Ask both whether a genuine improvement was measured and how its benefits are distributed.
Imagine a small business finishes the same amount of work in fewer hours after introducing AI. That is a promising operational result. What happens next is a different question. The owner might reduce prices, take on more customers, improve the service, shorten working hours, increase wages, or retain the savings. Several outcomes could happen together. Saying that “AI creates wealth” does not explain who receives it, who pays the transition costs, or whether the initial saving survives quality checks. Understanding those choices is essential to assessing AI’s effect on people rather than merely celebrating a faster workflow.
Productivity and pay measure different things
Productivity concerns output relative to inputs. Pay concerns what a worker receives. A firm’s revenue, its profit, a worker’s income, and a customer’s price are also distinct measures. A change in one can influence the others without producing a one-for-one transfer.
Consider an illustrative consultancy that saves four staff hours on a report. Those hours are not automatically four extra billable hours or a wage increase. There may be insufficient customer demand, fixed-price contracts, additional review costs, or another bottleneck. The example is hypothetical; its purpose is to identify which assumptions a claim about gains leaves unstated.
What the evidence can and cannot establish
The 2025 customer-support study by Brynjolfsson, Li and Raymond found larger benefits for less experienced and lower-skilled agents. That provides evidence that a particular assistant can narrow performance gaps within a workplace. It does not establish how economy-wide wages or wealth will be distributed.
An OECD paper published in 2024 examined 19 countries using data from 2014–2018. It found no indication of an effect on wage inequality between occupations in that period, alongside some evidence of lower inequality within occupations. These observations predate the recent generative AI wave; they cannot settle its eventual distributional effects.
Follow four possible channels
AI Affect’s framework is to follow the potential gain through four channels. This is a way to ask better questions, rather than a forecast of which group will win.
- Workers: Does the tool increase the value of their contribution? Do they share in better results through pay, training, autonomy, or working time?
- Customers: Does competition produce lower prices or better service, and can customers detect a decline in quality?
- Businesses: Can the firm turn saved effort into useful output, and does additional revenue exceed implementation and operating costs?
- Suppliers: How much of the benefit goes to model providers, computing infrastructure, software vendors, or other necessary inputs?
Looking at all four prevents a single company’s success story from becoming a claim about everyone’s living standards. A product can make a supplier valuable while delivering modest savings to some customers. A cheap tool can also help a small firm compete more effectively. Both are possibilities to investigate.
Questions to ask about a real rollout
Ask for the baseline, the period measured, and the quality standard. Check whether the claimed savings include staff training, integration, oversight, and correcting failures. Ask whether “adoption” means an experiment or regular operational use. For employees, ask how the organization plans to use any capacity released.
At the policy level, the IMF’s January 2024 staff discussion note describes distribution as dependent on how AI complements labor and affects capital returns. These are conditional economic mechanisms, not a guaranteed path for any country.
A useful test of an AI success story therefore has two parts: was a genuine improvement measured, and where did the benefit go? Keeping those questions separate makes the debate more accountable and gives workers, customers, and owners a clearer basis for negotiation.