The Fallacy of ROI

Somewhere right now a board is discussing AI. A slide shows projected costs. Someone asks about return on investment. The meeting ends with a decision to wait for more evidence, commission further analysis, or run another pilot before committing.

This meeting has happened before. Not a similar meeting, this one. It happened in the 1990s over ERP, in the 2000s over mobile, in the 2010s over cloud. The technology on the slide changes each decade. The meeting does not. That is not inefficiency, it is a structural pattern, and structural patterns have structural causes.

The cause is not that executives are slow or unimaginative. Most of the people in those rooms are experienced and are trying to decide well. The cause is that they are asking a question that sounds rigorous but cannot be answered in the timeframe it is asked. And because it cannot be answered, the decision can be deferred indefinitely without anyone having to say no.

A formula doing a job it was not built for

Return on investment was developed in the early twentieth century to evaluate capital equipment. A machine costs a defined amount and produces a defined output. The calculation is clean.

Applied to a transformational technology, the inputs become discretionary. What counts as the return, over what horizon, against which baseline, including which costs. Two people can produce entirely different ROI figures for the same AI programme, using the same formula, and both be correct. The number is not wrong. The frame is wrong.

None of this is new. The IT productivity paradox was named in the early 1990s, and economists had already observed that the computer age was visible everywhere except in the productivity statistics. The knowledge that the question was broken existed at every wave. The question was asked anyway. Which tells you it is not really about knowledge. It is doing something else.

What the question actually does

It does two things at once. It legitimises, and it brakes.

It legitimises because when a technology arrives that an organisation does not yet understand, ROI gives people a familiar language to engage with it. It makes the incomprehensible discussable. That is genuinely useful, and the people reaching for it are not being cynical. They are reaching for the most institutionally safe framework they have.

It brakes because the same question, asked before the definition is agreed and using a frame built for a different class of investment, structurally cannot be answered in time. So the organisation can engage with the technology forever without committing to a position. The ROI requirement is not a test that investments pass or fail. It is a condition that transformational investment cannot meet in advance, which makes it a permanent deferral mechanism wearing the clothes of an analytical standard.

That is why it is so hard to challenge. Arguing against it sounds like arguing against rigour, against accountability, against the responsible stewardship of capital. You are not challenging a bad question. You are challenging the framework the room is organised around.

When the number was clean and still wrong

The most instructive cases are not the ones where the maths was impossible. They are the ones where it was done cleanly and the structural cost stayed invisible until the environment changed.

In 2011 the Royal Netherlands Army disbanded its entire main battle tank force to save money after the financial crisis. The calculation was defensible. Tanks looked like Cold War relics and budgets were under pressure. What the number did not contain was the institutional knowledge to operate armoured formations, the crew skills that take years to build, and the strategic options that close when a capability is eliminated rather than reduced. By 2015 the Netherlands was leasing tanks from Germany to keep the knowledge alive, and could not deploy them without German approval. When Russia invaded Ukraine in 2022, it could not freely contribute tanks because it had none, and committed to rebuilding the capability at many times the original saving, with deliveries running years out.

A clean ROI calculation that captured what was visible and missed what mattered. The AI debate makes the opposite error: offshoring and the tanks had clean numbers and flawed outcomes, while AI has a complex reality that resists clean numbers and produces paralysis. The root cause is identical. The measurement framework is applied at the wrong level and the wrong moment, and whatever sits outside the frame is invisible until it is expensive.

The infrastructure threshold

Here is the question nobody asked. Nobody calculated the return on the first typewriters, or built a business case for giving every employee an email address. Those were absorbed as infrastructure. The question was never whether, only when and how.

The ROI question marks the exact moment a technology has not yet crossed that threshold. When a board asks for the return on AI, it is not being rigorous. It is revealing where the organisation sits on the adoption curve. The question is the diagnostic. It says: we still treat this as a discretionary project to be evaluated, not as something work will run on.

That is the trap. Projects get evaluated and infrastructure gets absorbed, and the path to absorption runs through comprehension, not calculation. The organisation is treating infrastructure as a project, and a project can be deferred forever.

The question underneath

The structural question is avoided because it is genuinely threatening. Not because the answer is bad, but because asking it seriously means admitting the current model might be inadequate: that the governance structures, decision rights, approval chains, and operating rhythms that define how the organisation works may have to change. The ROI question is not asked instead of that question. It is asked so that question never has to be.

This is the same incompatibility The Kinetic Enterprise describes: an industrial operating model meeting an environment it was not built for, and reaching for a financial instrument to avoid a structural decision. It connects directly to decision latency. The ROI cycle is itself a latency mechanism, a way for the decision to keep travelling without ever arriving.

Change the question

If the ROI question is functioning as a brake rather than an analysis, a better ROI calculation will not help. The move is to change the question.

Not: what is the return on this investment. But: what would have to be true for this technology to be infrastructure in three years, and what does the organisation need to build now to be ready. That is a design question, not a justification question. It shifts the room from defending a number nobody believes to defining a direction everyone can orient around. It is also, quietly, the question that the organisations which built lasting advantage from every previous wave were actually answering, even when they thought they were answering the ROI question.

CEZ Consulting works with executives to change that question, and to build the operating model that can act on the answer.