AI Productivity Gains May Lag in ‘Weak-Link’ Workflows Where Tasks Depend on Each Other

As artificial intelligence spreads from call centers to coding and logistics, the economic debate often focuses on a familiar promise: automate enough tasks and productivity will surge. Yet a growing body of research suggests the payoff can be slower and more uneven when work is organized as a chain of tightly connected tasks—what economists call a “weak links” production structure, where each step depends on the others.

In these settings, tasks behave as complements rather than substitutes. That means improving one task does not automatically translate into large gains in overall output if another task remains a bottleneck. A factory can install vision systems to spot defects at high speed, for example, but if downstream handling, compliance checks, or maintenance scheduling remain slow or error-prone, the line still runs at the pace of its weakest stage. Studies that model production as a set of interdependent tasks show that when substitution across tasks is low—sometimes represented by very small elasticities—automation in a single area can deliver surprisingly modest aggregate improvements.

This helps explain why firms can report striking successes with AI in specific functions while national productivity statistics move more gradually. Traditional AI, built on rule-based systems and machine learning trained for narrow tasks, has long improved precision and efficiency in domains like manufacturing inspection, credit scoring, and scheduling. More recently, generative AI has added a flexible layer: drafting text, generating code, producing designs, and interacting with customers in natural language. But even when AI executes its piece quickly and accurately, the end-to-end performance may remain constrained by complementary inputs—human approvals, data quality, organizational processes, or regulatory requirements—that the model does not remove.

Economists emphasize that complementarity changes investment incentives. If returns to automating one task depend on upgrading several others, firms may delay adoption until they can deploy AI and process changes in bundles. The result is a diffusion pattern that looks less like an immediate revolution and more like a series of staggered upgrades—especially in highly regulated sectors such as finance and healthcare, where transparency, accountability, and documentation can become binding constraints. Researchers studying technology adoption have long observed that “general purpose technologies” often require complementary innovations—new workflows, skills, and management practices—before their full benefits show up. AI appears to be following that script in many workplaces.

Complementarity has a darker implication as well: it can increase vulnerability. In a weak-links system, an attacker does not need to compromise everything; disrupting the most fragile task can degrade the whole output. Research exploring these models finds that when tasks are hard to substitute for one another—meaning production cannot easily route around a failure—even small shocks can have outsized effects. In practical terms, that can mean that AI-enabled systems, while efficient, may also create clearer targets. If a firm relies on an automated triage model for customer requests, poisoning the data feeding that model, knocking it offline, or manipulating its outputs can bottleneck the entire service operation. The same applies to AI-augmented supply chains: a single compromised forecasting or routing component can cascade into missed deliveries and idle inventory.

Generative AI adds a further twist because it can lower the cost of carrying out certain attacks. Research has highlighted how AI can scale convincing phishing messages, automate social engineering, and accelerate the discovery of system vulnerabilities. When production is complementary, the economic damage from these tactics can rise: an attack that delays just one critical stage—identity verification, payments processing, or dispatch—can halt the rest. In other words, the productivity gains from AI may arrive slowly in weak-links environments, but the potential for disruption can arrive quickly.

Labor markets, too, reflect this mix of gradual benefits and heightened stakes. Forecasts synthesized in international studies suggest substantial churn by 2030, with many roles disrupted by technology while even more new positions emerge. Complementarity helps reconcile these forces: automation may replace or compress certain tasks, but it can also raise the value of the remaining tasks that AI cannot easily substitute—oversight, judgment, compliance, and relationship-heavy work. That, in turn, increases demand for new roles to build, monitor, and govern AI systems, including engineers, data specialists, and ethics and risk professionals.

For policymakers and corporate leaders, the lesson is not simply to “adopt AI,” but to understand the system around it. Where production resembles a chain, the economic returns depend on strengthening multiple links: data pipelines, cybersecurity, employee training, and resilient fallback procedures. Otherwise, AI can make one segment world-class while leaving the overall process only marginally better—and potentially easier to knock off balance.

Why it matters

AI is often treated as a direct substitute for human work, but in many real organizations it functions as one component in a tightly coupled system. When tasks are complements, productivity gains depend on coordinated upgrades, while the same structure can amplify the harm from errors or attacks. Understanding where the bottlenecks sit may determine whether AI becomes a slow-burning efficiency tool—or a new source of systemic risk.