designed to improve themselves over time. Brynjolfsson et al. (2018) note: “Instead of requiring an inventor or developer to codify or code each step of a process to be automated, a machine-learning algorithm can discover on its own a function that connects a set of inputs X to a set of outputs Y as long as it is given a sufficiently large set of labeled examples mapping some of the inputs to outputs.” This improvement in AI technologies reflects greater measurability of task inputs and outputs, availability of large datasets that can be used to train algorithms, the absence of complex task logic, specialized skills, and task evolution over time, among others. Our measure of AI exposure, which instantiates the SML index developed by Brynjolfsson and Mcafee (2017) to the Indian context, incorporates these work attributes and is distinct from measures of IT investments in the literature (that is limited to measuring IT assets and expenses).

Our survey was administered in English and vernacular languages to 3,099 individuals employed across the 106 3-digit occupations defined in the 2004 National Classification of Occupations (NCO).⁹ The sample of respondents was selected to capture sufficient variation within a given occupation as specified in the 2017 Periodic Labor Force Survey (PLFS). The survey was administered between June and December 2019, well before the registration of the first COVID-19 case in India and enforcement of the nationwide lockdown (see Figure EC1.1 in the E-Companion for survey timeline). To create the occupation-level AI scores, we consider the mean SML score of all respondents from the focal occupation. Following prior work (Felten et al., 2021), we use the occupation-level AI exposure scores and the distribution of workers across the 106 occupations within each sector as per the 2017 PLFS to estimate sector-level exposure scores at the 2-digit NIC level. Specifically, we estimate ${AIIE}_{j}$ for each sector $j$ as the weighted average of occupation-level AI scores across all the occupations $o$ in sector $j$ :

{AIIE}_{j} = \sum_{o = 1}^{106} w_{o j} \times {AI}_{o}

(1)

where

w_{o j}

is the weight of occupation

o

j

, that is, the proportion of employees in sector

j

that work in occupation

o

{AI}_{o}

is the SML score of occupation

o

. The mean (s.d.) of sectoral scores is 3.53 (0.18). We report AIIE scores for all sectors in Table EC2.2 in the E-Companion. Whereas architecture, computer science, and consultancy services have high AIIE, the mining and agriculture sectors score low.

The next set of data in our analyses, high-frequency measures of economic activity, notably, daily payments data across sectors for the period January 2019 to April 2020, are obtained from the National Payments Corporation of India (NPCI). Specifically, we use data from the UPI, a real-time payment system developed by NPCI to facilitate electronic transactions. Using UPI, we track the volume and value of daily economic transactions at the $S t a t e \times S e c t o r$ level. The UPI reports sectors by four-digit Merchant Category Codes, which we map to two-digit NIC sectors manually. In the financial year 2019–20, UPI accounted for 47% by volume and 13% by value of all financial transactions across payment systems operational on the NPCI platform. We exclude individual- or household-level payments to confine our analysis only to business transactions. We present a summary of UPI payments across 45 sectors and 36 states (includes both states and union territories) in Table 1. We see a clear decline in the mean count and value of transactions during the post-lockdown period.

Table 1.

Summary: UPI payments data.

	Number of payments (’000)		Value of payments (INR million)
	Pre-lockdown	Post-lockdown	Pre-lockdown	Post-lockdown
	(1)	(2)	(3)	(4)
Sector	46	45	46	45
State	36	36	36	36
State $\times$ Sector	1,456	1,378	1,456	1,378
Mean	7,489	6,435	4,127	3,371
Std. Dev.	27,902	21,876	11,499	9,144
Median	88	34	87	50

Notes: The table gives a summary of sector-wise payments data from UPI. The pre-lockdown period is March 2–24, 2020, while post-lockdown is March 25–April 14, 2020. We exclude Holi (March 9, 2020) and public curfew (March 22, 2020) from the pre-lockdown period.

We draw on three additional data sources to uncover the mechanisms through which exposure to AI generates resilience. The first dataset (the AI and Future of Work Survey) comprises a survey of 301 executives and pertains to the adoption and applications of AI within their firms. The survey was administered by authors in 2019, before the lockdown. The survey also captured exposure to AI using the same rubric as Brynjolfsson et al. (2018). Therefore, using these data, we are able to assess specific business capabilities that AI investments relate to prior to the pandemic.

The remaining two datasets are used to assess whether firms with high exposure to AI reorganized in a manner that substituted labor or complemented it. The first of these datasets, which pertains to objective worker outcomes, comprises responses to an employee survey, called the Step-Up survey, conducted during the lockdown by Naukri.com, India’s largest online job search portal. The objective of the survey was to provide job seekers with information on companies and sectors after the lockdown (March–November 2020). The data captures variation across firms with regard to post-pandemic hiring and compensation shifts, as well as non-pecuniary benefits, and therefore, allows for an evaluation of whether these worker outcomes differed systematically across firms with varying exposure to AI. These data are available in aggregate form at the firm-level for 2,971 firms—employee responses are aggregated to provide firm-level estimates of the proportion of employees reporting the focal outcome. The second dataset, which pertains to subjective worker outcomes, consists of employee perceptions of their employing firms and was obtained from AmbitionBox, a subsidiary of Naukri.com. Analogous to Glassdoor.com, AmbitionBox allows employees to anonymously provide ratings, reviews and opinion texts of their employing firm across a range of dimensions, such as job security, work satisfaction, and company culture. The data comprises 119,060 reviews from 1,624 companies and allows for an assessment of whether subjective employee outcomes varied systematically across occupations with varying exposure to AI. Further details on our datasets are provided in Section EC2 in the E-Companion.

3. Impact of AI on Economic Activity

3.1. Empirical Framework

We estimate the following difference-in-differences specification to test the impact of the lockdown on payments received by a sector and whether such impact varies by the latter’s pre-pandemic exposure to AI:

r_{s, t}^{j} = α + β_{1} {Lockdown}_{t} + β_{2} {AIIE}^{j} \times {Lockdown}_{t} + δ_{s}^{j} + ε_{s, t}^{j}

(2)

where

r_{s, t}^{j}

is the log residual of the outcome variable in sector

j

in state

s

on date

t

.¹⁰ The outcome variables include the count and the total value of payments. The variable

{Lockdown}_{t}

is equal to one if the lockdown had been instated on date

t

, else it is equal to zero. In our case,

{Lockdown}_{t}

equals 1 for the time period March 25–April 14, 2020.

{AIIE}^{j}

is a pre-pandemic measure of industrial exposure to AI, as defined in Equation (1).

{AIIE}^{j}

is measured at the level of the two-digit NIC sector and is the same across all states

s

The term ${AIIE}^{j} \times L o c k d o w n_{t}$ captures the differential impact of the lockdown across sectors with different AI scores. The first difference results from comparing outcomes before and after the lockdown, while the second difference results from the variation in sectoral AI scores. Finally, $δ_{s}^{j}$ include $State \times Sector$ fixed-effects, which capture time-invariant heterogeneity in payments at the $State \times Sector$ level. For instance, if the payments prior to the lockdown are heterogeneous at $State \times Sector$ level, $δ_{s}^{j}$ would account for any such variation. All standard errors are clustered at the $State \times Sector$ level.

We expect the overall impact of ${Lockdown}_{t}$ on payments to be negative, which would be reflected in a negative value of $β_{1}$ . However, our primary variable of interest is the interaction term ${AIIE}^{j} \times {Lockdown}_{t}$ . Given our hypothesis that post-lockdown payments would be less impacted for sectors with high AIIE, we expect the coefficient $β_{2}$ to be positive. That is, high AIIE provides economic resilience against lockdown-induced disruptions.

To obtain an unbiased estimate of $β_{2}$ , we estimate Equation (2) by restricting our sample to 21 days before and after the lockdown. As described earlier, the lockdown was enforced nationwide and came into effect 4 hours after it was announced. The exact date, length, spatial coverage, and severity of the lockdown were unanticipated. Therefore, restricting our sample to a 21-day window on either side of March 25 rules out potential endogeneity concerns. For instance, our specification accounts for the possibility that sectors with higher values of digital payments are more likely to have high AIIE. The $δ_{s}^{j}$ term controls for any such non-time varying heterogeneity.

Our choice of a narrow 21-day event window rules out two other concerns. First, over a longer estimation window, both the lockdown and payments could be influenced by a third common factor, leading to a biased estimate of $β_{2}$ . Second, ${AIIE}^{j}$ is a pre-pandemic measure that is exogenous to the lockdown and can be taken as given during the first lockdown period. Over a longer period, sectors can invest in AI leading to changes in their AI scores. Such changes cannot be captured by a pre-pandemic AI score and would lead to erroneous measurement of AI and biased estimates.

Finally, we use residuals as our dependent variables as this allows us to filter out any seasonality in the payments data. We use a longer time-series, January 1, 2019–April 14, 2020, to address seasonality concerns through a battery of fixed-effects at the $State \times Sector$ level. Among other influences, these fixed effects control for fluctuations in payments due to seasonal demand or supply at the $State \times Sector$ level. Such extensive fixed-effects cannot be used directly in the final estimation in Equation (2), as our identification comes from changes in payments during the event-study window. Using residuals allows us to measure changes in payments sans any seasonality around the lockdown event. In the following subsections, we report our findings based on the residuals. However, the results remain similar with the use of raw payments data, and we report them along with other robustness checks in Section 3.3.

3.2. Less Economic Disruption in Sectors With High AI

The results for the impact of the lockdown on payments are reported in Table 2. The first three columns pertain to the count of payments received, while the last three columns relate to the value of payments. Column (1) reports the impact of the lockdown on the count of payments after controlling for $State \times Sector$ fixed-effects. The coefficient of $Lockdown$ is negative and significant, suggesting that the lockdown led to a significant decline in the count of payments. The coefficient $- 0.29$ corresponds to a 25% decline ( $e^{- 0.29} - 1 = - 0.25$ ) in the count of payments. These numbers are aligned with the broader decline in the Indian economy during this period as the GDP contracted by 24.4% in April–June 2020 quarter. Column (2) presents results for the specification that includes the interaction term. The coefficient of $Lockdown$ continues to be negative. The coefficient of $AIIE \times Lockdown$ is positive and significant, showing that sectors with high AIIE witness more transactions. Quantitatively, a one-standard-deviation increase in $AIIE$ , that is, 0.18 points, increases the count of payments by 8.6%. The aggregate impact on a sector is the sum of $β_{1}$ and $β_{2}$ , adjusted according to its AIIE score. We find that the aggregate impact of the lockdown is negative for the sector with the highest AIIE score. This sector witnesses a 15% decline in the number of payments. In contrast, for a sector with median AIIE, the decline is 25%. The results are robust to inclusion of $Date$ fixed-effects in our regression, which account for any daily variation in payments (column (3)). The magnitude and significance of the coefficient of the interaction term remains similar to that in column (2).

Table 2.
Impact of AI on sectoral payments.

Count of payments Value of payments

(1) (2) (3) (4) (5) (6)

Lockdown $- {0.287}^{* * }$ $- {1.960}^{ * }$ $- {0.299}^{ * }$ $- {2.107}^{ * }$

(0.015) (0.308) (0.015) (0.308)

AIIE $\times$ Lockdown ${0.465}^{ * }$ ${0.467}^{ * }$ ${0.502}^{ * }$ ${0.505}^{ * *}$

(0.085) (0.085) (0.085) (0.085)

Observations 47786 47786 47786 47786 47786 47786

R-Squared 0.20 0.21 0.26 0.08 0.08 0.11

State $\times$ Sector FE Yes Yes Yes Yes Yes Yes

Date FE No No Yes No No Yes

	Count of payments	Value of payments
Lockdown	$- {0.287}^{* * *}$	$- {1.960}^{* * *}$		$- {0.299}^{* * *}$	$- {2.107}^{* * *}$
	(0.015)	(0.308)		(0.015)	(0.308)
AIIE $\times$ Lockdown		${0.465}^{* * *}$	${0.467}^{* * *}$		${0.502}^{* * *}$	${0.505}^{* * *}$
		(0.085)	(0.085)		(0.085)	(0.085)
Observations	47786	47786	47786	47786	47786	47786
R-Squared	0.20	0.21	0.26	0.08	0.08	0.11
State $\times$ Sector FE	Yes	Yes	Yes	Yes	Yes	Yes
Date FE	No	No	Yes	No	No	Yes

Notes: The dependent variable in columns (1)–(3) is the residual of the log(count of payments) and in columns (4)–(6) is the residual of the log(value of payments). The results are based on estimating Equation (2). The lockdown dummy equals one for the post-lockdown period (March 25, 2020–April 14, 2020) and zero for the pre-lockdown period (March 2, 2020–March 24, 2020). We exclude the days of Holi (March 9) and the day of first public curfew in India (March 22) from the pre-lockdown period. The observations for regressions are at State $\times$ Sector $\times$ Date level. The State $\times$ Sector FE are at the two-digit NIC sector level. AIIE has a mean (standard deviation) of 3.53 (0.18). Clustered standard errors are at State $\times$ Sector level in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1.

Columns (4)–(6) present similar results for the impact of AIIE on total value of payments. Column (4) shows that the impact of the lockdown is negative and significant on the total value of payments. In column (5), the coefficient of $Lockdown$ is negative, but that of the interaction term $AIIE \times Lockdown$ is positive. A one-standard-deviation increase in AIIE increases the total value of payments by 9.4%. The results hold even after controlling for $D a t e$ fixed-effects, as reported in column (6). Once again, sectors with high AIIE see a lower decline in the total value of payments, and the interaction coefficient is similar in magnitude to that in column (5). To summarize, high AIIE provides economic immunity to the sectors after the lockdown, which manifests in a lower decline in the quantum and value of payments.

We next rule out the possibility that our results are driven by differential pre-lockdown trends across the sectors. Given that we use residuals in our specification and restrict the estimation period to 21 days before and after the lockdown, this concern is minimal. Nevertheless, we estimate the change in daily payments received by modifying Equation (2). We replace $AIIE \times Lockdown$ with $AIIE \times Period$ , where $Period$ belongs to half-month periods from February to April. We use the period March 1–15 as the baseline.

We plot the event study coefficients with log(value of payments) as the dependent variable in Figure 1. The vertical line divides the pre- and post-lockdown periods. For sectors with high AIIE, we find a sharp increase in the value of payments after the lockdown, which persists during the entire post-lockdown period. In contrast, most coefficients before the lockdown are insignificant, which rules out that sectors have diverging pre-trends before the lockdown. Similarly, E-Companion Figure EC1.2 reports the results for the log(count of payments). We find that the coefficients after the lockdown are positive and significant, indicating that high AIIE increases the count of sectoral payments.

Figure 1.

Impact of AI on value of payments over time. Notes: The figure plots the event-study coefficients on the interaction term $A I I E$ $\times$ $P e r i o d$ and their 99% confidence intervals with March 1–15, 2020, as the baseline. The dependent variable is the residual of log(value of payments). The vertical line corresponds to the lockdown and the coefficients to its right correspond to the post-lockdown period. We exclude the days of Holi (March 9) and public curfew (March 22) in the pre-lockdown period. Standard errors are clustered at State $\times$ Sector level.

3.3. Robustness

Our results are robust to alternate dependent and independent variables, other subsamples, and placebo tests. We detail these tests below.

Alternate Measure of AI Exposure: We assess the robustness of our results to alternate measures of AIIE; notably, the measures of AIIE developed by Brynjolfsson et al. (2018) and Felten et al. (2021). We use a cross-walk between the Standard Occupational Classification (SOC) of the US and the NCO to aggregate the occupation-level SML scores from the above papers and create sector-level AIIE scores. The results based on these alternate measures are reported in Table EC1.1 in the E-Companion. For both count (columns (2) and (3)) and value (columns (5) and (6)) of payments, the coefficient of the interaction term, $A I I E$ $\times$ $L o c k d o w n$ , is positive and significant. That is, sectors with high AI exposure have both a higher count and value of payments immediately after the lockdown.

Raw Payments Data: Our main results use residuals as outcome variables in our regressions as this allows for correction of seasonality in the payments data. As robustness, we alternatively estimate the impact of AIIE on the raw count and value of payment transactions in Table EC1.2 in the E-Companion. Our results remain robust to these alternate dependent variables.

Firm-Level Outcomes: Our primary analyses estimate the impact of AIIE on sectoral payments using high-frequency payments data, which serves as a proxy measure of the activity of the median firm within a sector. We additionally assess the robustness of the results based on payments data by evaluating the impact of AIIE on firm-level activity in the quarters following the lockdown. To this end, we utilize quarterly financial statements sourced from Prowess, covering the period October 2019–September 2020. We estimate a DID model to compare firms’ outcomes in the pre- versus post-lockdown period as a function of differential sectoral AI exposure. We report corresponding results in Table EC1.3 in the E-Companion. Our estimates show that in sectors with high AIIE, firm-level economic activity, as manifest in sales, expenses, and wages, remains elevated after the lockdown, as evidenced by the positive and statistically significant coefficient of the interaction term ( $AIIE$ $\times$ $Lockdown$ ). The results are invariant to using a battery of fixed-effects.

Placebo Lockdown 2019: We conduct a placebo test using pseudo-lockdown in the 2019 to assess if our results are driven by time trends that are not completely captured by our main specification. Specifically, we randomly draw 1,000 dates in 2019 as pseudo-lockdowns and re-estimate the baseline regression with a 21-day event window around the pseudo lockdown. The density plots of the coefficient estimates on the interaction term (AIIE $\times$ Lockdown) for both the count and amount variables of this simulation exercise are presented in Figure EC1.3 in the E-Companion. In general, there is no reason to expect that AIIE should differentially impact the payments in periods outside the lockdown. We indeed find that all the coefficients of the interaction term are close to zero in these regressions.

Non-Essential Sectors: Our primary results are based on 45 sectors. However, some sectors, like food and utility, were allowed to function normally during the lockdown, as they were deemed essential by the government. We exclude these sectors and re-estimate Equation (2) using only the 38 non-essential sectors. We find that our results are robust to the exclusion of these essential sectors from the sample (Table EC1.4 in the E-Companion).

Including Industry-Specific Measures of Resilience: We test the robustness of our results to including sectoral indicators (WFH, restricted labor index, leverage) that could also determine sectoral resilience after the lockdown. Our main results remain unchanged (Table EC1.5 in the E-Companion).

Post-lockdown Adoption of E-payment Technologies: We rule out the differential adoption of payment systems after the lockdown as an alternate explanation for our main results. Using additional data on e-payments adoption in India, we find no evidence of differential adoption of payment technologies in the 21-day window around the event window (Figure EC1.4 in the E-Companion). We also do not find differential adoption of payment systems as a function of sectoral AI exposure (Figure EC1.5 in the E-Companion).

3.4. AI Adoption of Neighboring Sectors Matters

The findings described so far relate to the direct impact of AIIE on the focal sector’s payments. In this section, we assess how linkages between the focal sector and other sectors impact payments received by the former. Owing to linkages between sectors within the economy, it is possible that a sector itself has high AI exposure but is unable to function during the lockdown due to dependency on other sectors.

Like before, we use the payments data to test the impact of average AI score of input (or upstream) and output (or downstream) sectors on economic activity of sector $j$ . We construct the input AI score, $Input$ ${AIIE}^{j}$ , for a sector $j$ by computing a weighted sum of the $AIIE$ scores of all sectors upstream to $j$ . The input AIIE score is given by: $Input$ ${AIIE}^{j}$ $= \sum_{i \neq j}$ $w_{i j} {AIIE}^{i}$ , where $w_{i j}$ is the weight of sector $i$ as input in sector $j$ . The summation is taken over all sectors $i$ , except $j$ . Similarly, output AIIE score, $Output$ ${AIIE}^{j}$ $= \sum_{i \neq j}$ $w_{j i} {AI}^{i}$ , is built using weighted $AIIE$ scores of sectors downstream to $j$ . The weights come from the sectoral input–output matrix at the two-digit NIC level constructed from the 2012 to 2013 supply-use tables, as in Venkatesh and Pal (2018). $AIIE$ score is based on the sectoral SML scores used in the previous subsection.

To estimate the impact of AI exposure of neighboring sectors on economic activity, we augment Equation (2) with input $A I I E$ (and/or output $A I I E$ ) scores as additional explanatory variables. The results are reported in Table 3, where all specifications are saturated with $State \times Sector$ and $Date$ fixed-effects. Panel (a) presents the results for the count of payments. In column (1), the coefficient of $Input$ $AIIE$ $\times$ $Lockdown$ is positive and significant, that is, a sector that buys inputs from sectors with high average $AIIE$ score receives higher payments during the lockdown.¹¹ The same is true for sectors that sell output to downstream sectors with high $AIIE$ (column (2)). The effect is much stronger in the latter case, as a sector is less likely to receive payments from its downstream sectors if they are adversely impacted by the lockdown. In column (3), we include the lockdown interaction with sectors’ own $AIIE$ and $Input$ $AIIE$ as explanatory variables. Here, the coefficient of $Input$ $AIIE$ $\times$ $Lockdown$ is insignificant, and the entire decline in the count of payments is captured by sector’s own $AIIE$ . In contrast, only the coefficient of $Output$ $AIIE$ $\times$ $Lockdown$ is significant in column (4). In column (5), we conduct a horse race between the three measures and find that $Output$ $AIIE$ and $Input$ $AIIE$ capture the increase in the count of sectoral payments. However, the $Output$ $AIIE$ score is more important. A one-standard-deviation increase in $Output$ $AIIE$ , equal to 0.085, leads to a 21% increase in the count of payments.

Table 3.
Impact of AI on payments of neighboring sectors.

(1) (2) (3) (4) (5) (6)

Panel (a): Count of payments

AIIE $\times$ Lockdown ${0.395}^{* * }$ 0.056 $- 0.016$ $- 0.025$

(0.094) (0.089) (0.095) (0.117)

Input AIIE $\times$ Lockdown ${0.861}^{ * }$ 0.387 ${0.505}^{ }$ ${2.120}^{ * }$

(0.218) (0.251) (0.255) (0.267)

Output AIIE $\times$ Lockdown ${2.500}^{ * }$ ${2.430}^{ * }$ ${2.306}^{ * }$ ${2.368}^{ * }$

(0.233) (0.254) (0.272) (0.289)

Observations 47786 46274 47786 46274 46274 41588

R-Squared 0.26 0.30 0.26 0.30 0.31 0.33

Panel (b): Value of payments

AIIE $\times$ Lockdown ${0.424}^{ * }$ 0.119 0.066 ${0.204}^{}$

(0.094) (0.091) (0.096) (0.116)

Input AIIE $\times$ Lockdown ${0.942}^{* * }$ ${0.433}^{}$ 0.371 ${1.478}^{* * }$

(0.216) (0.248) (0.259) (0.267)

Output AIIE $\times$ Lockdown ${2.410}^{ * }$ ${2.261}^{ * }$ ${2.170}^{ * }$ ${2.023}^{ * *}$

(0.209) (0.231) (0.248) (0.272)

Observations 47786 46274 47786 46274 46274 41588

R-Squared 0.11 0.12 0.11 0.12 0.12 0.14

State $\times$ Sector FE Yes Yes Yes Yes Yes Yes

Date FE Yes Yes Yes Yes Yes Yes

	(1)	(2)	(3)	(4)	(5)	(6)
Panel (a): Count of payments
AIIE $\times$ Lockdown			${0.395}^{* * *}$	0.056	$- 0.016$	$- 0.025$
			(0.094)	(0.089)	(0.095)	(0.117)
Input AIIE $\times$ Lockdown	${0.861}^{* * *}$		0.387		${0.505}^{* *}$	${2.120}^{* * *}$
	(0.218)		(0.251)		(0.255)	(0.267)
Output AIIE $\times$ Lockdown		${2.500}^{* * *}$		${2.430}^{* * *}$	${2.306}^{* * *}$	${2.368}^{* * *}$
		(0.233)		(0.254)	(0.272)	(0.289)
Observations	47786	46274	47786	46274	46274	41588
R-Squared	0.26	0.30	0.26	0.30	0.31	0.33
Panel (b): Value of payments
AIIE $\times$ Lockdown			${0.424}^{* * *}$	0.119	0.066	${0.204}^{*}$
			(0.094)	(0.091)	(0.096)	(0.116)
Input AIIE $\times$ Lockdown	${0.942}^{* * *}$		${0.433}^{*}$		0.371	${1.478}^{* * *}$
	(0.216)		(0.248)		(0.259)	(0.267)
Output AIIE $\times$ Lockdown		${2.410}^{* * *}$		${2.261}^{* * *}$	${2.170}^{* * *}$	${2.023}^{* * *}$
		(0.209)		(0.231)	(0.248)	(0.272)
Observations	47786	46274	47786	46274	46274	41588
R-Squared	0.11	0.12	0.11	0.12	0.12	0.14
State $\times$ Sector FE	Yes	Yes	Yes	Yes	Yes	Yes
Date FE	Yes	Yes	Yes	Yes	Yes	Yes

Notes: For Panels (a) and (b), the dependent variables are the residuals of log(count of payments) and log(value of payments), respectively. The lockdown dummy equals one for the post-lockdown period (March 25, 2020–April 14, 2020) and zero for the pre-lockdown period (March 2, 2020–March 24, 2020). For a given sector j, the variable AIIE measures the sector’s own exposure to AI, Input AIIE reflects the AI exposure of its input sectors and output AIIE reflects the AI exposure of its output sectors. We exclude the days of Holi (March 9) and the day of first public curfew in India (March 22) from the pre-lockdown period. The observations for regressions are at State $\times$ Sector $\times$ Date level. The State $\times$ Sector FE are at two-digit NIC sector-level. Columns (1)–(5) are for all sectors, while Column (6) is estimated on the subset of sectors that exclude construction, manufacturing, real estate, and utility. Clustered standard errors are at State $\times$ Sector level in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1.

Results pertaining to the total value of payments in Panel (b) mimic those in Panel (a) (columns (1)–(4)). Column (5) in Panel (b) shows that Output AIIE has maximal impact on the value of payments received among the three measures. Here, only the coefficient of Output AIIE is positive and significant. One possible reason for the insignificant effect of Input AIIE could relate to difficulties in accessing inputs due to mobility restrictions during the lockdown, which, in turn, reduce the positive effects of AI on the focal sector. This issue should be less relevant for sectors that rely less on physical inputs. To evaluate this possibility, we re-estimate the model after excluding sectors that rely extensively on physical inputs for production (i.e., the construction, manufacturing, real estate, and utility sectors) (Column (6)). We find that the coefficient of Input AIIE after lockdown is positive and significant in this case, highlighting the importance of both input and output AIIE.

Overall, Table 3 demonstrates that AI adoption by a sector has an impact on outcomes of neighboring sectors. The sectors connected to those with greater AI exposure show more economic resilience during the lockdown. These results relate to the work by Cui et al. (2022), who demonstrate how AI systems can create and deliver value in the context of procurement. These positive spillovers across the supply chain can occur in multiple ways as the integration of firms’ own systems with its partners can facilitate customer requirements management, demand forecasting, or minimization of supplier risk (Helo and Hao, 2022; Xu et al., 2023).

4. How Does AI Build Resilience?

In this section, we report analyses that shed light on the mechanisms underlying the observed economic resilience facilitated by AIIE during the pandemic. We perform three distinct sets of analyses, each of which relates to different aspects of the processes that account for the effects of AI on business continuity. Specifically, these analyses demonstrate that (i) AI facilitates business continuity in ways beyond those attributable to IT investments or digitization, (ii) pre-pandemic AI exposure relates to superior information management capabilities that arguably increases reorganizational capability for focal firms during the pandemic, and (iii) AI-enabled reorganization positively impacts a range of worker outcomes during the pandemic.

4.1. AI Builds Resilience Beyond IT/Digitization Exposure

The results reported in prior sections demonstrate the impact of AIIE on resilience in sectoral economic activity. Even though we conceptually distinguish AI and IT, it is possible that our primary results are driven by variation in IT investments and digitization levels across sectors, which, in turn, are correlated with AIIE. For instance, IT investments such as software applications or communication infrastructure can provide economic resilience to pandemic-induced lockdowns, especially during periods in which workers stayed at home to work (Park et al., 2023). Similarly, the digitization of products and services renders them modular, and in turn, amenable to generative and distributed innovation (Rahmati et al., 2020) that were valuable to product and service redesign during the pandemic. Given that IT also potentially contributed to resilience during the pandemic, it is important to examine whether the effects of AI we document persist after accounting for those attributable to IT.

Our baseline specification, through the inclusion of $State \times Sector$ fixed-effects, indirectly accounts for the average IT or digitization levels in a given state and sector. However, there could still be a differential impact of these measures on payments over time. To rule out such concerns, we empirically demonstrate that the contribution of AIIE to economic resilience is distinct from that of IT and digitization (Table 4).

Table 4.
Impact of AI vs. IT/digitization on sectoral payments.

Count of payments Value of payments

(1) (2) (3) (4) (5) (6)

AIIE $\times$ Lockdown ${0.353}^{* * }$ ${0.451}^{ * }$ ${0.244}^{ }$ ${0.390}^{ * }$ ${0.389}^{ * }$ ${0.158}^{}$

(0.083) (0.097) (0.097) (0.086) (0.096) (0.095)

IT $\times$ Lockdown ${0.887}^{* * }$ ${0.767}^{ * }$ ${0.687}^{ * }$ ${0.769}^{ * }$

(0.143) (0.150) (0.137) (0.141)

Digitization $\times$ Lockdown ${0.327}^{ * }$ ${0.501}^{ * }$ ${0.357}^{ * }$ ${0.552}^{ * *}$

(0.098) (0.106) (0.101) (0.101)

Observations 42256 45778 40248 42256 45778 40248

R-Squared 0.28 0.28 0.30 0.13 0.12 0.14

State $\times$ Sector FE Yes Yes Yes Yes Yes Yes

Date FE Yes Yes Yes Yes Yes Yes

	Count of payments	Value of payments
AIIE $\times$ Lockdown	${0.353}^{* * *}$	${0.451}^{* * *}$	${0.244}^{* *}$	${0.390}^{* * *}$	${0.389}^{* * *}$	${0.158}^{*}$
	(0.083)	(0.097)	(0.097)	(0.086)	(0.096)	(0.095)
IT $\times$ Lockdown	${0.887}^{* * *}$		${0.767}^{* * *}$	${0.687}^{* * *}$		${0.769}^{* * *}$
	(0.143)		(0.150)	(0.137)		(0.141)
Digitization $\times$ Lockdown		${0.327}^{* * *}$	${0.501}^{* * *}$		${0.357}^{* * *}$	${0.552}^{* * *}$
		(0.098)	(0.106)		(0.101)	(0.101)
Observations	42256	45778	40248	42256	45778	40248
R-Squared	0.28	0.28	0.30	0.13	0.12	0.14
State $\times$ Sector FE	Yes	Yes	Yes	Yes	Yes	Yes
Date FE	Yes	Yes	Yes	Yes	Yes	Yes

Notes: The dependent variable in columns (1)–(3) is the residual of the log(count of payments) and in columns (4)–(6) is the residual of the log(value of payments). The sector-level IT intensity measure is the ratio of IT assets to total capital and calculated from Prowess data by authors. The sector-level Digitization measure comes from Rahmati et al. (2020). The lockdown dummy equals one for the post-lockdown period (March 25, 2020–April 14, 2020) and zero for the pre-lockdown period (March 2, 2020–March 24, 2020). We exclude the days of Holi (March 9) and the day of first public curfew in India (March 22) from the pre-lockdown period. The observations for regressions are at State $\times$ Sector $\times$ Date level. The State $\times$ Sector FE are at the two-digit NIC sector level. Clustered standard errors are at State $\times$ Sector level in parentheses. ***p < 0.01, **p < 0.05, *p < 0.1.

To differentiate the effect of AIIE from general IT intensity, we include $IT \times Lockdown$ as an additional explanatory variable in Equation (2). Sector-level IT intensity is measured as the ratio of IT assets to total capital as in Xue et al. (2012).¹² We re-estimate the impact on payments and report corresponding results in columns (1) and (4). The coefficients of $AIIE \times Lockdown$ remain positive and significant. The coefficient of $IT \times Lockdown$ is also positive and significant, suggesting that as expected, IT intensity has an independent positive impact on payments after the lockdown.

Recent advances in Information Systems research posit that modern IT systems have evolved from being mere back-office technologies to integral components of diverse products and services (Henfridsson and Bygstad, 2013; Yoo et al., 2010). Therefore, we additionally assess the robustness of our results to the inclusion of a recent measure of digitization proposed by Rahmati et al. (2020)¹³ that captures the average ability of the sector to: (i) develop digital products and services, (ii) embed digital components into products and services, and (iii) reconfigure and recombine know-how into new classes of products and services. The measure encompasses a broader conceptualization of digitization that extends beyond investments in traditional IT, and captures the extent to which firms in a sector embed digital technologies within their products, services, and processes. This measure, therefore, encompasses the remote work and collaboration role of digitization.

The results pertaining to the estimation of the impact of AIIE after including $Digitization \times Lockdown$ as an additional explanatory variable in Equation (2) are reported in columns (2) and (5). The coefficient of $A I I E \times L o c k d o w n$ continues to be positive and significant even after accounting for digitization. The coefficient of $Digitization \times Lockdown$ is also positive and significant, indicating that digitization independently contributes to resilience as well. Finally, in columns (3) and (6), we include all three measures in our regression model and find that the coefficient of $AIIE$ $\times$ $Lockdown$ remains positive and significant.¹⁴

Overall, these results suggest that both AI and conventional IT contribute to resilience during the pandemic. Importantly, they are consistent with the proposal that AI-enabled novel forms of business process reorganization that contributes to resilience over and above that facilitated by IT and digitization, and underscore the stability and significance of the effects of AI on business continuity that we document in the previous section.

To further understand the nature of underlying integration of AI with business processes and associated reorganization, we conduct a systematic evaluation of media reports following the lockdown in India. The corresponding analysis also allows us to further differentiate the effects of AI from those that would correspond to traditional IT and digitization. Using Factiva, we downloaded AI-related articles from five major Indian business newspapers (Financial Express, Business Standard, Indian Express, Mint, and The Economic Times) published between March and October 2020. We used the topic filter “Artificial Intelligence/Machine Learning” and use three search keywords “covid”, “pandemic”, and “lockdown.” This results in an initial list of 272 articles. We then screen these articles manually to: (i) identify those that describe a specific application of AI-based technologies, and (ii) relate these applications to a business challenge brought about by the pandemic. We exclude articles that describe applications of AI to mitigate COVID-19-specific healthcare challenges and those that describe duplicate use cases. This process results in a total of 47 use cases from 37 articles. Each use case is additionally coded for whether or not it relates to the three domains of AI application - Demand forecasting and marketing, Production management, and Logistics.¹⁵ In this regard, we rely on recent work by Choi et al. (2022), who describe these as three broad areas in which AI technologies are driving significant reorganization of business processes. See Table EC1.6 in the E-companion for article details, a description of each of these use cases, and the application domains they are categorized in. Two noteworthy observations follow from this analysis.

First, the identified use cases cover applications across a wide range of domains. In terms of the application domains described by Choi et al. (2022), of the 47 use cases, 20 relate to Demand forecasting and marketing, 23 to Production management, and 9 to Logistics.¹⁶ This diversity in application domains suggests that AI builds resilience with regard to a wide range of business functions and helps firms deal with disruptions to both demand and supply side operations during the pandemic. Moreover, as the associated descriptions in Table EC1.6 indicate, the industries these applications are embedded in are similarly diverse and include retail services, financial and insurance services, oil and gas, education, professional services, and public administration, among others. These observations are consistent with the role of AI as a general-purpose technology whose applications transcend business domains and industries (e.g., Bresnahan and Trajtenberg, 1995; Brynjolfsson et al., 2018; Chen and Biswas, 2021; Choi et al., 2022).

Second, the applications documented in Table EC1.6 reflect a multitude of ways in which businesses drew on AI-based technologies to reorganize work processes in each of the three domains described by Choi et al. (2022). For example, applications in the area of demand forecasting and marketing included AI-based virtual agents that automated conversations traditionally handled by insurance agents and enabled contactless claims processing in the insurance industry, the use of ML to forecast in-time stock needed in stores in the retail sector, the use of virtual reality, ML and Deep Tech to provide virtual home tours and site visits to enhance the buying experience and maintain business continuity in the real estate sectors, and advanced computer vision technologies and deep learning models that allowed physical retail companies to provide contactless sizing for their customers. With regard to production management, applications include the use of computer vision technologies to automate car inspections in the insurance sector, the use of advanced audio and video analytics to automate exam proctoring in education or monitor employee productivity for remote systems, and the use of real-time analytics of sensor data in the oil and gas industry toward preventive and remote maintenance. In the area of logistics, AI platforms allow for dynamic route optimization, tracking insights, and automated order management and dispatch across diverse sectors.

The above applications of AI technologies during the pandemic in each of the three domains can be contrasted with those attributable to conventional IT. For example, in the area of marketing, with the widespread implementation of social distancing, companies turned to social media for marketing to and communicating with their customers (Cong et al., 2024). Intelligent sensing played an important role in the area of demand forecasting (Boh et al., 2023; O’Leary, 2020). For example, Boh et al. (2023) document the case of Nike, which faced significant challenges in aligning supply with demand during the pandemic. The company responded with a dynamic supply chain that leveraged demand sensing and real-time inventory visibility using RFID technologies along with data analytics to bring the right products to the right places.¹⁷

In the area of production management, organizations switched to online business models through participation in digital platforms that leveraged demand-side economies of scale to power interactions between different classes of consumers and suppliers (Boh et al., 2023). Diverse types of organizations drew on in-house and cloud services, peer-to-peer information and knowledge sharing systems, and online support service systems to drive the transition to remote work and remote delivery of traditional services (Cong et al., 2024; Rai, 2020). For example, Park et al. (2023) document how centralized IT investments of higher education institutions enabled and supported their transition to emergency remote teaching and maintaining better education services during the crisis. Enterprise systems enabled traditional in-person components of service delivery such as monitoring, assessments, and consultations to be performed virtually with relevant data uploaded from peripheral sites to a centralized integrated database and interpreted remotely (Ting et al., 2020; Vargo et al., 2021).

In the area of logistics, companies used technologies such as adaptive warehouse management systems that dynamically reassigned labor based on real-time demand signals and that in coordination with Internet-of-Things-enabled digital product twins to simulate different operating scenarios under pandemic conditions (Mohapatra et al., 2022).

The above examples emphasize the differential characteristics and roles of traditional IT relative to AI technologies. While the AI applications involved learning, adapting and autonomous decision-making using both structured and unstructured data, the traditional IT applications involved implementation of pre-programmed rules or logic without the learnability or adaptability of systems. They were also used for specific tasks with clear action-outcome linkages and structured data.

In sum, the qualitative analysis allows us to document vivid examples of how exactly AI accorded resilience and sustenance to businesses across various domains and lends further support to the proposition that the role played by AI during this time was distinct from that of IT and digitization.

4.2. Superior Pre-Pandemic Information Capabilities in Firms With High AI Exposure

A rich body of research in management (Cyert et al., 1963; Galbraith, 1973; Mendelson and Pillai, 1998) propounds the view of organizations as information processing systems, and finds that firms’ information management strategies aim to route the right information to the right decision-makers at the right time, and assist them to act on it quickly and effectively. Indeed, prior research in economics and management (Bresnahan et al., 2002; Brynjolfsson and Hitt, 2000; McAfee and Brynjolfsson, 2008) finds that the comparative advantage of information technology, in general, is that it redraws how information is acquired, stored and processed in firms to shape the optimal organization of work. In this regard, it is noteworthy that many of the use cases we document based on our assessment of media reports suggest that AI contributed to resilience via improved efficiency in terms of acquisition and processing of information (e.g., by aiding demand forecasting, monitoring of inventory and production, management of supply chains via enhancements such as real-time route optimization and post-delivery analytics, and provision of customer/stakeholder services such as the execution of digital audits and automated insurance claims handling). It is therefore possible that the buffering impact of AI exposure on business continuity that we document results from superior information management capabilities of firms with greater AI exposure that in turn, enabled them to reorganize business processes in the face of pandemic-induced disruptions.

To assess whether AI exposure relates to firms’ information management capabilities, we use the AI and Future of Work survey, wherein executives responded to the 23-item SML rubric, providing an estimate of AI exposure. In addition, the respondents provided information regarding the organizational changes within the focal firm after the adoption of AI. In terms of application areas that relate to their firms’ information management capabilities, respondents rated the extent to which AI played a role in enabling the following business capabilities within their firms on a five-point scale: linking the organization to customers and suppliers, integrating internal business processes, and providing timely information for executive management to develop strategies. To measure the effect of AI on these variables, we estimate the following specification:

y_{i, f} = μ_{0} + μ_{1} * A I_{i, f} + C o n t r o l s_{i} + C o n t r o l s_{f} + ε_{i, f}

(4)

where

y_{i, f}

is the score reported by respondent

i

working in firm

f

on each of the above mentioned business capabilities. Our primary coefficient of interest (

μ_{1}

) is that of

A I_{i, f}

, which measures the AI score of the firm

f

. We also include individual-level (

C o n t r o l_{i}

) and firm-level (

C o n t r o l_{f}

) controls based on other questions in the survey.

Columns (1)–(3) of Table 5 report results for the estimation of AI-enabled business capabilities. We find that AI exposure relates positively to the strength of linkages with suppliers and customers, greater internal business process integration, and timely sharing of strategic information with top management. That is, exposure to AI is associated with firm characteristics that relate to superior information management. These findings are consistent with the idea that AI exposure facilitated greater resilience during the pandemic because it resulted in superior information management capabilities at the firm level.

Table 5.

AI exposure and firms’ information capabilities.

	Linking to Customers/Suppliers	Integrating Internal Business Processes	Information to Management
	(1)	(2)	(3)
AI	${0.443}^{* * *}$	${0.742}^{* * *}$	${0.822}^{* * *}$
	(0.137)	(0.155)	(0.134)
Observations	301	301	301
R-Squared	0.52	0.49	0.54
Company Controls	Yes	Yes	Yes
Respondent Controls	Yes	Yes	Yes
Sector FE	Yes	Yes	Yes

Notes: All dependent variables are on a 1–5 scale. The variable AI is based on firm-level AI score. Company controls include 3-year revenue, foreign ownership, age, employees, and research and development budget, information technology budget, and marketing budget as a proportion of revenue. Respondent controls include gender, age, current and overall experience, designation, and department. Sector FE are at 2-digit NIC level. Robust standard errors in parentheses. $^{* * *}$ p < 0.01, $^{* *}$ p < 0.05, $^{*}$ p < 0.1.

4.3. The Impact of AI Exposure on the Labor Force

We additionally assess whether AI substituted or complemented labor during the pandemic. Not only does such an assessment facilitate an understanding of the mechanisms through which AI resulted in business continuity during the pandemic, it is also of independent interest and importance because the effects of AI on demand for and value of labor are theoretically ambiguous and inconclusive. Prior research (Frey and Osborne, 2017; Huang and Rust, 2018; Leonhard, 2016; Mfanafuthi et al., 2019) finds that occupational exposure to AI may result in displacement or reduction of human capital within that occupation. Indeed, Frey and Osborne (2017), in an early and detailed assessment of 702 occupations in the US, conclude that 47% of total jobs in the country are at risk of being substituted by AI. In contrast, a rich body of work (Brynjolfsson et al., 2018; Frank et al., 2019; LaGrandeur and Hughes, 2017) concludes that AI may complement human capital to increase productivity and reorganization of work. Brynjolfsson et al. (2018) submit that the latter outcome is more likely since only some tasks in most occupations are suitable for AI and others will continue to require human capital. Given the conflicting evidence, these effects of AI remain open empirical questions.

Our data on worker outcomes are obtained from AmbitionBox, a subsidiary of Naukri.com, India’s largest online job search portal, and include two data sources. The first dataset comprises responses to a survey of workers conducted during and after the lockdown to provide job seekers real-time information on companies and sectors that were actively hiring, salary shifts in their occupation and sector, and components of salaries that were impacted. Over 100,000 workers respond to questions on pecuniary and non-pecuniary benefits and support provided by their employers during the lockdown. Specifically, we estimate shifts in hiring (hiring continuity, offers cancelled), salaries (timeliness, cessation, reduction) and non-pecuniary benefits (employee care programs, work-from-home). These data, however, are available only at the firm-level, i.e., individual responses aggregated to firm-level estimates of fraction of workers reporting the focal outcome. We assess the impact of firm-level AI exposure on these outcomes for a sample of 2,971 firms.¹⁸ Accordingly, we estimate the impact of AI on these objective firm-level outcomes by:

y_{j} = μ_{0} + μ_{1} * A I_{j} + C o n t r o l s_{j} + λ_{s} + ε_{j}

(5)

where

y_{j}

is one of the objective worker outcomes and

A I_{j}

is the firm-level AI exposure score. Table 6 presents the results of this estimation of the effects of AI on objective worker outcomes. All models control for sector fixed effects,

λ_{s}

. In terms of hiring shifts (columns (1)–(3)), the coefficient of AI is positive and significant for continued hiring and negative for hiring freezes and cancellation of offers. It suggests that for all three measures, workers in firms with higher AI exposure experience superior outcomes in terms of changes in hiring-related outcomes post-lockdown. The magnitude of these effects is substantial. For instance, a one-standard-deviation increase in AI increases the proportion of employees reporting continued hiring by 0.14 (or 14%). Similarly, the results presented in columns (4)–(6), which pertain to provision of on-time salary, cessation of salary, and salary cuts, respectively, suggest that salary-related outcomes were also superior for workers in firms with higher AI exposure. Finally, the results in columns (7) and (8), which pertain to non-pecuniary benefits, indicate that workers in firms with higher exposure to AI were also more likely to be able to work from home and to benefit from employee care programs, respectively. It is worth noting that the higher ability of high AI firms to allow for work from home arrangements is also consistent with our proposition that these firms were able to reorganize work to a greater extent in the post-lockdown period than firms with low AI exposure.

Table 6.

Impact of AI on hiring, wages and non-pecuniary benefits.

	Hiring Shifts			Salary Shifts			Non-pecuniary Benefits
	Hire Currently	Hire Freeze	Hire Cancel	Salary On-Time	Salary Cessation	Salary Cut	WFH	Employee Care
	(1)	(2)	(3)	(4)	(5)	(6)	(7)	(8)
AI	${0.14}^{* * *}$	$- {0.11}^{* * *}$	$- {0.02}^{* *}$	${0.14}^{* * *}$	$- {0.10}^{* * *}$	$- {0.16}^{* * *}$	${0.33}^{* * *}$	${0.11}^{* * *}$
	(0.05)	(0.04)	(0.01)	(0.04)	(0.03)	(0.05)	(0.05)	(0.03)
Observations	2592	2592	2592	2592	2592	2592	2592	2592
R-Squared	0.22	0.18	0.12	0.27	0.18	0.27	0.40	0.21
Sector FE	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes

Notes: All regressions are based on firm-level outcomes. AI score is operationalized as firm-level AI exposure and computed using occupations of the respondents within the firm. The regressions control for two firm characteristics: average rating of the firm (sourced from the employee review data described below) and the number of employees. Section EC2 in the E-Companion provides more details regarding these data. WFH, Work From Home. Robust standard errors in parentheses. $^{* * *} p < 0.01$ , $^{* *} p < 0.05$ , $^{*} p < 0.1$ .

We also obtain workers’ perceptions of their firms from the AmbitionBox data. Analogous to Glassdoor.com, AmbitionBox allows employees to anonymously provide ratings, reviews and opinion (text/comment) of their employing firm, the management, and various aspects of work culture.¹⁹ These data comprise 119,060 reviews across 1,624 companies. We analyze four key employee outcomes captured in the reviews - overall satisfaction with the firm, sense of job security, satisfaction with work, and satisfaction with the company culture. Whereas the previous analyses pertaining to objective outcomes are feasible only for the period following the imposition of the nationwide lockdown, these employee reviews pertaining to subjective outcomes are available both pre- and post-lockdown period. We estimate the impact of AI on these subjective worker outcomes using the following specification:

\begin{aligned} y_{j, o t} & = μ_{0} + μ_{1} {AI}_{o} + μ_{2} {Lockdown}_{t} + μ_{3} ({AI}_{o} \times L o c k d o w n_{t}) \\ + {Controls}_{j} + {Controls}_{o} + λ_{f} + ε_{j, o t} \end{aligned}

(6)

where

y_{j, o t}

is one of four subjective outcomes of worker

j

in occupation

o

in period

t

{AI}_{o}

is the AI exposure for occupation

o

. E-Companion Table EC1.7 presents the results of the estimation of the effects of AI on subjective worker outcomes. The coefficient of lockdown is negative and significant for all outcomes, affirming that the lockdown led to a significant decline in subjective employee well-being. However, the coefficient of the interaction term,

AI

\times

Lockdown

, is positive and significant for all four outcomes, suggesting that an increase in exposure to AI improved these outcomes in the post-lockdown period. The aggregate impact on an occupation is the sum of these two coefficients; therefore, the aggregate impact is negative even for the occupations with the highest score for exposure to AI. However, higher exposure to AI relates to lower decline in subjective worker outcomes.

Overall, these results support the conclusion that higher exposure to AI is associated with superior objective and subjective worker outcomes, and that AI complemented, rather than substituted, labor during the lockdown. While prior work indicates that the impact of AI-based technologies on the labor market has so far not been substantial (Acemoglu et al., 2022), it is important to note that we assess these effects over a relatively short period of time. In the long run, as Brynjolfsson and Mitchell (2017) note, the impact of AI on the workforce is likely contingent on a range of factors, including but not limited to complementarities between tasks and occupations, elasticity of labor supply, and price elasticity of demand for different tasks and occupations. Moreover, the impact of AI on subjective worker outcomes in particular may differ substantially even in the short run in other empirical contexts that are not characterized by as much uncertainty and insecurity for workers as the pandemic. In such scenarios, effects of AI on the relationship between workers and technology and the structuring of work (Glikson and Woolley, 2020), may draw greater attention and consideration by workers and thus impact their evaluations more strongly. As AI-based technologies develop in the future and as the world moves definitively to the post-pandemic era, further research is needed to understand AI’s impact on labor.

It is also worth noting that several of the worker outcomes we studied arguably also serve as indicators of business continuity; in particular, objective outcomes such as continued hiring, continued payment of salaries, and provisions for work-from-home arrangements. Given that we observe overall positive effects of AI on these outcomes, these findings constitute further evidence in support of the conclusion that AI exposure was associated with business continuity and resilience during the pandemic.

5. Discussion and Conclusion

The COVID-19 pandemic and related lockdowns to curb its spread resulted in significant disruptions to demand, supply chains, work, and business models of firms. In this paper, we ask if these disruptions were asymmetric across firms—specifically, whether the exposure of occupations and firms to AI helped them systematically mitigate the negative effects of the crisis. We find robust evidence that sectors with high exposure to AI were resilient to lockdown-induced disruptions. Further, the observed benefits of AI extend to other connected sectors within the focal sector’s input–output network. These results remain robust to a range of different specifications, including those that utilize alternative measures of AI and business continuity. Additional findings inform an understanding of the processes that account for the documented positive impact of AI exposure on resilience by highlighting the distinct role of AI in driving the effects we observe, documenting an association between AI exposure and information management capabilities that arguably enabled reorganization of business operations, and showing that AI-enabled reorganization during the pandemic occurred in a manner that complemented labor.

Our study is among the growing set of studies within the domain of operational disruptions and risk management that assess the impact of crises events on firms (e.g., Bakshi and Kleindorfer, 2009; Kleindorfer and Saad, 2005; Mehrotra and Schmidt, 2021; Schmidt and Raman, 2021). The negative impact of operational disruptions on firm value is substantial (Hendricks and Singhal, 2005; Hendricks and Singhal, 2003). The COVID-19 pandemic was unique in terms of severity and the widespread nature of the associated disruptions. Yet, many of the specific challenges it produced for businesses, such as supply chain disruptions, cratering of demand, and the failure of existing business models, were similar to those produced by other disruptions, such as localized natural calamities, political instability, wars, terrorist attacks, and economic slowdowns. Given that we find buffering effects of AI exposure on business continuity within the particularly disruptive context of the COVID-19 pandemic, one can expect that the anticipated widespread adoption of AI technologies in the future is likely to result in a world wherein economic activity is, in general, more resilient to unanticipated disruptions.

Our findings also contribute more generally to the literature on the business value of AI. As noted earlier, the limited research that assesses the economic value of AI (e.g., Babina et al., 2024; Li et al., 2021; Lui et al., 2022; Rock, 2019) has yielded inconclusive evidence, and scholars have called for systematic assessments of the impact of AI on business performance. Our novel empirical context, the aftermath of the COVID-19 pandemic, yields important insights for this literature. The pandemic resulted in a range of supply- and demand-side disruptions, whose impact was heterogeneous across firms and industries. Yet, these disruptions uniformly resulted in the sudden and unanticipated need to rethink and redesign business processes to meet challenges associated with the pandemic. The uniformly positive effects of AI exposure on resilience that we observe within this context attest to AI’s potential to reorganize and transform business processes. At the same time, it is worth noting that there is possibly greater heterogeneity across firms and industries in terms of the potential upside of AI adoption in environments characterized by fewer widespread disruptions than the pandemic. It would therefore be useful for further research to facilitate a nuanced understanding of contextual antecedents of successful AI adoption that can inform managerial decisions regarding AI investments and deployment.

Relatedly, our results point to the possibility that the effects of AI investments on business performance and market reactions are contingent on the value that reorganization of business operations brings within the particular context (Brynjolfsson and Mcafee, 2017; O’Shaughnessy et al., 2023). Indeed, prior work that has documented negative effects of AI on business outcomes doesn’t necessarily consider samples of firms or industries wherein the need or benefits of reorganization of business processes are salient. For example, Lui et al. (2022), in what the authors note to be the first study to measure the value of AI adoption for firms, evaluates the effects of 119 announcements pertaining to AI investments by 62 firms and finds that the mean cumulative abnormal returns for these firms immediately after the announcement were negative and statistically significant ( $- 1.77$ % in magnitude). However, it is important to note that this study was based on a sample of firms that happened to announce AI investments during a specific time period (January 2015–February 2019), and not necessarily those that investors perceived to be poised to benefit from greater flexibility in terms of organizing business operations. Such a misalignment is also likely within the current context wherein there is a hype surrounding AI-based technologies that is unprecedented in terms of magnitude and social costs (Markelius et al., 2024). In fact, it is worth noting that Lui et al. (2022) also find that the negative effects they document are confined to firms with weak (but not strong) IT capabilities. The authors explain these findings by noting that IT capabilities are likely to be a precursor to successful implementation of more advanced AI-based technologies because IT capabilities enhance organizational agility, including operational adjustment agility. Thus, the authors’ findings appear consistent with the possibility that it was the absence of a perceived upside in terms of benefits of restructuring business processes via the adoption of AI that contributed towards the overall negative effects of AI investments on market value that they document.

Our findings pertaining to the impact of neighboring sectors’ AI exposure on a focal sector’s resilience suggest that the value of AI investments extends also to other sectors within the sector’s input–output network. These findings reflect that the resilience of neighboring sectors was necessary and important over and above a sector’s exposure to AI in terms of maintaining continuity during the pandemic. They are also consistent with prior work that relates to the value of adoption of technological innovations to the extent to which they align with customer capabilities (Afuah, 2004; McElheran, 2015). These findings suggest that the marginal returns to investments in AI are likely to increase as AI adoption increases. This entails the presence of positive externalities in the context of AI adoption, as has been noted in the context of other general purpose technologies (Bresnahan and Trajtenberg, 1995), and therefore, highlights the critical role that policymakers could play in aligning incentives to adopt AI with the value it brings for industries and economies.

Although our primary focus has been to assess the impact of AI on resilience during the pandemic, reports of a spike in AI adoption during this time indicate that the pandemic is likely to have shaped the evolution of AI as a mainstream technology. This premise is consistent with a substantial body of research that shows that diffusion of new technologies and innovations is a path-dependent process wherein early patterns of adoption and usage influence subsequent adoption and development of technologies (e.g., Aghion et al., 2014; Greve and Seidel, 2015; Karanam et al., 2024). The increased adoption of AI post-pandemic possibly led to phenomena that further accelerated its adoption, such as the development of regional and industrial hubs associated with generation of AI-related knowledge and increased transmission of such knowledge (Dahlke et al., 2024). At the same time, owing to the specific set of challenges associated with the pandemic, it likely also resulted in a qualitative shift in terms of the specific applications and domains within which development of AI occurred that are not necessarily optimized to economic and social needs in the long term. Such a possibility relates to interesting avenues for future research. For example, it would be worth examining whether greater integration of AI with business processes during the pandemic translated to a sustained source of value for certain classes of firms, AI technologies, or domains of reorganization. Such assessments are likely to guide policy that would support effective integration of emerging technologies with business processes during crises events and in their aftermath. Another important direction for future research would be to extend our analysis to the firm level in order to better understand the impact of AI investments on individual firm performance, as our current industry-level analyses limit causal insights into firm-specific heterogeneity that contributes to organizational resilience.

Footnotes

Acknowledgments

The authors gratefully acknowledge the support of the Office of the Chief Economic Advisor to the Government of India, Dr. Krishnamurthy Subramanian, and the National Payments Corporation of India for providing access to payments data. They thank Prof. Subodha Kumar, the Senior Editor, and the three anonymous reviewers for their valuable and constructive feedback, which significantly improved this paper. They also thank Aditya Singhai and Aadhil M Shaik for excellent research assistance. Most importantly, this work is dedicated to the memory of the late Nikhil Madan, whose insights and writing were instrumental in shaping this article. An earlier version of this paper was circulated as “Suitability for Machine Learning and Immunity to the COVID-19 Pandemic.”

Declaration of Conflicting Interests

The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The authors received no financial support for the research, authorship, and/or publication of this article.

ORCID iDs

Abhishek Bhatia

Deepa Mani

Shekhar Tomar

Supplemental Material

Supplemental material for this article is available online (doi: ).

Notes

How to cite this article

Bhatia A, Madan N, Mani D and Tomar S (2026) AI and Immunity to the COVID-19 Pandemic. Production and Operations Management 35(5): 1611–1629.

References

Acemoglu

Autor

Hazell

, et al. (2022) Artificial intelligence and jobs: Evidence from online vacancies. Journal of Labor Economics 40(S1): S293–S340.

Acemoglu

Carvalho

Ozdaglar

, et al. (2012) The network origins of aggregate fluctuations. Econometrica 80(5): 1977–2016.

Afuah

(2004) Does a focal firm’s technology entry timing depend on the impact of the technology on co-opetitors? Research Policy 33(8): 1231–1246.

Agca

Babich

Birge

, et al. (2022) Credit shock propagation along supply chains: Evidence from the CDS market. Management Science 68(9): 6506–6538.

Agca

Birge

Wang

, et al. (2023) The impact of COVID-19 on supply chain credit risk. Production and Operations Management 32(12): 4088–4113.

Aghion

Hepburn

Teytelboym

, et al. (2014) Path dependence, innovation and the economics of climate change. In: Handbook on Green Growth. Edward Elgar Publishing, Inc.

Babina

Fedyk

, et al. (2024) Artificial intelligence, firm growth, and product innovation. Journal of Financial Economics 151: 103745.

Bakshi

Kleindorfer

(2009) Co-opetition and investment for supply-chain resilience. Production and Operations Management 18(6): 583–603.

Ban

Rudin

(2019) The big data newsvendor: Practical insights from machine learning. Operations Research 67(1): 90–108.

10.

Berente

Recker

, et al. (2021) Managing artificial intelligence. MIS Quarterly 45(3): 1433–1450.

11.

Beyer

Jain

Sinha

(2023) Lights out? COVID-19 containment policies and economic activity. Journal of Asian Economics 85: 101589.

12.

Boh

Constantinides

Padmanabhan

, et al. (2023) Building digital resilience against major shocks. MIS Quarterly 47(1): 343–360.

13.

Bresnahan

Brynjolfsson

Hitt

(2002) Information technology, workplace organization, and the demand for skilled labor: Firm-level evidence. The Quarterly Journal of Economics 117(1): 339–376.

14.

Bresnahan

Trajtenberg

(1995) General purpose technologies ‘engines of growth’? Journal of Econometrics 65(1): 83–108.

15.

Brynjolfsson

Hitt

(2000) Beyond computation: Information technology, organizational transformation and business performance. Journal of Economic Perspectives 14(4): 23–48.

16.

Brynjolfsson

Mcafee

(2017) Artificial intelligence, for real. Harvard Business Review 1: 1–31.

17.

Brynjolfsson

Mitchell

(2017) What can machine learning do? Workforce implications. Science (New York, N.Y.) 358(6370): 1530–1534.

18.

Brynjolfsson

Mitchell

Rock

(2018) What can machines learn, and what does it mean for occupations and the economy? AEA Papers and Proceedings 108: 43–47.

19.

Bughin

Seong

Manyika

, et al. (2018) Notes from the AI frontier: Modeling the impact of AI on the world economy. McKinsey Global Institute 4.

20.

Chang

Pierson

Koh

, et al. (2021) Mobility network models of COVID-19 explain inequities and inform reopening. Nature 589(7840): 82–87.

21.

Chen

Biswas

(2021) Turning crisis into opportunities: How a firm can enrich its business operations using artificial intelligence and big data during COVID-19. Sustainability 13(22): 12656.

22.

Choi

Kumar

Yue

, et al. (2022) Disruptive technologies and operations management in the industry 4.0 era and beyond. Production and Operations Management 31(1): 9–31.

23.

Cohen

(2018) Big data and service operations. Production and Operations Management 27(9): 1709–1723.

24.

Cong

Yang

Zhang

(2024) Small and medium enterprises amidst the pandemic and reopening: Digital edge and transformation. Management Science 70(7): 4564–4582.

25.

Cui

Zhang

(2022) AI and procurement. Manufacturing & Service Operations Management 24(2): 691–706.

26.

Cyert

March

, et al. (1963) A behavioral theory of the firm. Englewood Cliffs, NJ 2(4): 169–187.

27.

Dahlke

Beck

Kinne

, et al. (2024) Epidemic effects in the diffusion of emerging digital technologies: Evidence from artificial intelligence adoption. Research Policy 53(2): 104917.

28.

Ding

Levine

Lin

, et al. (2021) Corporate immunity to the COVID-19 pandemic. Journal of Financial Economics 141(2): 802–830.

29.

Dohale

Akarte

Gunasekaran

, et al. (2022) Exploring the role of artificial intelligence in building production resilience: Learnings from the COVID-19 pandemic. International Journal of Production Research 62(15): 5472–5488.

30.

Dohmen

Merrick

Saunders

, et al. (2023) When preemptive risk mitigation is insufficient: The effectiveness of continuity and resilience techniques during COVID-19. Production and Operations Management 32(5): 1529–1549.

31.

Dong

Tomlin

(2012) Managing disruption risk: The interplay between operations and insurance. Management Science 58(10): 1898–1915.

32.

Ellis

Rao

Raju

, et al. (2018) RFID tag performance: Linking the laboratory to the field through unsupervised learning. Production and Operations Management 27(10): 1834–1848.

33.

Felten

Raj

Seamans

(2021) Occupational, industry, and geographic exposure to artificial intelligence: A novel dataset and its potential uses. Strategic Management Journal 42(12): 2195–2217.

34.

Frank

Autor

Bessen

, et al. (2019) Toward understanding the impact of artificial intelligence on labor. Proceedings of the National Academy of Sciences 116(14): 6531–6539.

35.

Frey

Osborne

(2017) The future of employment: How susceptible are jobs to computerisation? Technological Forecasting and Social Change 114: 254–280.

36.

Galbraith

(1973) Designing Complex Organizations. Reading, MA: Addison-Wesley Pub. Co.

37.

Geva

Saar-Tsechansky

(2021) Who is a better decision maker? Data-driven expert ranking under unobserved quality. Production and Operations Management 30(1): 127–144.

38.

Glikson

Woolley

(2020) Human trust in artificial intelligence: Review of empirical research. Academy of Management Annals 14(2): 627–660.

39.

Graves

Tomlin

Willems

(2022) Supply chain challenges in the post-COVID era. Production and Operations Management 31(12): 4319–4332.

40.

Greve

Seidel

(2015) The thin red line between success and failure: Path dependence in the diffusion of innovative production technologies. Strategic Management Journal 36(4): 475–496.

41.

Gursoy

Chi

, et al. (2019) Consumers acceptance of artificially intelligent (AI) device use in service delivery. International Journal of Information Management 49: 157–169.

42.

Hale

Angrist

Kira

, et al. (2020) Variation in government responses to covid-19. Blavatnik School of Government Working Paper, Version 6.0 (BSG-WP-2020/032).

43.

Helo

Hao

(2022) Artificial intelligence in operations management and supply chain management: An exploratory case study. Production Planning & Control 33(16): 1573–1590.

44.

Hendricks

Singhal

(2003) The effect of supply chain glitches on shareholder wealth. Journal of Operations Management 21(5): 501–522.

45.

Hendricks

Singhal

(2005) An empirical analysis of the effect of supply chain disruptions on long-run stock price performance and equity risk of the firm. Production and Operations Management 14(1): 35–52.

46.

Henfridsson

Bygstad

(2013) The generative mechanisms of digital infrastructure evolution. MIS Quarterly 37(3): 907–931.

47.

Huang

Rust

(2018) Artificial intelligence in service. Journal of Service Research 21(2): 155–172.

48.

Karanam

Mani

Saha

(2024) Growing technological relatedness to the ICT industry and its impacts. Information Systems Research 36(1): 344–369.

49.

Kleindorfer

Saad

(2005) Managing disruption risk in supply chains. Production and Operations Management 14: 53–68.

50.

Kleindorfer

Belke

Elliot

, et al. (2003) Accident epidemiology and the U.S. chemical industry: Accident history and worst-case data from RMP* Info. Risk Analysis 23(5): 865–881.

51.

LaGrandeur

Hughes

(2017) Introduction: An Overview of Emerging Technology and Employment in the Early Twenty-First Century. Palgrave Macmillan, Cham.

52.

Lau

RYK

Zhang

(2018) Parallel aspect-oriented sentiment analysis for sales forecasting with big data. Production and Operations Management 27(10): 1775–1794.

53.

Leonhard

(2016) Technology Vs. Humanity: The Coming Clash Between Man and Machine. FutureScapes Lodge, UK.

54.

Sethi

(2021) Does CSR reduce idiosyncratic risk? Roles of operational efficiency and AI innovation. Production and Operations Management 30(7): 2027–2045.

55.

Lui

Lee

Ngai

(2022) Impact of artificial intelligence investment on firm value. Annals of Operations Research 308(1): 1–16.

56.

Markelius

Wright

Kuiper

, et al. (2024) The mechanisms of AI hype and its planetary and social costs. AI and Ethics 4(3): 727–742.

57.

McAfee

Brynjolfsson

(2008) Investing in the IT that makes a competitive difference. Harvard Business Review 86(7–8): 98.

58.

McElheran

(2015) Do market leaders lead in business process innovation? The case(s) of e-business adoption. Management Science 61(6): 1197–1216.

59.

McKendrick

(2021) AI adoption skyrocketed over the last 18 months. Harvard Business Review. https://hbr.org/2021/09/ai-adoptionskyrocketed-over-the-last-18-months (accessed November 18, 2024).

60.

Mehrotra

Schmidt

(2021) The value of supply chain disruption duration information. Production and Operations Management 30(9): 3015–3035.

61.

Mendelson

Pillai

(1998) Clockspeed and informational response: Evidence from the information technology industry. Information Systems Research 9(4): 415–433.

62.

Mfanafuthi

Nyawo

Mashau

(2019) Analysis of the impact of artificial intelligence and robotics on human labour. Gender and Behaviour 17(3): 13877–13891.

63.

Mithas

Chen

Saldanha

, et al. (2022) How will artificial intelligence and industry 4.0 emerging technologies transform operations management?. Production and Operations Management 31(12): 4475–4487.

64.

Mohapatra

Tripathy

Singhal

, et al. (2022) Significance of digital technology in manufacturing sectors: Examination of key factors during COVID-19. Research in Transportation Economics 93: 101134.

65.

Nenova

Shang

(2022) Chronic disease progression prediction: Leveraging case-based reasoning and big data analytics. Production and Operations Management 31(1): 259–280.

66.

O’Leary

(2020) Evolving information systems and technology research issues for covid-19 and other pandemics. Journal of Organizational Computing and Electronic Commerce 30(1): 1–8.

67.

O’Shaughnessy

Schiff

Varshney

, et al. (2023) What governs attitudes toward artificial intelligence adoption and governance?. Science and Public Policy 50(2): 161–176.

68.

Park

Son

Angst

(2023) The value of centralized IT in building resilience during crises: Evidence from U.S. higher education’s transition to emergency remote teaching. MIS Quarterly 47(1): 451–482.

69.

Pournader

Kach

Talluri

(2020) A review of the existing and emerging topics in the supply chain risk management literature. Decision Sciences 51(4): 867–919.

70.

PwC. (2020) AI: An opportunity amidst a crisis. Technical report, PwC.

71.

Queenan

Cameron

Snell

, et al. (2019) Patient heal thyself: Reducing hospital readmissions with technology-enabled continuity of care and patient activation. Production and Operations Management 28(11): 2841–2853.

72.

Rahmati

Tafti

Westland

, et al. (2020) When all products are digital: Complexity and intangible value in the ecosystem of digitizing firms. MIS Quarterly 45(3): 1025–1058.

73.

Rai

(2020) Editor’s comments: The COVID-19 pandemic: Building resilience with is research. Management Information Systems Quarterly 44(2): iii–vii.

74.

Raj

Seamans

(2019) Primer on artificial intelligence and robotics. Journal of Organization Design 8(1): 1–14.

75.

Reinhart

(2020) This time truly is different. Project Syndicate 23(3).

76.

Rock

(2019) Engineering value: The returns to technological talent and investments in artificial intelligence. Available at SSRN 3427412.

77.

Schmidt

Raman

(2021) Operational disruptions, firm risk, and control systems. Manufacturing & Service Operations Management 24(1): 411–429.

78.

Sharifi

Khavarian-Garmsir

Kummitha

RKR

(2021) Contributions of smart city solutions and technologies to resilience against the COVID-19 pandemic: A literature review. Sustainability 13(14): 8018.

79.

Swaminathan

(2018) Big data analytics for rapid, impactful, sustained, and efficient (rise) humanitarian operations. Production and Operations Management 27(9): 1696–1700.

80.

Ting

DSW

Carin

Dzau

, et al. (2020) Digital technology and COVID-19. Nature Medicine 26(4): 459–461.

81.

Tomlin

Wang

(2011) Operational strategies for managingsupply chain disruption risk. In Kouvelis P, Dong L, Boya-batli O and Li R (eds) Handbook of Integrated Risk Managementin Global Supply Chains. Wiley, 79–101.

82.

Tomlin

Wang

(2011) Operational Strategies for Managing Supply Chain Disruption Risk. John Wiley & Sons.

83.

Vargo

Zhu

Benwell

, et al. (2021) Digital technology use during COVID-19 pandemic: A rapid review. Human Behavior and Emerging Technologies 3(1): 13–24.

84.

Venkatesh

Pal

(2018) Compilation of an input–output table and social accounting matrix for India: 2012-13. Center for Study of Science.

85.

Webb

(2019) The impact of artificial intelligence on the labor market. Available at SSRN 3482150.

86.

Wooldridge

(1996) Introductory Econometrics: A Modern Approach. 2nd ed.

87.

Sethi

Chung

, et al. (2023) Reforming global supply chain management under pandemics: The GREAT-3Rs framework. Production and Operations Management 32(2): 524–546.

88.

Xue

Ray

Sambamurthy

(2012) Efficiency or innovation: How do industry environments moderate the effects of firms’ IT asset portfolios?. MIS Quarterly 36(2): 509–528.

89.

Yang

Feng

Whinston

(2022) Dynamic pricing and information disclosure for fresh produce: An artificial intelligence approach. Production and Operations Management 31(1): 155–171.

90.

Yoganarasimhan

(2020) Search personalization using machine learning. Management Science 66(3): 1045–1070.

91.

Yoo

Henfridsson

Lyytinen

(2010) Research commentary—The new organizing logic of digital innovation: An agenda for information systems research. Information Systems Research 21(4): 724–735.

92.

Zhu

Ninh

Zhao

, et al. (2021) Demand forecasting with supply-chain information and machine learning: Evidence in the pharmaceutical industry. Production and Operations Management 30(9): 3231–3252.

Supplementary Material

Please find the following supplemental material available below.

For Open Access articles published under a Creative Commons License, all supplemental material carries the same license as the article it is associated with.

For non-Open Access articles published, all supplemental material carries a non-exclusive license, and permission requests for re-use of supplemental material or any part of supplemental material shall be sent directly to the copyright owner as specified in the copyright notice associated with the article.

0.51 MB

0.00 MB