Book a call (opens Google Calendar)
Book a call (opens Google Calendar) hey@pronovahq.com

Jevons' Paradox and AI: Why Cheaper Work Makes Some Teams Busier

By Izzy Aly, Co-Founder, Systems and Architecture

77% of employees using AI say it added to their workload. Jevons' Paradox explains why, and which of your workflows will give time back instead.

In July 2024, the Upwork Research Institute asked executives and employees the same question from two sides. 96% of C-suite leaders expected AI to raise productivity. 77% of employees using AI said it had added to their workload (Upwork).

Both groups can be right at once. That's not a rollout failure. It's a 160-year-old piece of economics doing exactly what it predicts. The useful part is that it doesn't predict the same thing for every job. Some of your workflows will hand time back. Others will quietly double in volume. You can tell which is which before you automate anything.

Key takeaways

  • In July 2024, 96% of executives expected AI to raise productivity, while 77% of employees using AI said it added to their workload (Upwork Research Institute).
  • Jevons' Paradox only holds when demand for the work grows as it gets cheaper. Invoice chasing doesn't. Proposals do.
  • Sort each workflow before you automate it. Inelastic work frees capacity. Elastic work needs a review step designed in first.

What does Jevons' Paradox actually say?

In 1865, William Stanley Jevons argued in The Coal Question that more efficient steam engines would make Britain burn more coal, not less. Cheaper energy made new uses worth paying for. In January 2025, Microsoft CEO Satya Nadella applied it to AI: "Jevons paradox strikes again!" (Fortune).

Most takes stop there and conclude that AI will create more work for everyone. That skips the condition the whole idea rests on.

Jevons only kicks in when demand is elastic: when people buy a lot more of something once it gets cheaper. If demand is flat, efficiency just means you need less input. Economist James Bessen showed both outcomes inside the same industries. In 19th-century cotton textiles, labour productivity rose roughly 30-fold and cloth consumption rose about 100-fold, so jobs grew. Later, demand saturated, and productivity gains, along with trade, cut production jobs. Steel followed the same arc (Bessen, NBER Working Paper 24235, 2018).

Bank tellers are the version everyone quotes. Teller numbers held up while ATMs spread, because banks used the savings to open more branches. That story has run out. The US Bureau of Labor Statistics now projects teller jobs falling 13% between 2024 and 2034, from 347,400 to about 302,500, and names self-service automation as the cause (BLS). Same technology, different decade, opposite result.

So the honest version is this. When AI makes a piece of work cheaper, you either get more of that work or you need fewer hours for it. Which one depends on whether anyone wants more of it.

Which of your workflows will get busier, and which will free up time?

In 2024, 39% of employees using AI told Upwork they now spend more time reviewing or moderating AI-generated content (Upwork). That review load lands hardest on elastic work, where cheaper output means more output. Sort your workflows first.

Elastic or inelastic: the same efficiency gain sends work two ways A workflow gets cheaper with AI. Ask whether you would do ten times more of it at a tenth of the cost. If no, it is inelastic work with fixed volume, such as invoice chasing, timesheet checks, CRM clean-up and monthly client reports, and your team gets hours back, so automate it first. If yes, it is elastic work whose volume grows, such as proposals, outreach, research and content, and the result is more output to review, so design the review step before scaling. A dashed arrow shows inelastic work turning elastic when someone raises the standard, for example monthly reports becoming weekly. A workflow gets cheaper with AI Would you do 10x more of it at a tenth of the cost? NO YES INELASTIC WORK Volume stays fixed Invoice chasing, timesheet checks, CRM clean-up, monthly client reports ELASTIC WORK Volume grows Proposals, outreach, research, analysis, content raise the standard monthly to weekly Your team gets hours back More output to review Automate these first Design the review step before scaling
Elastic or inelastic: the same efficiency gain sends work two ways A workflow gets cheaper with AI. Ask whether you would do ten times more of it at a tenth of the cost. If no, it is inelastic work with fixed volume, such as invoices, timesheets, CRM clean-up and monthly reports, and hours come back, so automate it first. If yes, it is elastic work whose volume grows, such as proposals, outreach, research and content, and there is more to review, so design review first. A dashed arrow shows that raising the standard turns inelastic work elastic. A workflow gets cheaper with AI Would you do 10x more at a tenth of the cost? NO YES INELASTIC Fixed volume Invoices, timesheets, CRM clean-up, monthly reports ELASTIC Volume grows Proposals, outreach, research, content dashed: raise the standard Hours come back More to review Automate first Design review first
Fig. 1 — The same efficiency gain sends work in opposite directions, depending on whether anyone wants more of it.

Inelastic work has a fixed volume. You have as many invoices as you have clients. Cheaper chasing doesn't create more invoices. The same goes for timesheet reconciliation, CRM clean-up, onboarding paperwork and the monthly client report. Automate these and the hours come back.

Elastic work grows when it gets cheaper. If a proposal takes 20 minutes instead of a day, you'll bid on more work. If outreach costs nothing, you'll send more of it. Research, analysis and content behave the same way. Automate these and the team gets more output, plus a pile of output to check.

There's a trap between the two. Inelastic work turns elastic the moment someone raises the standard. The monthly report is fixed volume until a partner says, "Now that it's automatic, can we send it weekly?" That one decision moves a workflow from freeing time to consuming it. It's a management choice, not a law of nature, and you can make it on purpose.

A quick test for each job on your list: if this took a tenth of the time, would we do ten times more of it? If the answer is no, it's inelastic. If the answer is "probably, yes", plan for the volume.

What does the paradox cost when you don't plan for it?

In February 2026, researchers Aruna Ranganathan and Xingqi Maggie Ye at UC Berkeley Haas reported in Harvard Business Review that AI didn't reduce work at the tech company they studied. It intensified it (HBR). They followed 200 employees for eight months. Here's where that intensity shows up.

The output treadmill

The Berkeley team saw three things: people took on more tasks, work bled into lunch breaks and the minutes before leaving for the day, and people ran more jobs at once. The tools made one more task feel free. The authors recommend that teams agree an explicit "AI practice": norms for when and how to use the tools. The study is still running, so treat these as early findings.

The review bottleneck

When drafting gets fast, checking becomes the slow part. One person can now produce more than another person can review.

Faster drafting moves the queue to review Two pipelines, each running from drafting to a queue to review to done. Before AI, drafting and review keep pace, and only two items wait in the queue. After AI, drafting runs several times faster while review has the same capacity, so ten items wait in the queue in front of review. The queue is the bottleneck, and done work comes out at the same rate as before. BEFORE AI Drafting Review Done Review keeps pace with drafting AFTER AI Drafting several times faster Review same capacity Done same rate THE BOTTLENECK Drafts pile up in front of a review step that hasn't grown
Faster drafting moves the queue to review Two pipelines, each running from drafting to a queue to review to done. Before AI, drafting and review keep pace, and two items wait. After AI, drafting runs faster while review has the same capacity, so ten items queue in front of review. The queue is the bottleneck, and done work comes out at the same rate. BEFORE AI Drafting Review Done Review keeps pace with drafting AFTER AI Drafting faster Review same capacity Done same rate THE BOTTLENECK Drafts queue at a review step that hasn't grown
Fig. 2 — Faster drafting moves the queue to review. The constraint shifts; it doesn't disappear.

People also misjudge where the time goes. In a 2025 randomised trial by METR, 16 experienced open-source developers took 19% longer on real tasks when allowed to use AI tools. They still believed the tools had made them about 20% faster (METR). It's a small study using early-2025 tools, so don't read it as "AI makes people slower". Read it as "people can't feel the difference, so measure it".

The maintenance tail

Every workflow you automate is one more thing that can break when a tool changes its API or a process changes underneath it. In 2025, Google's DORA research found AI adoption now goes with higher delivery throughput. In announcing the report, Google added that AI adoption "does continue to have a negative relationship with software delivery stability", because acceleration "can expose weaknesses downstream" (Google Cloud). The report describes AI as an amplifier of whatever a team already does well or badly (DORA). More automations means more upkeep. Budget for it from day one.

What happens to employees, especially junior ones?

In 2025, Stanford's Digital Economy Lab found that US workers aged 22 to 25 in the most AI-exposed occupations saw a 16% relative decline in employment, using ADP payroll data (Brynjolfsson, Chandar and Chen, "Canaries in the Coal Mine?"). The declines cluster where AI automates the work rather than assisting with it.

That's the uncomfortable part of this topic. I'm not going to tell you it works out fine for everyone, because the evidence doesn't say that.

Here's what it does say. The work that grows under the paradox is deciding what's worth doing, writing down what good looks like, checking output and handling the exceptions. Upwork found 47% of employees using AI didn't know how to achieve the productivity gains their employers expected. That's a skills gap, and it's fixable.

The junior problem is sharper. Juniors traditionally learned the job on exactly the inelastic, repetitive work that firms automate first. Take that away and you lose the training ground for your next seniors.

One option: make juniors the reviewers. When an agent drafts and a person approves, the person approving learns what good work looks like on every single case. Pair that with a senior who checks the checker, and the junior still learns the job. They just learn it from the review side instead of the typing side.

For the business, the move is the same either way. Decide what the freed hours are for before they appear. If you don't, the treadmill decides for you.

Where does ProNova fit in this?

In 2025, Eurostat found that among EU firms with 10 or more employees that considered AI but didn't use it, 70.89% cited lack of relevant expertise as a reason (Eurostat). That's the gap we work in. Here's how our process maps to each cost above.

I'll be straight about one thing first. We're part of the paradox. Every workflow we ship is one more thing your business has to keep running. So our process is built around picking the right jobs, keeping a human on the decisions that matter, and planning the upkeep, not around shipping as much as possible.

Four costs of the paradox, and what answers each one Four rows. Output treadmill, where volume rises to fill the time, maps to the AI Workflow Lab, which sizes and prices each workflow before any build. Review bottleneck, where drafts outrun checking, maps to agents that draft while your team decides, with sign-off on the actions you choose. Maintenance tail, where every workflow needs upkeep, maps to agents that run in your own account, get re-tested when platforms change and have a trained owner. Junior pipeline, where there is less entry-level work to learn on, maps to team training on two to four of your real tasks, timed before and after. COST OF THE PARADOX WHAT ANSWERS IT Output treadmill volume rises to fill the time AI Workflow Lab Sizes and prices each workflow before any build Review bottleneck drafts outrun checking Agents draft, your team decides Sign-off on the actions you choose Maintenance tail every workflow needs upkeep Runs in your account Re-tested when platforms change; trained owner Junior pipeline less entry-level work to learn on Team training on real work 2–4 of your tasks, timed before and after
Four costs of the paradox, and what answers each one Four rows. Output treadmill maps to the AI Workflow Lab, which sizes and prices each workflow. Review bottleneck maps to agents that draft while you decide, with sign-off on actions you choose. Maintenance tail maps to agents that run in your account and get re-tested when platforms change. Junior pipeline maps to training on your real work, with tasks timed before and after. COST WHAT ANSWERS IT Output treadmill volume fills time AI Workflow Lab sizes and prices each workflow Review bottleneck drafts outrun checks Agents draft, you decide sign-off on actions you choose Maintenance tail upkeep per workflow Runs in your account re-tested when platforms change Junior pipeline less work to learn on Training on real work tasks timed before and after
Fig. 3 — Each cost of the paradox, and the part of our process that answers it.
  1. The AI Workflow Lab sorts the list. It's a 2–3 week sprint. It ends with priced specs for each workflow, a training plan and a working prototype. We size every workflow by how many systems it touches, whether it needs an approval step, and how many exception paths it has. That's where you decide which side of the paradox each job lands on, before you pay to build it.
  2. Agents draft, your team decides. A custom AI agent only acts on its own once its work matches your team's. You choose which actions need sign-off and which data it can touch. Every run has a spending limit.
  3. Upkeep is planned, not discovered. The agent runs in your own account, the code lives in your repository, and one person on your team is trained to manage it. We re-test when the platform changes, and ongoing support is optional.
  4. Training measures instead of guessing.Team training picks 2–4 real tasks, times them as they run today, then times them again after. That answers the METR problem: you see the actual difference, not the one people feel.

We haven't got customer results to quote yet, so I won't invent any. What I can show you is the process, and the questions it forces you to answer before anything gets built.

If your list has one clearly defined job on it, you don't need the Lab. Go straight to an agent build.

Frequently asked questions

Is Jevons' Paradox real for AI and jobs?

It's real, with a condition. In 2018, economist James Bessen showed that automation grew jobs in textiles and steel while demand kept rising, then cut them once demand saturated. AI will add work wherever cheaper output creates more demand, and remove hours where demand stays flat.

Why is my team busier after rolling out AI?

In 2024, 77% of employees using AI told Upwork it had added to their workload, and 39% spent more time reviewing AI output. Cheaper drafting usually means more drafts, more checking and more side tasks. Without agreed norms for using the tools, the saved time gets absorbed.

Which work should a small firm automate first?

Start with inelastic work: jobs with fixed volume, like invoice chasing, CRM clean-up and recurring reports. Automating them returns hours instead of creating more output to check. Leave elastic work, like proposals and outreach, until you've designed the review step.

Will AI replace junior staff?

Stanford's 2025 payroll analysis found a 16% relative employment decline for 22-to-25-year-olds in the most AI-exposed occupations. The risk is real. Firms that keep juniors as reviewers of AI output still build their next seniors. Firms that cut the entry-level work entirely lose that pipeline.

Sources

Related: if your worry is the AI bill rather than the workload, the same paradox drives it. See why falling token prices won't save your budget.

Topics
Author
Izzy Aly · Co-Founder, Systems and Architecture
Share

Work with us

Got more workflows than time to fix them?

The AI Workflow Lab maps where AI fits across your CRM and the tools around it, and tells you what to build first and what each build will cost.

See how the Lab works →
or book a scoping call (opens Google Calendar)

Keep reading

All posts →