What does the academic evidence say about whether AI investment improves firm performance? Compare supporting and conflicting findings, explain the conditions under which the relationship changes, and cite the key studies behind each conclusion.

1. Overall conclusion

The academic evidence generally supports a positive relationship between AI investment or adoption and firm performance, but the effect is conditional rather than automatic. Positive findings are strongest when performance is measured at the firm level, when AI is integrated into business processes, and when firms possess complementary capabilities such as skilled employees, data infrastructure, managerial support, and coherent technology strategies. The evidence is considerably weaker when AI is measured indirectly through patents or textual “AI focus,” when investment has not yet matured, or when organizations lack the capacity to reorganize around the technology.

The apparent disagreement across studies therefore reflects differences in what is being measured. “AI investment” may refer to patents, AI-related labor, managerial attention, adoption, or intensity of use; “performance” may refer to productivity, sales, profitability, market value, operational efficiency, employment, or resilience. These measures do not capture the same stage of the value-creation process.

2. Evidence supporting performance improvements

Firm-level productivity studies provide the clearest support. Using German firm data and instrumental-variable estimators, Czarnitzki, Fernández, and Rammer [1] find that both AI adoption and the intensity of AI use are positively associated with productivity. Evidence from European firms is also positive: AI innovation is followed by productivity gains of approximately 6.2–17% in an event-study framework, although the estimated premium is smaller—about 2.1–6%—in a distance-to-frontier specification [2]. This difference suggests that the size of the effect depends partly on the comparison group and the empirical design.

The strongest recent estimates come from studies that identify adoption directly rather than inferring it from patents. Pastor-Merino et al. [3] identify AI adoption among 62,525 Spanish firms using web data and large language models, then use entropy-balanced instrumental-variable estimations. They report that adopters have approximately 53–55% higher sales and 48% higher value added than comparable non-adopters; firms using AI intensively across multiple functions show larger premiums, including up to 74–76% higher sales and 67% higher value added. These estimates should not be treated as universal treatment effects, but they provide evidence that deeper, cross-functional integration is associated with greater performance gains.

Other studies identify improvements in specific dimensions of performance. AI-related patenting has an additional positive effect on labor productivity, particularly among small and medium-sized firms and service firms, where rapid adjustment may be easier [4]. AI-driven decision-making is positively related to firm performance, with big-data-powered AI contributing to the development of AI capabilities and decision quality [5]. AI adoption is also associated with stronger operational performance in manufacturing, although the magnitude of the effect depends on strategic orientation [6]. In marketing, AI use is linked to profitability, customer satisfaction, and customer acquisition, with customer outcomes acting as important mediating mechanisms [7].

The benefits are not restricted to ordinary productivity periods. Using natural-disaster shocks, Han et al. [8] find that AI investment can improve corporate resilience: a firm with AI-related skills in approximately 2.4% of its jobs could, on average, recover the valuation damage associated with disasters over a short event window. This finding indicates that AI may create value through flexibility and shock absorption, not only through routine cost reduction.

3. Conflicting and weaker findings

Several studies challenge the claim that AI investment produces an immediate or general performance premium. Parteka and Kordalska [9] find that AI innovation has a negligible role in officially recorded productivity growth at the macro level, describing this as evidence of a “modern productivity paradox.” This result does not necessarily contradict positive firm-level estimates: national productivity statistics may lag behind firm experimentation, may dilute gains concentrated in a subset of adopters, and may fail to capture organizational investment required to realize AI benefits.

Measurement also matters. AI patents indicate technological invention, not necessarily successful operational deployment. By contrast, adoption measures capture implementation, while intensity measures capture the extent to which AI is embedded in processes. The more favorable findings tend to use adoption or process integration, whereas patent-based measures produce more heterogeneous results. A lagged analysis of AI and machine-learning patent intensity finds positive effects on return on assets and operating margins, particularly after approximately five years, but limited effects on net profit margin [10]. This suggests that AI may improve productive capacity before it produces durable bottom-line profitability.

The evidence is not uniformly positive even at the firm level. Kazakis [11] finds no direct association between labor-based AI investment and firm efficiency in the baseline model. Efficiency gains appear only when AI investment is combined with capable managers, competitive pressure, stable institutional ownership, and access to long-term debt. Similarly, the evidence on AI implementation is positive for financial performance and market value but not uniformly positive across first movers and better-performing firms [12]. These findings indicate that selection into AI investment and the capacity to finance its implementation are central to the observed relationship.

There are also potential costs. Knesl [13] shows that firms with a high share of displaceable labor are negatively exposed to automation shocks; following such shocks, employment and profitability decline, especially in highly competitive industries. Hudson and Morgan [14] likewise find that AI exposure can increase idiosyncratic risk under particular combinations of high-technology industry conditions and board-network structures. Thus, AI can raise productivity while simultaneously increasing adjustment costs, employment risk, competitive pressure, or earnings volatility.

4. Conditions that change the relationship

The central mechanism is therefore complementary investment rather than AI spending alone. Lee et al. [15] find that revenue gains appear only beyond a sufficient level of AI-adoption intensity and are larger when firms also invest in complementary technologies and internal R&D. This explains why simple comparisons between adopters and non-adopters can be misleading: some adopters are merely experimenting, whereas others have redesigned processes around AI.

Organizational capability is equally important. Evidence from resilience research shows that firms may benefit more from AI in principle while realizing less of that potential when they lack complementary organizational designs [8]. Studies of investment efficiency and ESG performance point to related mechanisms: AI improves information transparency, internal control, financing conditions, innovation, and external oversight, which can convert technological investment into broader firm-level gains [18][19]. These mechanisms imply that the performance effect is mediated by what AI enables the organization to do, rather than by the technology as an isolated asset.

Finally, the relationship depends on whether AI complements or substitutes for labor and existing capabilities. Evidence from Italian firms indicates that AI combined with advanced information technologies is more likely to support human–machine complementarity, whereas AI combined with robotics is more likely to reinforce labor substitution; only the former combination is associated with stronger investment in AI-related training [20]. AI investment is therefore most likely to improve sustainable firm performance when it expands employees’ capabilities, improves coordination, and supports better decisions. Where it primarily substitutes for labor in highly competitive markets, gains may accrue through cost reduction while employment, profitability, or organizational stability deteriorate.

Taken together, the literature supports a conditional productivity and performance premium, not a universal AI premium. The most defensible conclusion is that AI investment improves firm performance when it is sufficiently intensive, integrated with complementary technologies and skills, supported by capable management, and evaluated over a time horizon long enough for organizational change to occur. The relationship becomes weak, delayed, or negative when AI is adopted symbolically, measured only through innovation proxies, introduced without complementary capabilities, or deployed mainly as a labor-substitution technology under intense competitive pressure.