Vortex IQ
← Back to blog

Jevons paradox and AI: why cheaper, faster AI means more use, and how an ecommerce team keeps it in check

Jevons paradox and AI: why cheaper, faster AI means more use, and how an ecommerce team keeps it in check

The Jevons paradox is the finding that when something becomes cheaper and more efficient to use, total use of it often goes up rather than down. William Stanley Jevons described it for coal in 1865. It now applies to AI. Each call to a language model costs less and comes back faster than it did a couple of years ago, so teams make far more calls, run more agents and publish more content. The price of each task falls while the monthly bill, the review workload and the energy behind it can all rise.

For an ecommerce lead or a finance owner, the answer is to run AI the way you already run paid search. Give every agent a job with a measurable outcome, a budget and a cap. Ask for approval before work multiplies. Measure what each pound of spend produces, use the smallest model that does the job, and switch off whatever does not earn its place.

What the Jevons paradox is

In 1865 William Stanley Jevons, a British economist, published The Coal Question. Britain's industry ran on coal, and Jevons was worried about what would happen as its reserves were used up. A common hope at the time was that better engineering would ease the problem: steam engines were getting more efficient, each unit of work needed less coal, so the country would burn less.

Jevons argued the opposite. A more efficient engine made steam power cheaper, and cheaper power made it worth using in more places and for more jobs. Coal use per engine fell, but the number of engines and the work they did grew much faster, so total consumption rose. In his words: "It is wholly a confusion of ideas to suppose that the economical use of fuel is equivalent to a diminished consumption. The very contrary is the truth."

Economists now call the general pattern the rebound effect. When something gets more efficient, part of the saving is eaten up because people use more of it. The Jevons paradox is the strong form, where the extra use outweighs the saving and total consumption goes up. It does not happen with everything. It happens when a lot of demand is waiting for a lower price, and AI work has plenty of that.

How the Jevons paradox shows up with AI

Model providers keep cutting prices and releasing smaller, faster models, so a task that was too expensive to automate last year can be cheap enough this year. When the Chinese lab DeepSeek released a low-cost model in January 2025, Microsoft's chief executive Satya Nadella posted that the "Jevons paradox strikes again", arguing that as AI gets more efficient and accessible, its use will soar.

For the companies selling AI, that is good news. For a store paying for it, it means the cost line behaves differently from how it looks on the pricing page.

Efficiency per task versus total consumption

Inside an ecommerce business, the rebound tends to look like this.

  • More calls. Rewriting a product description used to take a copywriter real time, so it was done where it mattered. Now it takes seconds, so the whole catalogue gets rewritten, then rewritten again for each channel, each market and each test.
  • More agents. Once one agent works, adding another feels free: one for reviews, one for ad copy, one for support replies, one for internal reports. Each runs on its own schedule, and few get switched off.
  • More calls per task. An agent rarely makes a single request. It plans, looks things up, uses tools, checks its own work and retries, and each step is another call to a model, often carrying a long block of context with it.
  • More content. More pages, more variants, more emails. Some of it earns traffic or sales. All of it has to be checked by someone.
  • More spend and more energy. Most model access is billed by usage, so the bill follows volume. Every call also runs on hardware in a data centre that draws electricity.

The least visible resource is your team's attention. Agents can produce work much faster than people can review it, and unreviewed output at volume is where the expensive mistakes happen: a wrong rule applied across a whole category, or the same inaccurate claim repeated on hundreds of pages.

What it means for an ecommerce budget

Most AI spend behaves like ad spend or a utility bill rather than a software licence. It rises with use, and use rises when the unit price falls. A finance owner can watch every line item get cheaper while the total climbs month after month.

Ecommerce teams already know this pattern from paid search. Cheaper clicks do not mean a smaller bill if the budget is open, and nobody sensible runs a campaign without a daily cap and a target return. AI agents need the same treatment, for the same reason.

Three things make AI spend harder to read than ad spend:

  • Usage is counted in different ways. Some tools charge per seat, some per task, some in credits, and model APIs charge per token, the small units of text going in and out. A longer prompt or a chattier agent costs more without anyone changing a setting.
  • The costs are spread around. Model fees, the subscriptions for tools built on top of them, and the hours your team spends reviewing output often sit in three different budgets.
  • The value is often unmeasured. A dashboard can tell you how many descriptions were written. It rarely tells you whether any of them sold more.

The number worth tracking is cost per outcome: what you paid, in fees and in people's time, for each product page fixed, each support ticket resolved or each change your team approved and shipped. If you are still building the case for agents, the business case for AI agents covers how to set that measure up before you start.

Practical controls that keep AI use in check

None of these controls needs specialist software. They are decisions about how work is chosen, paid for and stopped.

Pick jobs with a measurable outcome

Start with a job you can describe in one sentence and measure with one number: product pages with missing attributes fixed, broken links cleared, low-stock alerts acted on. A brief like "use AI for content" has no natural end, so it grows to fill whatever budget it is given. If you need a starting list, AI agent use cases, job by job sets out jobs that each have a clear result.

Set budgets and usage caps per agent

Give each agent its own budget, as you would each ad campaign, and set a hard monthly cap with alerts well before it is reached. Many model providers let you set spending limits and alerts on an account or project, so give each agent its own project or key where you can, which also shows you exactly what each one costs. Cap the obvious multipliers too: items per run, retries and how often it runs.

Require approval before work multiplies

The cheapest moment to stop wasted work is before it starts. Have agents propose a plan with a clear scope, such as rewriting the descriptions in one category, and have a named person approve it before anything runs. Approve a small batch first, check the results, and only then let the same job run wider. Anything that writes to the live store, spends money or reaches customers should wait for that sign-off.

Measure outcome per unit of spend

Review each agent every month on two numbers side by side: what it cost and what it produced. Cost includes your team's review time as well as the invoice. Produced means the outcome you set at the start, measured in your own data; a count of outputs does not qualify. If cost per outcome rises month on month, the agent is drifting into volume for its own sake.

Use smaller models where they do the job

Routine work rarely needs the largest model. Tagging products, sorting support emails, checking a page against a list of rules and pulling fields out of a supplier file are often handled well by smaller models that cost less and respond faster. Test a smaller model on a sample of real work, compare the results, and keep the bigger one for the tasks where the difference shows. Two habits cut use further without cutting results: store and reuse output for anything that has not changed instead of generating it again, and send work that can wait to the batch options that several providers offer at a lower rate.

Switch off what does not earn its place

Agents are easy to start and easy to forget. Keep a simple register of every agent and automation: the job it does, who owns it, what it cost and what it produced last month. Any agent without an owner or a measured result gets paused. Pausing costs nothing, and if nobody notices for a month, you have your answer.

Sustainability and energy use

Running AI uses electricity. Many data centres also use water for cooling, and the hardware has to be manufactured and eventually replaced. Models are becoming more efficient per task, but, as Jevons would have predicted, total demand is growing. In its 2025 report Energy and AI, the International Energy Agency estimated that data centres used around 1.5% of the world's electricity in 2024, and projected that their consumption would more than double by 2030, with AI the main driver of that growth.

Keep that in proportion. The figure covers all data centres, not only AI, and the share that any one store accounts for is very small. The decisions that matter most at that level, such as how data centres are powered and cooled, sit with cloud and model providers and with governments.

What a store controls is how much it asks for, and the controls above are the same ones that cut energy use. A call you never make uses no electricity. A smaller model generally uses less than a large one for the same task. Output that is reused is not processed twice.

If sustainability is part of your reporting, three further steps are worth taking:

  • Ask your AI vendors and model providers what they publish about energy use, emissions and where their workloads run.
  • Include AI use as its own line in your footprint estimates, using provider figures where they exist and saying clearly where you have estimated.
  • Do not claim that AI has made your business greener unless you have measured it. A saving per task is not a fall in total use, which is exactly what Jevons pointed out.

The same discipline runs through how Vortex IQ, the AI workforce for ecommerce, works. Specialist crews take one job at a time, find the problem in your own store data, prepare the exact change and wait for a decision. It proposes. Your team approves. Nothing goes live without your say-so. Each approved change is applied with an undo point and checked afterwards, so the work you pay for is work someone chose.

Frequently asked questions

Is the Jevons paradox a reason to avoid AI?

No. It is a warning that efficiency on its own will not hold total use down. What holds it down is a set of decisions: which jobs to run, at what budget, with whose approval, and when to stop.

Who first described the Jevons paradox?

William Stanley Jevons, a British economist, in his 1865 book The Coal Question. He argued that more efficient steam engines would lead Britain to burn more coal in total, because cheaper steam power would be put to more uses.

What is the difference between the Jevons paradox and the rebound effect?

The rebound effect is the general pattern in which part of an efficiency saving is lost because people use more. The Jevons paradox is the case where the extra use is larger than the saving, so total consumption rises.

Does cheaper AI always mean a bigger AI bill?

Not always. It means the bill follows how much you use, and lower prices make it easy to use more. Teams that set a budget and a cap for each agent, and pay only for jobs with a measured result, can see their cost per outcome fall as prices drop.

How should a store set its first AI budget?

Pick one job, estimate how much work it covers, and set a monthly cap with an alert well below it. Run it for a month, then compare the full cost, including review time, with the outcome you measured. Raise the cap only when the cost per outcome holds up.