Artificial intelligence is producing real gains for some workers, particularly when it helps people complete bounded tasks, find information, or apply the practices of stronger colleagues. PwC’s 2026 AI Jobs Barometer is the latest bullish signal: companies in the most AI-exposed sectors recorded productivity growth about 40% higher than the least exposed group. But that is a comparison of revenue per employee across groups of firms, not proof that AI alone caused the difference. The rewards are also concentrated. Reuters’ reporting on Meta’s Project OT offers the counterexample: internal agents and coding tools generated more activity but brought reliability and security problems and did not produce the hoped-for productivity lift. The macroeconomic data point to promise, not a settled AI boom. Productivity is becoming more uneven, more dependent on workflow design, and more demanding of human judgment.
The latest AI productivity headline is encouraging, not conclusive
For a technology that is supposed to reshape the economy, AI has generated an awkwardly mixed record. There are workers who can now draft a first-pass customer reply, summarize a long document, write a routine bit of code, or search a large internal knowledge base much faster than they could a few years ago. There are also organizations paying for model access, data infrastructure, security reviews, and training without a clear line from the new tool to a better business result.
PwC’s 2026 Global AI Jobs Barometer makes the optimistic case forcefully. Its headline finding is that productivity growth at the most AI-exposed companies was 40% higher than at the least exposed companies. In the underlying comparison, the most exposed group recorded 33.5% growth in productivity from a 2018 baseline, versus 24.0% for the least exposed group. PwC defines productivity here as turnover, or revenue, per employee, using Orbis company data.
PwC’s more interesting finding may be how uneven the results are. The top fifth of companies within the most AI-exposed group achieved average productivity growth of 163% from the same baseline, while a separate PwC performance study cited in the report found 20% of companies claiming 74% of AI-driven gains. Only 8% of CEOs in the Barometer’s cited survey said AI had generated more than a slight rise in revenue over the past year. The lesson is not that the tools are useless. It is that a small number of firms have learned how to turn them into commercial results, while many others are still experimenting.
A faster task is not automatically a more productive company
Productivity is often used as a synonym for speed. Economists use it more carefully: it is the amount of output produced for a given amount of input. The Bureau of Labor Statistics tracks output relative to labor, capital, energy, materials, and services. It also notes that total factor productivity includes effects that cannot be cleanly measured, such as technology, efficiency, scale, and resource reallocation.
That distinction sounds fussy until an AI rollout goes wrong. A programmer can produce more code while a software team produces fewer useful features. A customer-service worker can close chats faster while customers return with unresolved problems. A lawyer can receive a rapid summary while spending extra time checking its citations. In each case, an AI tool may improve a visible subtask and still fail to improve the final output after quality control, rework, risk, and coordination are counted.
The most persuasive evidence for AI is therefore often narrow. A well-known field study of 5,179 customer-support agents found that access to a generative AI conversational assistant increased issues resolved per hour by 14% on average. The effect was 34% for novice and lower-skilled workers and minimal for experienced, highly skilled workers. The researchers found evidence that the tool helped newer workers adopt the practices of their stronger colleagues.
The difference is especially important with so-called agents. A chatbot gives a person text to evaluate. An agent is designed to take actions, such as changing a record, calling an internal service, or creating an application, with less step-by-step instruction. That can remove bottlenecks. It also gives a faulty instruction, a bad inference, or an insecure permission more opportunities to do damage before a person catches it.
Meta showed why the denominator matters
The sharpest recent corrective to the productivity narrative came from Reuters’ investigation into Meta’s Project OT, short for Organization Transformation. The project imagined an “AI native” organization in which agents would take over much of the daily work of thousands of employees, with smaller groups of humans supervising. In scenario planning, executives considered reducing some teams by as much as 60%, Reuters reported.
The project was not simply a thought experiment. Meta reorganized teams into smaller pods and used AI tools in product and engineering work. But Reuters reported that Meta cancelled planning for a second restructuring wave after internal data suggested autonomous agent technology was not delivering the expected productivity gains. Meta confirmed Project OT existed, described it as a year-long scenario-planning effort involving cost cutting, team redesign, and redeployment, and said it did not proceed with every scenario.
The operational details are more revealing than the headline. Reuters reported that AI-assisted work led to a 220% year-over-year increase in changes to internal software platforms and infrastructure. Changes that resulted in new or upgraded features reaching users were up 36%. That gap is not an indictment of every coding assistant. It is a reminder that code volume is an input to a product process, not the finished product.
The costs appeared elsewhere too. Internal posts reviewed by Reuters warned of reliability problems from the AI coding surge and of agents taking large, disruptive actions. Major technical and security incidents, including service disruptions and possible data leaks, rose 40% year over year, while employee time spent firefighting them rose 70%, according to the reporting. Meta declined to comment on those internal disruption figures. In a July town hall, Zuckerberg said agentic technology had not accelerated as quickly as he had hoped, though he expected improvement within three to six months.
This does not prove that agents cannot deliver productivity. It proves something more useful: a company cannot infer productivity from activity, headcount plans, or a tool demo. An organization has to include the people who verify outputs, resolve incidents, protect data, retrain colleagues, and repair workflows. Leave those costs outside the calculation and almost any automation can look miraculous for a while.
Company results and national results answer different questions
The PwC and Meta examples describe companies. They cannot tell us whether AI has lifted productivity across the whole economy. National productivity statistics are slower, noisier, and more stringent. They must include the early adopters, the firms that never adopted, the failed experiments, and industries where AI has little immediate use.
The aggregate picture is promising but still thin. The Federal Reserve Bank of Kansas City found that industries with higher reported AI adoption tended to have faster productivity growth between late 2023 and mid-2025. But the relationship explained only a small share of the aggregate improvement. The authors put it plainly: the pattern is suggestive rather than definitive, because causality can run both ways and other sector forces may move with adoption.
The San Francisco Fed has made a related point: most macro studies still find limited evidence of a large AI effect, even though case studies show time and cost savings. General-purpose technologies do not transform an economy merely because they exist. Firms have to reorganize work around them, which is usually slower and more expensive than buying the technology.
The work is changing before the productivity verdict is final
One thing can be true before the economy-wide data settle: AI is changing what employers ask of people. PwC found that the skills in the most AI-exposed jobs are changing more than twice as fast as in the least exposed jobs. New tasks added to AI-exposed roles were 2.5 times more likely to depend on empathy, judgment, and creativity. Its U.S. analysis also found that highly exposed entry-level roles were seven times more likely to request skills traditionally associated with senior workers, such as leadership and strategic thinking.
That distinction matters for any employer, including an AI Recruiting Company, trying to distinguish a genuinely redesigned role from one that has merely had a few tasks automated. The durable premium is unlikely to belong simply to the person who can type a prompt. It is more likely to belong to the worker who knows when an output is wrong, which information should not leave the organization, and what decision remains too consequential to delegate.
What the evidence says
AI is making some workers more productive now. The strongest evidence is in well-defined tasks with useful data, clear feedback, and a human who remains responsible for the outcome. It can help less experienced workers climb a learning curve faster, and the best-performing companies appear to be finding ways to combine it with growth rather than treating it only as a headcount lever.
AI is not yet proven to be a broad, automatic productivity machine. PwC’s 40% figure is an important company-level correlation, not a clean causal estimate for the global economy. Meta’s Project OT shows why the missing parts of the calculation matter: more output from a model can coexist with more verification, more incidents, and less value delivered to customers. Economy-wide measures point to an early, uneven effect.
The nuanced answer, then, is not that AI either works or fails. It works when a firm can turn quicker machine output into better human decisions, reliable operations, and useful final products. When it cannot, the technology may simply move the work around, often to the people tasked with catching its mistakes.
