The AI Productivity Gap No One Is Closing

Despite heavy investment and experimentation with artificial intelligence, only 37% of companies report any earnings impact—and that share has not moved in a year. The persistence of this gap suggests challenges in translating technical advances into measurable business results, requiring different leadership, measurement and implementation strategies.

By Where Should I Learn Editorial · 2026-09-22

When companies tally the benefits of artificial intelligence, the numbers are surprisingly small: only 37% report any observable earnings impact, and that share has not budged in a year. That single statistic now stands as a stark indicator that technical adoption alone is not enough to deliver financial returns. It raises a question that should preoccupy executives and educators alike: why does the route from pilot to profit remain so fraught?

The first answer lies in the distinction between adoption and integration. Organizations may acquire platforms, hire specialized talent and run proofs of concept, yet those activities do not automatically change the routines, incentives and processes that generate revenue. When AI tools exist alongside legacy workflows instead of being woven into them, their outputs can become curiosities rather than drivers of financial performance. The unchanged 37% figure suggests many companies are still in that latter category.

Measurement practices also blunt claims of impact. Earnings are the hardest form of evidence to produce: they require not only a successful model but also sufficient scale and time to affect top-line or bottom-line metrics. Firms that report operational improvements—faster processing, fewer errors, or higher customer satisfaction—may still see no detectable change in earnings within reporting periods. The persistence of the statistic across a year implies that such measurement gaps remain widespread.

Another factor is selection bias in projects. Early AI initiatives often target technically interesting problems rather than business-critical ones. When investments flow to exploratory work or to infrastructure that yields indirect benefits, the link to immediate earnings weakens. For the 63% of companies not seeing earnings effects, it is plausible that their portfolios contain too many low-leverage experiments and not enough efforts tightly aligned with revenue or cost levers.

Organizational capacity is a further constraint. Delivering measurable earnings impact requires cross-functional teams that can translate model outputs into commercial decisions—marketing campaigns, pricing moves, inventory choices, or automation of billable work. Where such coordination is absent, AI outputs may be technically sound but operationally unusable. The static percentage over a year suggests many organizations have yet to develop the sustained processes needed to convert outputs into outcomes.

Data and infrastructure weaknesses also help explain the gap. Producing a model is one task; generating clean, integrated datasets that can feed production systems every day is another. Companies that have not invested in the plumbing of data—governance, lineage, pipelines, and monitoring—will struggle to scale pilots into revenue-driving services. A flat earnings-impact metric hints that those foundational investments remain incomplete in many places.

Timing and cost structure matter as well. Some AI projects require significant upfront expenditure and pay off only after a period of scale. When earnings cycles are quarterly or annual, such payoffs may fall outside the window that companies use to judge success. The fact that the proportion of companies reporting earnings impact has not moved in a year could reflect a mismatch between project timelines and conventional financial reporting periods.

Regulatory, ethical and customer-trust considerations introduce additional friction. Even when a model could enhance a product or cut costs, businesses may hesitate to deploy it broadly because of compliance risk, reputational concerns or uncertain customer acceptance. Those deployment delays can keep potential earnings benefits latent, visible internally as capability but invisible on the income statement. The static 37% figure therefore also signals caution in the fielding of high-impact systems.

Closing the gap requires a different set of priorities than many current AI programs emphasize. First, measurement must be explicit: pilot success criteria should include concrete financial KPIs and clear hypotheses about how model outputs will move them. Second, portfolio choices should favor initiatives with direct links to revenue or controllable costs, rather than exploratory research that lacks a commercialization path. Third, organizations need to build the operational interfaces—interfaces between models and business decisions—that allow outputs to trigger pricing, replenishment, service routing or other commercial actions.

Leadership and skills must adapt in tandem. Executives should insist on accountable owners for the decisions an AI system will affect, not just technical leads for model development. Training should focus less on abstract model accuracy and more on decision design: how a prediction will be interpreted, who will act on it, and how acting will change outcomes. The unchanged earnings-impact figure shows that technical proficiency without decision-oriented implementation is insufficient.

If the 37% statistic is a warning, it is also a useful one. It reframes the conversation from whether AI can be built to how it can be embedded so that its effects show up where boards care: earnings. Achieving that will not be quick or easy; it will require changes in governance, measurement, skill development and project selection. But treating the statistic as a prompt rather than as a verdict will help companies move beyond the first wave of experimentation toward the harder work of operationalizing AI for measurable financial returns.