AI makes work cheaper. That does not mean there will be less of it.

In 1865, William Stanley Jevons argued against a simple claim about steam engines. A more efficient engine used less coal for each unit of work. Many people assumed that better engines would therefore reduce total coal use. Jevons saw the opposite pressure. Cheaper power made engines profitable in more mines, factories, and trades. More engines could consume more coal even when each engine needed less.

That is now called the Jevons paradox. The name makes it sound like a trick. It is ordinary demand. When the cost of doing something falls, people can do more of it, use it in more places, and invent uses that were not worth the old price.

The effect is not automatic. If demand is fixed, efficiency can reduce total use. If demand grows enough after the price falls, total use rises. Jevons gives us a question to ask about AI. What happens when the cost of producing an answer, a draft, a prediction, or a line of code falls toward zero?

The labor version needs more care. A person is not a resource that a company simply consumes. A job contains many tasks, relationships, decisions, and duties. Software can make one task faster without making the occupation disappear.

Economist James Bessen studied how demand changed employment in textiles, steel, and cars. His conclusion was conditional. Automation can increase employment while it lowers prices if customers buy much more of the product. Employment can later fall when demand stops responding as strongly. The result depends on demand, not on efficiency alone.

AI can therefore pull work in opposite directions. A company may need fewer hours to produce the same number of reports. It may also decide to produce ten times as many reports, serve customers who previously received no help, or test many more product ideas. Some of that new volume needs human judgment, review, sales, support, and repair. Some of it does not. Jevons cannot settle the balance for us.

This is why a demo that saves thirty minutes does not prove a headcount plan. Management still decides the amount of output, the quality bar, the price, the staffing model, and how much risk a worker must absorb.

A 2025 study followed 5,172 customer-support agents after a company introduced an AI assistant. The tool suggested replies and linked workers to internal documents. People remained responsible for the conversation and could ignore or edit each suggestion.

Agents resolved 15% more issues per hour on average. Less experienced and lower-skilled workers improved by about 30%. Newer workers moved through the learning curve faster. Customers were also less hostile, and the company recorded lower attrition among newer workers.

Those results matter. They show a tool helping people learn and do better work in a real company. They do not show what happened to total employment or wages. The authors say their data cannot answer those questions. A manager could use the gain to reduce queues, shorten shifts, serve more customers, raise targets, or cut staff. The study measures the tool. It does not approve the management choice.

The International Labour Organization and Poland's NASK institute reviewed almost 30,000 tasks in 2025. They estimated that one in four jobs worldwide had some exposure to generative AI. Clerical work had the highest exposure, and women faced a larger share of the most exposed jobs in high-income countries.

The report did not call one in four jobs replaceable. It found that transformation was more likely because most occupations still include tasks that need people. That difference matters. A changed job can become safer, better paid, and easier to learn. It can also become a thinner job made of exceptions, complaints, and constant checking after the easier work moves to software.

Counting exposed tasks tells us where change may occur. It does not tell us whether a worker will keep a job, receive training, share the productivity gain, or carry more work for the same pay.

The most immediate Jevons effect at work may be neither more jobs nor fewer jobs. It may be more work per person. When a draft takes five minutes instead of an hour, the saved time often becomes five more drafts. When software clears routine cases, the worker receives a day made only of difficult cases.

OECD surveys found that 75% of surveyed finance workers and 77% of manufacturing workers who used AI said it increased their pace of work. The report also describes a manufacturer that chose not to automate every easy task. Its managers understood that easy work gave people a mental break between harder problems.

That detail gets closer to the human cost than a productivity score. A queue can expand until it fills every minute the tool saves. The worker then supplies more attention, more correction, and more emotional effort. The company records higher output. The person experiences a tighter day.

This outcome is a policy choice inside the company. A model does not set the daily target, remove recovery time, monitor keystrokes, or decide that a fifteen-percent gain belongs only to the employer.

Before a company deploys AI, workers should know what the saved time is for. A useful policy states which work will stop, which service will expand, how quality will be checked, and how the gain will affect staffing, hours, and pay.

The policy also needs a total-use measure. Cost per task can fall while the number of tasks rises much faster. Measure total compute, total review time, total customer contacts, total failed work, and total hours. A per-task number can hide the rebound that matters.

• Give workers a role in choosing the tasks and limits before rollout.
• Count review, correction, and failed attempts as work.
• Keep a quality floor that does not fall when volume rises.
• Set a queue or workload limit instead of treating every saved minute as new capacity.
• State how workers share the gain through pay, shorter hours, training, or added staff.
• Publish total energy and total task use, not only efficiency per request.

A job seeker does not need to ask a hiring manager about nineteenth-century coal economics. Ask what happened after the tool became faster. The answer will tell you whether the company treats efficiency as room for better work or as permission to stretch every worker.

• Which tasks did the tool remove, and which tasks increased after adoption?
• Did targets, team size, hours, or pay change when output rose?
• Who reviews AI output, and is that time included in capacity plans?
• Can a worker reject a suggestion without hurting a performance score?
• What happens when the system fails during a busy period?
• Does the company measure customer results and worker health, or only volume?

Jevons helps explain why cheap AI may lead to more AI, more output, and more electricity use. It also warns us not to turn a task-level saving into a system-level promise.

The labor result remains open. Demand can grow. New work can appear. Existing work can become easier. Employers can also use the same tool to raise targets, narrow jobs, cut staff, or transfer risk to contractors. Acemoglu's 2024 estimate of AI's economy-wide productivity effect was modest, at no more than 0.66% over ten years under his assumptions. Large claims about growth or job loss still run ahead of measured results.

The useful question is not whether AI is efficient. It is who controls the extra capacity. If workers have a say, the gain can buy time, learning, access, and better service. If they do not, the queue will usually find the empty space first.