# More efficiency = less work? AI's Jevons paradox > Efficiency gains do not disappear, they shift. Why the same technology removes work in one area and creates new work in another. - **Author:** Dominic Asche - **Date:** 08 / 2026 - **Reading time:** 7 min - **URL:** https://fse-group.de/en/think-tank/the-jevons-paradox-of-ai With AI a task takes ten minutes instead of an hour. So the company needs fewer working hours? For that one task the arithmetic holds. For the system as a whole, the opposite can be true. The arithmetic looks so reasonable that hardly anyone questions it: with AI a task takes ten minutes instead of an hour. So the company needs fewer working hours. For that one task the arithmetic holds. For the system as a whole, the opposite can be true. William Stanley Jevons described this contradiction back in the 19th century, using coal consumption as his example. More efficient steam engines needed less fuel for the same output. That made steam power cheaper and attractive for more applications. Coal consumption rose as a result, even though every single engine ran more frugally. [1] So more efficiency can trigger more demand. Applied to the title, that means: more efficiency can also mean more work. Applied to AI it means this. When the effort for a task drops sharply, that task is not necessarily just done more cheaply. It may well be done far more often. New applications become viable, expectations rise. A monthly analysis turns into daily monitoring, a semi-annual test into a continuous review. What matters is the ratio between two developments: how sharply does human effort per task fall? And how sharply does the number of tasks grow? If volume grows faster than productivity, more efficiency can lead to more work. If volume stays largely constant, the need for human work falls. The Jevons paradox therefore protects no job. What it does explain is why the same technology expands the volume of work in one area and reduces it in another. ![Infographic on AI's Jevons paradox, titled When efficiency creates demand. A four-step chain of effects: first, effort per task falls; second, new applications become viable; third, the number of tasks grows; fourth, work shifts instead of disappearing. Below it a line chart titled AI is eating the billable hour, showing valuation multiples for IT services against the S&P 500 from 2018 to 2026 on a consensus NTM P/E basis. The IT services line rises to 38 by 2022 and falls to around 10 by 2026, while the S&P 500 runs at about 20. IT services thus trade at a decade low, around one third of their early-2025 level. Source CapIQ and Morgan Stanley, 27 July 2026, chart after a16z. At the foot, three questions: which tasks get cheaper? Does volume grow as a result? What becomes the new bottleneck?](https://fse-group.de/media/think-pieces/20260825_thinktank-jevons-paradox-en.png) ### Software development shows the contradiction Software development is a particularly good test case. AI intervenes directly in the writing of code. At the same time, demand for software has barely any fixed ceiling. In most companies there is no shortage of ideas for new applications, interfaces, automations or data products. What is short is time, budget and development capacity. So when AI lowers the cost of a single implementation, the result is not simply the same software built with less effort. More gets built. A developer may need less time for an individual feature. In exchange, additional prototypes appear, more tailored solutions, more integrations and shorter release cycles. Existing systems also get changed more often, because changes become economical that would previously have stayed in the backlog. More output does not automatically mean more stability or more security, though. Quite the opposite: as the volume of generated code grows, so does the effort needed to review it. A chart published by a16z in August 2026 shows at the same time how sharply the capital market is questioning the classic business model of IT service providers. The valuation multiples for IT services examined there stood at only around a third of their early-2025 level. Gagan Chawla summed the development up pointedly: “AI is eating the billable hour.” [3] The chart is not an employment statistic. It does not prove that two thirds of IT work is disappearing. What it does show is an expectation held by the market: a business model whose revenue grows mainly with the number of hours sold comes under pressure when AI takes over a growing share of the work that used to be billable. Fewer hours per result can nevertheless lead to more software work. What changes is the nature of that work. Routine code, standard migrations and simple tests lose value. Architecture, integration, security and accountability gain in importance. ### A profession is more than its most easily automated task This thought also shaped the discussions in our AI harnessing course at Stanford. In economic terms, a profession can be described as a bundle of individual tasks. In practice, work is more than the sum of those tasks. Between them sit context, experience, coordination and accountability. A person recognises when a case does not fit the standard. They weigh the consequences for customers, colleagues and adjacent systems. They decide when a result is solid enough to keep working with. Charles I. Jones describes professions in *AI and Our Economic Future* as bundles of complementary tasks. Some of them are easy to automate. Others remain the bottleneck for now. Jones speaks of “weak links”, tasks that cap the productivity of the entire process. Once the remaining activities are automated, it is precisely these tasks that can become more valuable. [2] Anyone who simply wants to automate everything also skips a prior question: does the value chain make sense at all, and is it worth automating? Jones points to radiology. In 2016 it was publicly predicted that AI would take over medical image analysis so quickly that training radiologists would hardly be worthwhile. Almost ten years later, according to Jones, there were more people working in the profession and salaries were higher. That does not prove AI caused the growth. It does show how wrong it is to reduce a profession to its most visible task. Radiologists do not just read images. They place findings in a clinical context, coordinate with other physicians, carry out procedures and bear responsibility for decisions. For companies this leads to a pragmatic approach: AI should first take on clearly delimited tasks rather than replacing whole roles wholesale. The more precisely a task is defined, the better its quality, its limits and its escalation paths can be checked. ### More code is not yet better software The falling cost of software production has a flip side. The first draft of the code becomes cheap. The responsibility for it does not. What the content world calls “AI slop” has long had its counterpart in software development. AI produces code that looks plausible but can be functionally wrong, insecure or hard to maintain. More code therefore also means more tests, more reviews and, in the worst case, more technical debt. The Jevons mechanism shows up here too. Higher productivity creates additional work. That work is not automatically valuable, though. A team that produces ten times more code and then has to correct a large share of it has raised its throughput, not its benefit. Our FSE Pentest Agent shows how the effect can be put to good use. It lets applications be checked for vulnerabilities daily instead of twice a year. According to our published case study, the time needed to run a test falls by 85 per cent. [4] The decisive change lies less in the time saved than in the higher test frequency. Because the individual test gets cheaper, the security standard rises. Results have to be assessed, risks prioritised and vulnerabilities fixed. Running the test repeatedly gets automated. The specialist work shifts towards interpretation and accountability. If a company kept testing only twice a year, it would indeed need less working time. A Jevons effect only arises once the lower cost leads to higher demand. ### In the energy sector the exception stays decisive In the energy sector both developments become visible at the same time. A typical clarification case does not consist of a single entry in SAP or another core system. Contract details have to be checked, entries from CRM, ERP, billing and market communication reconciled, and dependencies understood. In the end someone has to answer for the correction being right, both functionally and in regulatory terms. AI can take over individual steps of that sequence. It can classify cases, gather information, compare data, prepare proposed solutions and, within clear permissions, execute changes in the system. Case handling does not disappear as a result. Its share of repeatable routine work can fall considerably, though. That should not be glossed over. In a project for SENEC we built an agent for inside sales. It takes on defined steps in the post-processing of orders, recognises error patterns and applies predefined solutions in the ERP system. Throughput time fell by 80 per cent as a result. At the same time, capacity equivalent to one full-time position was saved. [5] Where order volume is largely fixed, higher productivity does indeed lead to less routine work. For energy companies that is exactly where the strategic dividing line runs. In standardised processes with a largely fixed volume, in certain bookings, checks or clarification cases, human work will decline. In tasks with demand that has gone unmet so far, it can grow. That includes continuous quality assurance, cybersecurity, market monitoring, data maintenance and the integration of system landscapes that have grown over time. ERP and CRM systems, billing and market communication do not disappear because of AI. They get connected more tightly. That increases the need for people who understand these connections and take responsibility for the overall result. ### Jevons protects no job The *Future of Jobs Report 2025* expects 170 million new and 92 million displaced positions worldwide by 2030. These figures relate to several economic and social developments, not to AI alone. The growing occupational groups include software developers as well as specialists in data, AI and information security. For administrative and classic case-handling roles, declines are expected. [6] The arithmetic net gain of 78 million positions is therefore no reason to relax. The new jobs do not automatically appear where others vanish. They call for different skills and may arise in other companies, regions or pay brackets. For company leadership it is not enough to define a blanket savings target for full-time positions. They have to understand which tasks get cheaper, whether that grows the volume of work, and which human activity then becomes the bottleneck. Just as important is the question of how much additional work is created by review, integration and error correction. More efficiency can mean less work. In many administrative routines that is exactly what will happen. More efficiency can also trigger more software, more tests, more market monitoring and more decisions. The contradiction resolves as soon as we stop equating a minute saved with a job saved. What AI changes first is tasks. What that turns into for a profession is decided by the interplay of demand, process design and accountability. The Jevons paradox gives no reassuring answer to that. It does pose the right question: what does AI make available in abundance, and what does that make genuinely scarce? In the end an old entrepreneurial question remains. What value do we create when a previously scarce capability is suddenly available in abundance? The answer decides which work will count in future. #### Sources - [1] William Stanley Jevons: [*The Coal Question*](https://files.libertyfund.org/files/317/Jevons_0546_EBk_v6.0.pdf), first published 1865. - [2] Charles I. Jones: [*AI and Our Economic Future*](https://web.stanford.edu/~chadj/AIandEconomicFuture.pdf), *Journal of Economic Perspectives*, Vol. 40, No. 3, 2026. - [3] a16z: [*Charts of the Week: Bookslop*](https://www.a16z.news/p/charts-of-the-week-bookslop), section “IT Services, Down (But Not Out?)”, 7 August 2026. Gagan Chawla: [his reading of the chart on LinkedIn](https://www.linkedin.com/posts/gagankchawla_a16z-charted-it-services-multiples-at-decade-share-7495710895218745344-LQ-w/). - [4] FSE Group: [*FSE Pentest Agent*](https://fse-group.de/en/unsere-arbeit/fse-pentest-agent/). - [5] FSE Group: [*SENEC Vertriebsinnendienst Agent*](https://fse-group.de/en/unsere-arbeit/senec-vertriebsinnendienst-agent/). - [6] World Economic Forum: [*The Future of Jobs Report 2025*](https://www.weforum.org/publications/the-future-of-jobs-report-2025/in-full/2-jobs-outlook/).