RDP 2023-10: Adoption of Emerging Digital General-purpose Technologies: Determinants and Effects 5. Adoption and Firm Profitability

5.1 Empirical framework

We employ a panel event study framework to estimate the effect of GPT adoption on firm-level outcomes. The framework allows us to analyse changes in firm outcomes before and after the adoption of GPT, accounting for the fact that firms adopt GPT at different points in time which can create issues around overlapping treatment periods in difference-in-difference models (Goodman-Bacon 2019). The approach is also similar in nature to the dynamic analysis undertaken in Babina et al (2024) for AI adoption.

The variable Adopt_i indicates the period when the technology was first referenced by firm i. The outcome of interest is denoted as y_it, and the panel event study specification is as depicted in Equation (2).

where:

The adoption event's lags and leads are defined as binary variables signifying that a specific firm was a given number of periods away from the adoption event. The coefficients of interest are the betas related to the lags and leads. We focus on up to four years before adoption and three years after.^[8] While a longer post-adoption window could be appropriate if these investments have very long payoff windows, we are constrained by the sample period available.

We control for firm fixed effects in ${\mu }_i$ and industry-time fixed effects ${\theta }_{s,t}$. The former allows us to control for time-invariant unobserved heterogeneity at the firm level, while the latter accounts for industry-level patterns in adoption, demand conditions, tax incentives and reporting standards. We also control for firm size, measured as log assets, given the earlier findings around its importance for adoption. We cluster standard errors at the firm level to account for correlation of outcomes within firms over time.

For the analysis, our reference or base period, relative to which all other results are interpreted, is two years before adoption. We choose this rather than the identified year of adoption as there is some evidence that firms may be investing or changing their behaviour ahead of adoption.

This approach allows us to consider how outcomes change pre- and post-adoption. If there is no trend in outcomes beforehand, it will provide some evidence that there are no other factors driving both future profitability and the decision to adopt, which would make our results invalid. That said, our results should still be thought of as correlations, rather than as strictly causal. Ultimately firms are selecting into adoption (treatment), rather than it being allocated randomly, and they may do so in expectation of some future outcome.

5.2 Aggregate results

We start our analysis focusing on non-IT firms over the full sample period. This includes any adoption event occurring from 2013 to 2022, and any balance sheet data back to 2009 to account for the pre-adoption behaviour of firms. Profitability generally remains broadly stable following adoption, showing neither statistically nor economically significant differences compared to the pre-adoption period (Figure 3).

There is a large literature documenting that the diffusion of GPT tends to be slow because of the need to create complementary inventions, or make organisation changes, that facilitate profitable adoption of the GPT (e.g., Agrawal et al 2023a). In the case of cloud computing, this could reflect the developments in cloud-based programs or cloud services. For AI/ML, this could include the development of simple machine learning models and packages.

As such, it is worth examining whether profitability evolves differently for early and late adopters.^[9] Figure 4 shows that there is a significant downturn in profitability in the years following the adoption event for early adopters, defined as those adopting from 2013 to 2016. Conversely, profitability of late adopters (adopting from 2017 to 2022) seems to remain broadly stable post-adoption. To test whether the difference in post-adoption outcomes for early and late adopters is significant we adopt a staggered difference-in-difference-in-difference model. We do find a marginally significant difference in outcomes, suggesting that late adopters had better post-adoption outcomes than early adopters (Appendix F).

Importantly, there is no significant trend in outcomes in the pre-adoption period across all samples, providing some evidence that the findings do not reflect differing prospects and trajectories for the firms. That said, there is some volatility in profitability in the years pre-adoption. However, the results are robust to changing the exact period definitions, for example by lumping all pre-and post-adoption periods together (Appendix E).

It is also possible that there are some other differences between early and late adopters that are driving the differences in their outcomes. One is that changes in accounting standards could potentially play a role. From 2019 to 2020 there was a clarification in accounting standards indicating that most cloud computing expenses should be expensed, though a small share could be recorded as investment (Department of Finance 2022). This is likely to have led more firms to record GPT-related outlays as expenses, which should weigh on their profitability and work against our findings. More generally, the adoption of cloud technology could lead firms to shift from capital expenditure with subsequent depreciation expenses, to operating expenses, making it look like firms' profitability is falling post-adoption when profitability is measured using EBITDA (as we do).

We take two simple approaches to considering these issues. The first is to deduct capital purchases from profitability. The second is to account for depreciation expenses by measuring profitability as EBIT (earnings before interest and tax). The results remain broadly unchanged, though the measure including capital expenditure is noisier (Appendix H).^[10]

One further concern might be the inclusion of the COVID-19 pandemic in the late adopters sample. It may be that those firms that had adopted cloud or AI/ML technologies were better able to navigate the COVID-19 shock, either because they were better managed or because the technologies helped navigate the shock (consistent with the large spike in adoption). As a simple test, we ended our sample in 2020. Doing so leads to, if anything, stronger evidence that late adopters increase profitability following post-adoption in terms of the magnitude of the effect, though the response is not significant (Appendix I).^[11]

Taken together these results suggest that profitable adoption has become easier over time, consistent with further developments in these technologies.^[12] If this is the case, it suggests that there may be more scope for effective adoption of these technologies going forward, which could improve productivity. That said, there still appears to be a large degree of variation in outcomes, as reflected in the wide error bands. This suggests other factors may dictate whether adoption is profitable, which we explore below.

5.3 The role of Board of Directors

In this section, we delve into the characteristics of the Board of Directors that correlate with the profitable adoption of technologies. We conduct the same event study looking at profitability post-adoption, but split the sample based on whether the Board has members with certain characteristics or not. We focus on late adopters (post-2016), given this is when the bulk of adoption occurs. Focusing on the most recent adoptions also should help to limit issues around the cross-sectional snapshot of directors becoming less accurate over time.

Focusing first on whether the Board has some members with a background in GPT, there is evidence of an increase in profitability for adopting firms whose current Board has some degree of GPT expertise, at least temporarily (Figure 5). This is not the case for boards with no members with GPT experience. The difference in outcomes is statistically significant (Appendix F) and robust to period definitions (Appendix E). The difference in outcomes appears to be driven by a reduction in operating expenses for firms with a GPT-experienced Board, while revenue remains stable post-adoption for all firms (Appendix D). This contrasts to Babina et al (2024) who find that AI adoption leads to increased sales, but relatively few cost savings. The difference may reflect that the adoption of cloud technologies is far more frequent in our sample and appears to drive much of our results.^[13]

These results provide some evidence that boards with relevant experience may help profitable adoptions to GPT. Such a finding is somewhat consistent with evidence from Alekseeva et al (2021) and Calvino et al (2022) that having managers with AI skills is highly valued in terms of wage premia or firm productivity, respectively.

One concern with this finding may be that Board members add mentions of GPT to their biographies if their company profitably adopted the technology. However, spot checks of the biography data suggest that this is not a key driver, and that we are mainly picking up the directors' educational background and previous work experience.

We also observe substantial variation in profitability based on the gender composition of the Board. Figure 6 indicates that there is a statistically significant (if temporary) increase in return on assets for firms with a female Board member. However, there is no evidence of any change in profitability among late adopters with no female representation on the Board of Directors. The difference in post-adoption outcomes across the groups is statistically significant (Appendix F). This finding aligns with and bolsters earlier findings that gender diversity in senior management roles is associated with stronger firm performance (Gordini and Rancati 2017; EmadEldeen et al 2021) and dynamic capabilities (Wilson et al 2023). In terms of magnitude, the increase in profitability for firms with female representation is far smaller than that for those with GPT skills, suggesting the latter is likely a more important driver of company outcomes.

Since there is a strong positive correlation between firm size and female representation on the Board, the increase in profitability post-adoption for firms with a female director may be driven by firm size.^[14] To check for this, we first conduct the event study separately for larger firms (in the 4th and 5th quintiles of assets) and smaller firms (in the bottom three quintiles of assets). Profitability is unchanged post-adoption for larger firms, whereas it tends to increase for smaller firms though not significantly so (Appendix G). Given the prevalence of female Board members on larger firm boards, this would work against the above findings. To consider this further, we split firms by a combination of firm size and female representation and find that smaller firms with a female director appear to drive the increased profitability (though the increase is not significant, likely reflecting the small sample size). Taken together, this suggests that it is female representation (or some other correlated factor), not size, that is driving the result.^[15]

None of the other Board characteristics were associated with differing profitability post-adoption. This is somewhat surprising given the strong correlation between adoption and, for example, having someone with an IT background on the Board. In part, it may reflect the use of a snapshot of the Board and their skills and experience. Future work could explore the question using longitudinal board information.

Footnotes

We restrict our attention to the four periods prior to adoption only as there is high volatility and noise in the more distant history. [8]

Early adopters tend to be smaller but have similar profitability pre-adoption to late adopters (see Table B1). They are also more likely to be in certain sectors, as shown in Figure 2. [9]

Another approach considered was to use coarsened exact matching, to match our firms across several observables. We considered this approach, and while some of the patterns were similar, the small sample size meant that the results were very volatile and imprecisely estimated. Moreover, given the pre-trends are quite flat in our main regressions, and given the main concerns around causality likely arise around selection on unobservables, we have not pursued this approach further. [10]

The difference in outcomes between firms with and without GPT skills on their board is also evident, though the former no longer shows a significant increase. Removing mining firms, which may have very different technology needs, also does not change the results (Figure I1). [11]

We can't rule out the possibility that the nature of the early and late adoptions differed substantially. For example, early adoptions may have been larger investments, more experimental and more disruptive. They may also have had positive spillovers or greater upside risk in terms of skills formation, learning by doing or pushing the knowledge frontier forwards. Nevertheless, the evidence suggests that more recent adoptions were easier to turn into profit (or at least to avoid a loss), even if this simply reflected the potentially less ambitious nature of these adoptions. [12]

That said, many cloud technologies have in-built AI/ML features. For example, Loucks (2018) estimates that around 70 per cent of firms that adopt AI capabilities will do so via cloud-based enterprise software. As such, we should be cautious in examining the technologies separately. [13]

While we control for size in our baseline regressions, we do not allow the effect of adoption to vary by size. [14]

Sample size of across subgroups is broadly comparable, so the results do not simply reflect very small sample sizes in some sub-samples of the data. [15]