AI and the workplace: what the evidence actually shows

The most careful global measurement available comes from the International Labour Organization’s 2025 index of occupational exposure to generative AI. Its headline is that one in four workers worldwide is in an occupation with some exposure, while 3.3% of global employment falls into the highest exposure category. Clerical occupations remain the most exposed, and exposure has begun extending into specialised professional and technical tasks.

Exposure is not evenly distributed

Two divides in the data matter for planning. The first is income: total exposure is 11% of employment in low-income countries against 34% in high-income ones, because exposure tracks how digitised the work already is. The second is gender: 4.7% of female employment sits in the highest exposure category against 2.4% of male employment, widening further in high-income countries. Where clerical and administrative work is disproportionately done by women, an untargeted automation programme will land unevenly whether or not anyone intended it to.

What transformation means for an employer

• —Map tasks rather than job titles — exposure sits at the task level, and a role is usually a bundle of very differently exposed work.
• —Expect the saved time to move rather than disappear: DORA’s 2025 research found time saved generating code is often reallocated to verifying it.
• —Redesign the role around what is left, and say so explicitly, before people conclude the tool is being introduced to replace them.
• —Invest in the judgement work — validation, exception handling, client relationships — because that is what remains scarce.
• —Measure quality of output after adoption, not just volume; faster wrong answers are not productivity.

The training question is not optional

A workforce that has never been taught what these tools do badly will use them as though they do everything well. That is where the risk sits: a confident answer accepted without checking, a client document containing something nobody verified, a decision justified by an output nobody can explain. Practical training on limitations, on what may not be entered, and on how to check a result is a control as much as it is a benefit.

Lower measured exposure is not the same as lower urgency. It reflects the current shape of the work, and that shape changes the moment the work is digitised — which is what most organisations are already doing for other reasons.