The pitch for AI investment is almost always the same. Automate the repetitive work, reduce cost per task, free up your people for higher-value activity. The business case is built on efficiency, and the measure of success is whether costs came down.

That framing is not wrong. It is just incomplete in a way that sets most programmes up to look like failures even when they are working.

An old idea that keeps being true

In 1865, the economist William Stanley Jevons noticed something counterintuitive happening with coal. As steam engines became more efficient, he expected total coal consumption to fall. It did not. It rose, sharply. More efficient engines made coal economically viable across a far wider range of applications. The efficiency gain did not reduce demand. It created it.

He called this a paradox. We now call it the Jevons Paradox, and it has been observed in energy, transport, computing, and every other domain where making something cheaper to do makes it more common to do it.

AI is not exempt.

What it looks like in practice

Give a development team an AI coding tool and they do not ship the same features in half the time. Scope expands to meet capacity. More gets attempted because more can be attempted. The cost per feature falls. The number of features grows.

Automate a customer service function and contact volume does not drop. Friction drops. Interactions that previously were not worth initiating start happening. The cost per contact falls. Volume rises.

Give a marketing team access to AI content generation and they do not produce the same content for less money. They produce more content, across more channels, at a pace that was impossible before.

In every case the economics are the same. Cost per task falls, number of tasks rises, total spend is flat or higher. And the programme that was supposed to reduce costs looks, on paper, like it has not delivered.

Why the measurement is set up to mislead

The standard ROI framework for AI investment assumes a fixed demand baseline. Measure the task before, measure it after, calculate the saving. That assumption is wrong by design.

Efficiency tools do not operate on fixed demand. They change what is economically rational to attempt. When the cost of doing something falls far enough, people do more of it. That is not a failure of the tool. It is a consequence of the tool working.

When the savings do not appear, the programme gets questioned. The business case gets relitigated. Leaders end up defending an investment that is performing exactly as it should, against a benchmark that was never a fair measure of performance.

The paradox as a signal

Here is the reframe. If usage grows after you introduce an AI capability, that is not a cost problem. It is evidence the capability is genuinely useful. The Jevons Paradox is not a warning. It is a success indicator.

The question it surfaces is not why consumption is going up. It is whether the induced demand is creating value. That is a more interesting question, and a more answerable one. But it requires a different kind of measuring.

The headcount version of this

There is a specific version of this argument running through boardrooms right now. AI will drive efficiency, which will reduce the need for headcount. It is being presented simultaneously as a business case and a strategic outcome.

The Jevons framing complicates that directly. If efficiency induces demand rather than containing it, the more likely outcome is not the same work done by fewer people. It is more work, of greater scope, done by the same people with better tools. The productivity floor rises. The ceiling moves.

That is not the same as saying headcount is never affected. Some tasks that get automated do not regenerate in another form, and the skills required to operate at a higher level are genuinely different. Not everyone will make that transition at the same pace, and that deserves honest attention.

But framing AI investment primarily as a headcount reduction story creates the same measurement problem as framing it as a cost reduction story. It sets a baseline the technology is likely to make irrelevant, and judges the programme against outcomes it was not designed to produce. Organisations that frame the investment around capability and value creation will find it far easier to demonstrate that it is working.

What to measure instead

Efficiency metrics are not useless. Cost per task, time saved, throughput: these are worth tracking as indicators. But they should not be the primary frame for whether a programme is succeeding.

What you actually want to measure is the value generated by the additional capacity. Revenue enabled or protected. Time to market. Quality of output. Customer retention. Decision speed. These connect the investment to business outcomes rather than task economics.

There is also a forcing function embedded here. To measure value, you have to define what valuable output looks like before you automate. That means deciding, in advance, what is worth doing more of. That discipline is not a measurement detail. It is the strategy. The measurement question and the strategy question turn out to be the same question.

The programmes that get this right do not just measure better. They make better decisions about what to build in the first place, because they have asked the harder question first.

Leaders who understand the paradox will build programmes that demonstrate value. Leaders who do not will spend their time explaining why the savings did not materialise, even when the underlying investment is performing exactly as it should.

Efficiency is a means. Value is the point. The programmes that conflate them will struggle to make the case for what they have built, even when what they have built is right.