In 1865, the British economist William Stanley Jevons published The Coal Question and made an observation that annoyed everyone: James Watt’s new steam engines were far more efficient than the old ones. They burned less coal per unit of energy produced. Logically, total demand for coal should have fallen. The opposite happened. British coal consumption multiplied tenfold between 1830 and 1865.
Why? Because when a resource becomes cheaper to use, people invent new ways to use it. The efficient steam engine doesn’t replace the old one at a constant volume. It makes viable uses that didn’t exist before. New railway lines, new factories, new industries. The per-unit gain is absorbed, then overtaken, by the explosion in usage.
📌 Jevons’ paradox applied to AI: Economists call this the rebound effect. And it’s playing out right now with generative AI. Except the resource at stake isn’t coal. It’s the working time of skilled professionals.
Freed-up time never stays free
The 2023 Harvard/BCG study (Navigating the Jagged Technological Frontier), conducted on 758 Boston Consulting Group consultants, measured gains of 12 to 25% in speed and up to 40% in quality on creative and analytical tasks. An MIT study on professional writers shows a 37% gain in speed on writing tasks. For developers, GitHub reports 55% more speed on code completion with Copilot.
These gains are real. The problem is what we do with them.
The Microsoft Work Trend Index 2025 shows that the introduction of AI increased administrative tasks by more than 90% while reducing deep work by nearly 10%. Users don’t take the time they save to do less. They do more. More meetings, more emails, more documents, more micro-decisions. The volume of work expands to fill the freed-up capacity. And then exceeds it.
Cal Newport, the author of Deep Work and Slow Productivity, has documented this trap. His analysis is blunt: speeding up tasks has a way of pulling new ones in their wake. The result is not a productive utopia but a more intense frenzy. You don’t work less. You work differently, across a larger number of things, with the constant sense of moving fast without making progress on what matters.
Three mechanisms of the rebound effect
The AI rebound effect doesn’t work uniformly. It operates through three distinct channels.
The inflation of deliverables. When producing a document takes four hours instead of two days, you don’t produce one in four hours. You produce five in two days. The proposal that would never have been written because the prospect seemed lukewarm becomes viable. The presentation variant you wouldn’t have had time to test becomes mandatory. The quality standard expected by the client rises automatically, because the client knows AI exists and that their service providers are using it. Per-unit time falls, total volume rises, overall time stays constant or increases.
The expansion of scope. AI makes accessible tasks that were beyond the reach of a single person. A freelancer can now produce a complete SEO strategy, an editorial plan, landing-page mockups and a technical audit in a week. Before AI, they would have outsourced three of those four tasks. The result: they work on more fronts simultaneously, manage more complexity, and the time they save on each task is reinvested into opening up new ones. It’s the exact equivalent of the new railway lines made possible by Watt’s engine.
The compression of client deadlines. This mechanism is the most insidious because it comes from outside. When an entire market accelerates, expected deadlines compress. An audit that used to take three weeks can now be delivered in five days? The client expects five days. Then three. Then something halfway done in 48 hours. The productivity gain is gradually captured by the market, not by the producer. Economists call this rent dissipation: the competitive advantage of an innovation is redistributed to buyers as competitors adopt it.
The paradox at the macro scale: 89% of companies see nothing
The individual rebound effect is compounded by a macro problem that should give every AI evangelist pause. I’m one of them, so I’m telling myself this too.
McKinsey confirms this from another angle: of the 88% of organizations using AI, only 6% derive a significant financial impact (more than 5% of attributable EBIT). BCG puts the figure at 5%.
We’ve seen this movie before. Robert Solow, the Nobel laureate in economics, wrote in 1987 that you could see the computer age everywhere except in the productivity statistics. It took until the late 1990s, 15 to 20 years after the first massive computing investments, for the gains to show up in the macro figures. Paul David had shown a similar precedent with electricity: a 40-year lag.
Erik Brynjolfsson (Stanford) explains this lag through the J-curve: general-purpose technologies require massive complementary investments (process redesign, training, new work organization) that don’t appear in traditional indicators. During this phase, measured productivity can even fall. A Census Bureau study shows that manufacturing firms that adopted AI initially see their productivity drop by 1.33 points, before outperforming over a four-year horizon.
At the individual level, the gains are real but absorbed by the rebound effect. At the organizational level, they exist in potential but remain blocked by the absence of structural transformation. Plenty of activity, few measurable results.
What this changes for legal professionals
Thomson Reuters reports that equipped professionals save roughly 5 hours per week, and large firms see reductions of 30% in contract review time.
None of these figures says that lawyers work less or earn more.
What we observe in practice: firms that use AI handle more cases, produce longer briefs, cover more case law in their analyses. The standard of diligence rises. A lawyer who presents research covering 15 decisions where their competitor cites 50 thanks to AI loses credibility, even if the 15 were enough. The race is for volume and apparent thoroughness, not for time saved.
The phenomenon also shows up in pricing. Corporate clients, who know AI exists, are starting to challenge billing by the hour. A contract that took 8 hours to draft now takes only 3? The client expects to pay for 3 hours. The efficiency gain is transferred to the client. The lawyer has to take on more cases to maintain their revenue. Jevons, again.
Escaping the trap: three operational principles
Identifying the paradox isn’t enough. You still have to respond to it concretely.
Distinguish production from progress. Producing more deliverables isn’t the same thing as moving toward a strategic goal. The reflex when AI frees up time is to produce: one more document, one more email, one more proposal. The discipline is to ask the question before starting a task: does this deliverable advance a 6-month objective, or am I producing it because it has become easy? If it’s the second answer, the time would be better invested elsewhere.
Bill for value, not time. This is the only structural defense against deadline compression. As long as a professional bills by the hour, every productivity gain is automatically captured by the client. Migrating to fixed-fee, value-based or outcome-based billing becomes an economic necessity as AI accelerates production. An SEO audit that generates 40% more organic traffic is worth the same price whether it took two weeks or two days to produce. The conversation about price should be about the outcome, not the effort.
Protect non-productive time. The most counterintuitive piece of advice, and probably the most useful. The time saved by AI shouldn’t be entirely reinvested into production. Some of it should go to strategic thinking with no tool, no prompt, no expected deliverable. Paul Graham calls this the Maker’s Schedule: long, uninterrupted blocks for deep creative work. Ethan Mollick (Wharton) reaches the same conclusion by another route: the real AI lever isn’t prompting, it’s the ability to delegate and direct. And that ability develops in moments of stepping back, not in production sprints.
The real question is organizational
The steam engine didn’t make Victorian companies more profitable by running the old factories faster. It made possible a new kind of industrial organization. Factories changed shape, size, location. Paul David shows that this process took decades, and that the companies that simply plugged the new machine into the old organization are the ones that benefited least from it.
McKinsey and BCG converge on this point: what sets apart the 5 to 6% of companies that create value with AI is neither the sophistication of the tools nor the quality of the prompts. It’s the redesign of their processes. Leaders are three times more likely to have redesigned their workflows than laggards (McKinsey, 2025).
⚠️ Plugging ChatGPT into a law firm organized the way it was in 2019 is like plugging Watt’s engine into a coal mine dug by hand. It goes a little faster, but the structural gain is zero. Real transformation means rethinking what gets produced, how it’s produced, how it’s billed, and how professionals’ time is split between production, direction and reflection.
Jevons is not inevitable
Jevons’ paradox is not a law of physics. It’s a behavioral tendency. The rebound effect happens when you don’t see it, when you let freed-up time fill itself by default.
The professionals who will extract the most value from AI over the next two years won’t be the fastest. They’ll be the ones who understood that speed of production is a means, not an end, and who reinvested their time into what AI can’t do: seeing far ahead, deciding under uncertainty, building systems that create value without their constant presence.
Sources
- W.S. Jevons, The Coal Question (1865)
- Harvard Business School / BCG, Navigating the Jagged Technological Frontier (2023)
- Microsoft Work Trend Index 2025
- NBER Working Paper 34836, Firm Data on AI (February 2026)
- McKinsey Global Survey, The State of AI in 2025 (March 2025)
- BCG, Are You Generating Value from AI? The Widening Gap (September 2025)
- Lamy Liaisons / OpinionWay, Impact de l’IA sur les avocats et juristes en 2025
- ABA, 2024 Artificial Intelligence TechReport
- Thomson Reuters, Generative AI in Professional Services Survey (2025)
- Erik Brynjolfsson, The Productivity J-Curve (AEJ, 2021)
- Ethan Mollick, Management as AI Superpower, One Useful Thing (2025)
- Cal Newport, Slow Productivity (2024)
- Paul Graham, Maker’s Schedule, Manager’s Schedule (2009)