Who Gets the Gains from AI: Workers, Customers or Shareholders?

If AI makes a task faster, who gets the saved time? A worker might finish earlier, a customer might receive a cheaper service, or an employer might expect more output from the same staff. The technology does not decide between those outcomes.

That makes AI adoption a distribution question as well as a technical one. Workers need to know what counts as a gain, what new work appears and who can influence the decision about sharing the benefits.

Exposure does not mean a job disappears

The ILO’s 2025 assessment of generative AI and jobs estimates that one in four workers globally is in an occupation with some degree of exposure. It stresses transformation as the more likely broad outcome than wholesale redundancy, because human input remains necessary.

These are estimates of potential exposure, not counts of layoffs caused by AI. A task that software can assist with is not the same thing as a complete occupation that can be safely automated.

Count the work that remains

Suppose a fictional support team spends 100 hours drafting responses. An AI tool cuts drafting to 60 hours, but review, correction and handling exceptions add 20 hours. The net saving is 20 hours, not 40. Subscription, integration and training costs also belong in the evaluation.

If employees spend the saved hours fixing mistakes that the dashboard does not count, the apparent gain may be exaggerated. If the quality genuinely improves, measuring only speed may understate the benefit. Decide the measurement before attaching rewards or staffing cuts to it.

Four ways to share a real gain

These uses can be combined. A company may need to recover investment costs, but that does not settle how later benefits should be distributed. Nor should employees be expected to absorb all disruption while every upside remains discretionary.

Put numbers on the proposed split

Return to the fictional team’s 20-hour net saving. Here are three possible agreements, measured in hours per comparable work period:

All three rows use the same saving. Only the agreement determines who benefits. The figures assume the saved hours can actually be rescheduled; scattered minutes across shifts may not translate into an afternoon off. Track whether the promised time was delivered, alongside quality and errors.

What to put in an adoption agreement

Ask for a baseline, a trial period and a way to report errors without being penalised for undermining a productivity target. Establish who reviews sensitive outputs and how workers receive paid time to learn the system.

Then ask what changes if the tool works: pay, hours, staffing, responsibilities or performance expectations. “We will become more efficient” is a goal; it is not a worker benefit until the terms explain how employees share in it.

Research the supplier and the employer separately

Review Microsoft’s company research and Alphabet’s company score when examining major technology suppliers. Those scores cannot tell you whether your own employer has implemented a fair AI policy.

Our productivity–pay gap guide explains why higher output need not translate automatically into higher typical pay. The practical challenge is to negotiate the connection before a temporary productivity experiment becomes a permanent increase in expectations.

Research checked 24 September 2026. Examples labelled hypothetical are calculations, not observed company results.