Not Everywhere, Not Always

Incumbent firms across R&D-intensive industries have invested heavily in AI, with the expectation to increase innovation performance by accelerating development cycles, reducing costs, and enabling new knowledge recombination- Yet, observed innovation performance effects remain modest and uneven. Prior research provides limited guidance for resolving this puzzle. Existing studies tend to (i) treat AI’s impact as homogeneous across all innovation stages, (ii) focus predominantly on efficiency gains, and (iii) fall short in explaining why the anticipated impact of AI on innovation performance fails to materialize in practice. As a consequence, managers face persistent uncertainty about where, when, and how AI meaningfully contributes to innovation outcomes.

To address these blind spots, following study examines the scope and boundary conditions of AI’s impact on innovation performance from a process-level perspective. It draws on a revelatory singlecase study of a large European-US pharmaceutical incumbent with high R&D intensity and advanced AI readiness. The analysis focuses on AI use across the early drug discovery process, a strategically critical activity that both drives innovation performance, attracts substantial AI investment and comprises five sequential stages: target identification, target validation, hit discovery, hit-to-lead, and lead optimization. The findings show that AI’s impact on innovation performance is highly uneven and strongly conditional on stage-specific characteristics.

In target identification, AI improves efficiency and, in some cases, quality, yet only when data availability and mechanistic understanding are strong. In target validation, AI plays a minimal role, despite this stage accounting for a substantial share of early-stage project failure. In hit discovery and hit-to-lead, AI accelerates search and iteration, but performance gains remain localized and primarily efficiency-based. In lead optimization, AI supports multiparameter optimization, yet improvements in compound quality depend heavily on predictive accuracy. Because failures at this stage often stem from pharmacokinetic, pharmacodynamic, and toxicity limited quality gains constrain downstream success. Overall, AI improves local process efficiency but does not generate proportional improvements in innovation success or quality across the full process. Performance effects are asymmetric, stagedependent, and fragile, rather than cumulative.  The study contributes to innovation research by demonstrating that AI does not function as a broadly enabling innovation capability. Instead, its strategic value is contingent on task structure, data regimes, and the alignment between algorithmic strengths and process-critical decision points. These findings help explain why AI adoption by incumbents does not automatically translate into sustained performance effects: efficiency gains at isolated stages are insufficient to overcome systemic uncertainty in innovation processes. For managers, the results underscore the need for selective AI deployment aligned with specific stages rather than firm-wide adoption strategies.

‍