The easiest AI costs to count may become the least important part of the story. SemiAnalysis gives the better name to the harder problem: dark output, the economic value created by AI that remains invisible or badly distorted in GDP, price measures, labor data, and industry accounts.
A company can see its data-center bill, software subscription, token spend, and headcount changes. It may fail to see the research brief that took five minutes instead of five hours, the sales proposal created before a call that would otherwise have gone unmade, or the internal analysis that replaces a paid consultant without creating a comparable transaction.
This problem starts with the way the economy gets counted. The Bureau of Economic Analysis describes the service economy through national income and product accounts built from production, income, spending, investment, and price measures. That framework works reasonably well when goods and services move through observable markets at observable prices. AI scrambles that logic when a task that used to cost hundreds of dollars becomes an internal prompt, a model call, or a workflow step buried inside a broader software bill.
The issue deserves attention because AI use has moved beyond a small experimental fringe. The U.S. Census Bureau found that AI adoption among businesses hovered between 17% and 20% from December 2025 through May 2026, with higher rates in information and finance and insurance. Larger firms use it more, which matters because large firms also have more administrative, analytical, software, and compliance work that can vanish inside internal workflows before official statistics can see the output.
Executives already face a measurement trap. McKinsey’s 2025 global survey found that nearly nine in ten respondents said their organizations regularly use AI, while nearly two-thirds had not begun scaling AI across the enterprise and only 39% reported enterprise-level EBIT impact. That gap between broad usage and narrow financial visibility explains why leaders can understate AI returns when the benefits show up as speed, quality, decision readiness, or avoided outside spending rather than clean revenue growth.
Usage data points in the same direction. Anthropic researchers analyzed more than four million Claude conversations and found that AI use concentrated heavily in software development and writing, while roughly 36% of occupations used AI for at least a quarter of their associated tasks. They also found that 57% of use looked like augmentation and 43% looked like automation. That means AI output often appears as help embedded in human work rather than a separate product with a separate invoice.
Real productivity gains can still exist inside that fog. A Quarterly Journal of Economics study of 5,172 customer-support agents found that access to a generative AI assistant increased issues resolved per hour by 15% on average, with larger gains for less experienced and lower-skilled workers. That kind of AI productivity can improve throughput and customer experience while remaining hard to separate from staffing choices, training effects, demand shifts, and quality changes.
The opposite risk also exists: AI can look productive while slowing work down. METR’s randomized trial with experienced open-source developers found that early-2025 AI tools increased task completion time by 19%, even though developers expected the tools to reduce completion time and later believed they had helped. That finding exposes an AI measurement weakness inside many companies: self-reported time savings often substitute for audited workflow outcomes.
The token bill will tempt executives and economists into false precision. A 2026 paper on token spend argues that tokens have become a practical accounting unit for foundation-model services, yet token expenditure and economic value remain distinct because value depends on the workflow, marginal productivity, risk, and downstream effects. One thousand tokens can produce useless text, a better customer reply, a flawed legal summary, or a strategic insight. The unit of computation tells us little unless we know the business outcome it changed.
Benchmarks help but cannot solve the accounting problem alone. The GDPval benchmark evaluates economically valuable tasks across 44 occupations in the top nine U.S. GDP sectors and reports that frontier models are approaching industry experts in deliverable quality on some tasks when paired with oversight. That gives leaders a better picture of capability, yet capability does not equal realized value. A model may pass a task test and still fail inside a messy approval process, regulated environment, or incentive system that discourages employees from changing how work gets done.
Open measurement efforts can improve the situation. One 2026 paper proposes an open source economic index using public user-LLM chat data and O*NET tasks to measure both AI adoption and capability across occupations. That type of task-level evidence matters because the unit of change in AI work usually comes below the job, below the department, and below the line item. The job title may stay fixed while the work inside it changes dramatically.
A newer form of the productivity paradox may follow. A 2026 survey of Korean workers found that 51.8% used generative AI for work and that AI reduced working time by 3.8%, while the correlation between time savings and output changes was near zero. The authors argue that workers often captured gains as on-the-job leisure rather than increased output. That productivity paradox does not make AI useless. It shows that the destination of saved time matters as much as the saved time itself.
Business leaders should stop asking whether AI creates value in the abstract. They should ask where the value lands. Does it cut purchased services? Does it increase output? Does it reduce cycle time? Does it raise quality? Does it help weaker performers close skill gaps? Does it free managers for higher-value work? Does it create new work that nobody previously bought because the human version cost too much? Each answer needs a different metric.
That means AI governance should include a measurement layer, not just a risk layer. Teams should document the task, baseline time, review burden, output quality, rework rate, failure cost, human oversight, and business result before declaring success. Leaders should also watch for invisible work losses: junior employees losing learning opportunities, experts spending more time reviewing low-quality AI output, and employees using AI privately because the official tools feel too slow or restrictive.
At the policy level, GDP will remain essential, but AI will pressure it in exactly the places where services already strain measurement. At the firm level, the better response starts with disciplined workplace AI adoption: redesign workflows, measure outcomes at the task level, and treat human behavior as central to the production system. AI value will increasingly appear before the accounting system can classify it. Leaders who wait for perfect aggregate proof will move late, while leaders who measure the invisible work inside their own organizations will see the real ledger sooner.
