AI got cheaper, so why is our bill rising? Jevons paradox and 6 steps

Your AI bill rises because cheaper AI gets used in more places and more often, and that grows faster than prices fall. Economists call this the Jevons paradox, and it usually doesn't mean anyone is being careless.

Plenty of companies see the same pattern. Last year, AI translated a few foreign-language emails, at a barely noticeable monthly cost. This year, customer service uses it to answer messages on LINE (the messaging app most Taiwanese customers use), sales uses it to draft quotes, and accounting uses it to check statements. Each use costs less than it did last year, yet the total is several times higher.

Below are prompts for an AI assistant and six steps to see where the money goes and budget better, then why it happens and how to manage costs as usage grows.

Ask AI first

An AI assistant such as ChatGPT, Claude or Gemini can total the bills or usage details (CSV or PDF) from each AI service's billing page by department and process, and work out the cost per task. Copy these and swap in your own data (these are examples).

Total by department and process

Attached are the last 3 months of bills and usage details from our AI services, plus a lookup sheet showing which department and process each account or API key belongs to. Total the monthly cost by department and process, list the 5 that grew the most, and check your totals against the bills. List accounts or keys missing from the lookup sheet separately. Give me a table I can paste into a spreadsheet.

Work out the cost per task

I've also attached how many tasks each process completed last month (for example, 1,200 orders entered and 3,500 customer replies). Work out the AI cost per task for each process, then build a sheet for next year's budget: expected tasks × cost per task, plus the new uses we expect. Leave blank and flag any numbers I need to fill in.

Check that the AI's totals match the bills; if they don't, ask it to list the rows it left out. For accounts missing from your lookup sheet, ask around to find out who uses them. If the billing details include employee names, emails or the last digits of a credit card, delete those columns or replace them with codes before uploading.

6 things you can do now

1. List all of your AI spending in one sheet

In Excel or Google Sheets, list every AI service and subscription: monthly cost, who uses it and for which process, pulled from credit card statements and each service's billing page (the first prompt above can total it). Then it's clear which uses are largest and which could be combined.

2. Work out the cost per outcome for one process

Pick a process that already uses AI, such as order entry or customer replies. Its monthly AI cost divided by tasks completed is the cost per task; compare it with the time and error rate of doing the task by hand. If you haven't started with AI yet, where to start with AI can help you pick a process.

3. Set monthly quotas for departments and applications

On the same sheet, give each department and each application a monthly quota and an owner, remind the owner at 80%, and review once a month. The point isn't to restrict use; it's to make sure every dollar has someone accountable for it.

4. Budget by how much work you want done

Lower unit prices don't mean the budget can shrink in proportion. A steadier estimate is the second prompt's formula: tasks you want AI to complete next year × cost per task, plus new uses. Then check monthly whether the cost per task is going down.

5. Test small models on real cases for simple work

Small models are often good enough for well-defined work like classification, field extraction and summarization. Take 50–100 real cases, run each through a small and a large model, record accuracy and cost in a spreadsheet, then decide which jobs move to the small model. Set a rule: anything the small model isn't confident about goes to a larger model or a person.

6. Ask your provider or developer about caching and batching

If your AI applications connect through an API, ask about these two:

• Is repeated content cached? Company policies, product catalogs and standing instructions get sent with every call. Major AI providers offer prompt caching; with Anthropic, for example, reading from the cache costs about a tenth of the normal input price.
• Is non-urgent work batched? Month-end reconciliation, bulk product-description rewrites and cleanup of old records don't need instant answers. OpenAI and Anthropic both offer batch APIs that finish within 24 hours at roughly half the price of real-time calls.

What is the Jevons paradox? The coal question of 1865

In 1865, British economist William Stanley Jevons published The Coal Question. Many people assumed that James Watt's improved steam engine, which burned far less coal, would reduce Britain's coal consumption. Jevons pointed out the opposite: because the engines were cheaper to run, more factories, mines and railways used them, and the nation's total coal use soared.

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."

The mechanism comes down to three steps:

1. Efficiency improves: the same job takes fewer resources.
2. Unit costs fall: uses that weren't worth it before now pay off.
3. Uses and volume multiply: the new demand outweighs the savings, and total consumption rises.

This isn't an iron law. When demand is already saturated (switching to LED bulbs doesn't make you install ten more lamps), efficiency really does save resources. The Jevons paradox holds when there's a large pool of demand that went unmet because it was too expensive. AI is exactly that kind of case.

Why AI fits the Jevons paradox

DeepSeek and Nadella's post

In January 2025, Chinese startup DeepSeek released R1, a reasoning model known for its low cost, and markets began to worry that AI might not need so many chips after all. On January 27, Nvidia's stock fell about 17% in a single day, wiping out nearly $600 billion in market value, the biggest one-day loss ever for a U.S. company. In the middle of that turmoil, Microsoft CEO Satya Nadella posted on X: "Jevons paradox strikes again! As AI gets more efficient and accessible, we will see its use skyrocket, turning it into a commodity we just can't get enough of."

Unit prices keep falling, totals keep rising

More than a year and a half later, the numbers largely back Nadella up:

The International Energy Agency's April 2026 report describes the same pattern: AI's energy efficiency per task is improving at a pace unprecedented in energy history, but more people are using AI and energy-hungry uses such as AI agents are spreading, so total electricity demand keeps climbing.

Inside a company: work that never paid off now does

Things you set aside because staff were stretched or the cost was too high become worth doing once AI prices drop:

Each item is cheap on its own, but together they can multiply the company's AI usage within a year.

Heavier use

Usage also grows because AI is used in more demanding ways. With AI agents that look things up and call tools on their own, a single task can mean dozens of model calls. In a post about its research feature, Anthropic reported that agents typically use about 4× as many tokens as a chat, and multi-agent systems about 15×. Cheaper tokens don't automatically mean cheaper tasks.

Spending more isn't necessarily bad. If the extra money buys faster replies, fewer errors and more orders, a bigger budget makes sense. What matters is knowing where the money goes and what it buys, which is exactly what the six steps above are for.

How will your staff's work change?

The paradox has entered the jobs debate too. In April 2026, Apollo chief economist Torsten Slok described a "Jevons employment effect": as AI makes legal, accounting and consulting work cheaper, more people will buy those services, and employment in those fields could grow rather than shrink (as reported by Fortune).

The go-to example is radiology. In 2016, AI pioneer Geoffrey Hinton said we should stop training radiologists now. Nearly a decade later, an analysis in Works in Progress found that despite hundreds of FDA-approved AI imaging tools, radiologists remain in short supply across the U.S., and residency programs offered a record number of positions in 2025.

The other side of the data matters just as much. An August 2026 update from the Stanford Digital Economy Lab found no sign of widespread job displacement across the economy. But in the occupations most exposed to AI, employment of workers aged 22 to 25 is about 19% below where it would be had it kept pace with their peers in less-exposed jobs. The gap comes mainly from companies hiring fewer young people, not from layoffs, and experienced workers show no similar gap.

The most defensible reading: the total amount of work may grow, but its makeup will change, and the first work affected is the routine work that new hires used to learn on. In the projects we've seen, AI takes over steps, not jobs: data entry, formatting and first drafts go to AI, while people review, handle exceptions, talk to customers and make the final call. A few practical suggestions:

• Take stock of tasks, not people: list the tasks in each role and decide which suit AI and which need a person.
• Redesign how new hires learn: juniors used to learn the business through data entry and tidying records; now they can review AI output alongside senior colleagues.
• Give the saved time somewhere to go: faster replies, more orders and markets you never had time for are where the real return on AI comes from.

When to consider the next step

The six steps above only need an AI assistant, a spreadsheet and the services you already use. It's worth looking at the next step when:

• AI is used in more departments and processes, and downloading the bills, sorting usage and pasting it into an AI to total takes longer every month.
• You want to see the cost per task for each department and process at any time, without waiting for month-end.
• You want simple work to go to small models automatically, and anything uncertain to go to a larger model or a person.

The next step: route AI requests through one gateway

The idea is to add an "AI gateway" between your company's applications and the AI services, so every department's AI requests pass through it. What you did above with a spreadsheet and an AI assistant now happens automatically:

1. Quotas: usage is tallied by department and application, and the owner is notified as a quota gets close.
2. Caching and batching: repeated policies and catalogs go through prompt caching; non-urgent work goes into a batch queue and runs overnight.
3. Routing: requests go to a small or large model by difficulty; anything uncertain goes to a larger model, and important decisions go to a person to confirm.
4. Cost per outcome: a dashboard records the cost of each outcome, next to the previous labor cost and error rate.

You can do this in phases. Start with your highest-volume application; the others, and your current AI services and accounts, stay as they are. Once the numbers check out and the team is comfortable, connect the rest. With the cost per outcome in hand, you can tell whether a bigger budget means more valuable work or usage you can trim.

AI will keep getting cheaper, and your company will probably keep spending more on it. That's fine, as long as every dollar buys results you can see. To work out which models your processes need and how to set up routing and cost tracking, start with a free AI Health Check, or look at our AI integration services and system integration. You're also welcome to tell us what you're working on.