What CEOs are seeing — and what the numbers miss

At the surface level, the claim is straightforward: big AI spending, lots of pilots, fewer headline-grabbing layoffs than expected, and no clear spike in productivity statistics. But beneath that simplicity lie multiple explanations that are often conflated.

• Measurement lag and mismatches. Traditional productivity metrics are blunt instruments. GDP and labor productivity measures were built for an industrial era when output was physical and easy to count. AI often improves quality, reduces response times, or enables new business models — gains that can take years to diffuse or to be captured by price indexes.
• Diffusion and complementary investments. AI rarely delivers full value on its own. Models need labeled data, integrations, process redesign, retraining, and governance. Without complementary capital — better data pipelines, change management, and updated workflows — an AI model can sit as a pilot forever.
• Compositional effects in employment. Rather than broad-scale replacement, AI changes task composition. Some roles shrink, others expand, and many jobs are reconfigured. That can mean no big change in headline employment rates even as work looks very different inside organizations.
• Pilots, safety, and regulatory caution. Risk teams and regulators rightly slow down rollouts for systems that affect safety, fairness, or privacy. Those checks lower the immediate velocity of adoption and thus delay measurable gains.

Why this moment echoes Solow

Solow’s paradox wasn’t just an observation about delayed benefits. It highlighted a deeper point about general-purpose technologies: the payoff is rarely instantaneous or evenly distributed. Steam, electricity, and computing each required waves of complementary investments before productivity rose sharply. AI is following a similar arc.

Three dynamics stand out:

1. Time for re-architecting operations. Bringing AI into production isn’t a plug-and-play process. It demands new data foundations, monitoring, retraining cadences, and human-in-the-loop systems. Companies are still learning how to redesign workflows so AI amplifies, rather than obstructs, human labor.
2. Invisible gains precede visible output. Early benefits often look like reduced cognitive load, fewer repetitive tasks, and small time savings aggregated across many workers. Those gains are real but diffuse, and accounting systems don’t always translate them into higher productivity numbers quickly.
3. Misaligned incentives and misallocation. Not all AI dollars buy the same thing. A flood of funding into flashy prototypes, rather than durable infrastructure, can create a mirage of progress without durable gains. Allocation of capital and attention matters as much as the models themselves.

Sectoral and scale differences

AI’s impact is uneven. In industries where work is standardized and data-rich, automation and augmentation can show measurable gains sooner. Manufacturing and some parts of financial services often realize returns faster because processes are already instrumented and outputs measurable. In labor-intensive, service-oriented sectors with high human judgment, such as healthcare or education, value accrues in subtler ways.

Large organizations also face a different challenge than nimble startups. Scale brings potential for big payoff, but it also amplifies frictions: legacy systems, procurement cycles, and cultural barriers. Smaller players can sometimes iterate faster, but lack the scale to make widespread productivity claims.

Signs that the tide is turning

All is not static. Several leading indicators suggest that when the right pieces come together, productivity gains can accelerate:

• Platformization of AI. As robust model-serving platforms, observability tooling, and data contracts become widely adopted, the cost of going from prototype to production falls.
• Standardized interfaces and APIs. Common standards make it easier to swap in models and to embed them into products without custom engineering for every use case.
• Task-level metrics. Organizations shifting attention to task completion rates, cycle times, and error reduction are finding performance wins that aggregate into larger productivity improvements.

What to do next — for builders, managers, and the curious community

Reviving Solow’s paradox is not an indictment of AI; it’s a call to action. The pathway to broad productivity growth isn’t simply better models — it’s thoughtful systems design and patience. Practical steps that accelerate impact include:

• Measure differently. Supplement GDP-style metrics with task-level, outcome-based KPIs. Track time saved, error reductions, customer satisfaction, and new revenue streams emerging from AI-enabled products.
• Invest in complements. Data engineering, user experience, change management, and continuous monitoring are the scaffolding that allow AI to scale its benefits.
• Run wide, not just deep. Instead of sinking enormous budgets into one monolithic initiative, spread smaller experiments across workflows, learn quickly, and scale the successes.
• Foster interoperability. Prioritize modular systems and open contracts so investments don’t lock organizations into brittle architectures.
• Report transparently. Share both successes and the friction points. A richer public record of what works will speed diffusion and reduce repeated mistakes.

Why patience is not complacency

Optimism about AI’s long-term potential is compatible with realism about the present. The historical arc of transformative technologies suggests several years — sometimes decades — of organizational and institutional change before broad productivity gains appear. That interim is not wasted: it is when durable foundations are laid.

Think of today’s investments as invisible wiring. The visible effects may be subtle now — faster approvals, fewer clerical errors, enhanced creativity in niche teams — but together they build the circuits that will carry larger flows of value in the years ahead.