Business AI usage in the US reached 17-20% between December 2025 and May 2026. While AI impacts white-collar work, many firms report no change in staffing needs. Employment among 22–24-year-olds fell by 12% in AI-exposed industries, as generic tasks get automated. Firms must now focus on protecting entry-level roles through training.
AI adoption is relatively slow and uneven, constrained by transition costs, organizational inertia, the need to redesign workflows and the difficulty of training systems on proprietary and incumbent data. In the US, business AI usage was only about 17-20% between December 2025 and May 2026. These early observations should, therefore, not be mistaken for AI’s final effects. Nonetheless, they offer some useful stylized facts.
First, AI is exposing cognitive and white-collar work more than earlier automation, which focused largely on routine physical tasks. Generative AI is particularly effective at language, coding, information processing and analysis.
The International Labour Organization estimates that one in four workers globally is in an occupation exposed to generative AI, rising to 34% in high-income countries. This does not mean white-collar jobs will disappear, but that the boundary between ‘skilled’ and ‘automatable’ work is shifting.
Second, the risk is greatest at the entry level. AI can perform the generic tasks through which young workers traditionally enter firms and gain experience. US Census research found that employment among 22–24-year-olds fell by about 12% in the most AI-exposed industries over the 10 quarters after ChatGPT’s introduction, largely because hiring weakened.
This makes firm-specific knowledge increasingly valuable in an AI-dominated ecosystem. A senior manager who understands a firm’s customers, systems and risks is harder to replace than a junior programmer writing standard code. The same applies to junior analysts, accountants and researchers. The more generic the task, the easier it is to automate; the more firm-specific the knowledge, the more AI can complement the human worker.
This brings us to the third stylized fact: firm-specific knowledge is becoming an important source of protection against displacement. So far, there is little evidence that widespread AI adoption has produced economy-wide mass unemployment.
An Organisation for Economic Co-operation and Development survey found that 83% of small and medium firms (SMEs) using generative AI reported no change in their overall staffing needs. Only 9% reported a decline while 6% reported an increase. At the same time, 65% of AI-using SMEs reported that AI improved employee performance while 20% said it increased their demand for skills.
Therefore, deployment matters. AI can replace workers doing generic tasks while enabling professionals and entrepreneurs to do far more. Productivity gains can create new markets, demand and jobs. The policy challenge is to capture AI’s productivity gains while keeping its net employment impact broadly neutral.
One way of doing this is to pursue a ‘net-zero labour loss’ strategy. The basic principle would mirror the logic of net-zero carbon policy. Firms adopting AI would bear a measurable labour-displacement obligation, while firms that generate employment would get labour credits that could be traded to offset that obligation.
To achieve net zero, the first step would be to establish a standardized labour-equivalent token measure. In simple terms, we would need to estimate how many AI tokens, or perhaps units of AI compute, are, on average, equivalent to replacing one full-time worker performing a standardizable task per year.
The next step is to tax each AI company with a reasonable rate of labour losses based on the number of labour-equivalent tokens it sells. The precise rate, of course, needs to be calibrated empirically against observed employment displacement.
Further, we need to differentiate between AI that substitutes workers and AI that makes them more productive. A company that uses AI to expand production and create additional jobs should be able to generate labour equivalent token credits. These credits could be sold to AI companies or other firms with an outstanding labour-displacement obligation. However, if sufficient credits are unavailable, that obligation could be settled through a labour-transition levy.
The government could use the proceeds to finance retraining, apprenticeships and skill development. Similarly, firms would need to report their token consumption and the corresponding labour displacement at a specified rate, as well as their net labour displacement. This would incentivize firms to use AI to raise productivity and create new jobs. It would also encourage firms to move workers along the AI value chain, ensuring net-zero labour loss rather than displacing them outright.
Therefore, the objective would be to make firms internalize part of the social cost of displacement while rewarding those that use AI to expand employment. There is no doubt that the smarter response to the risk of AI-driven labour displacement would be to design a system that manages it without stifling innovation. Regulating the AI industry to protect existing jobs could create costly distortions. A principle of net-zero labour displacement could help strike a balance.