The “problem-first” trap with AI

Why are leaders questioning AI investment decisions?

A question came up recently on an industry network I'm a member of. A peer asked what IT leaders were doing to prevent AI initiatives from being launched without clearly defined business problems and an identified return on investment (ROI).

It's a question I've heard more than once and at first glance, it seems entirely reasonable. We should be cautious of expensive AI programmes that are searching for a justification rather than solving a genuine business need.

But the more I thought about it, the more I realised it wasn't the answer that bothered me. It was the question itself.

The question assumes that every technology initiative should begin with a clearly defined business problem and measurable ROI. It treats that as established discipline and asks how leaders can enforce it. I don't think that's right, and I want to explain why.

AI is different from traditional technology projects

When there are clearly bounded problems with known solutions, a problem-first framing is the right approach.

Traditional technology projects, such as finance system replacements, tenant portals, and workflow automation, all have clear scope and forecastable returns. For this class of investment, NPV and ROI work well because the value is predictable: a period of build cost followed by a roughly stable stream of benefit. Most organisations built their governance mechanisms for this shape, and those mechanisms work.

AI is a General-Purpose Technology (GPT), not a point solution. The use of GPTs tends to follow a very different value curve.

Costs rise as firms invest in the infrastructure and services required to make the technology useful: process redesign, retraining, new managerial practices, platform build, data and workflow accumulation. Value rises later, often steeply, once that intangible investment matures. The value profile is typically more back‑loaded and uncertain than traditional delivery, which is more forecastable from the outset.

The intervening thirty years were spent figuring out that you couldn’t simply bolt electric motors onto a production line. The real gain came from redesigning the factory from the ground up. A generation of intangible investment, invisible to the productivity statistics of the time, eventually produced one of the largest productivity booms in industrial history. The internet showed the same pattern at the enterprise level through the mid-90s, even as consumer adoption ran ahead.

The governance challenge organisations often miss

I believe the original question is a category error. It assumes the discipline that’s appropriate for solution investment is the same as that required for capability investment, yet the two behave differently.

Worse, it tends to produce fake problem statements reverse-engineered to satisfy a governance gate. The question I think we should be asking isn’t how to prevent AI initiatives without defined problems and ROI; it’s actually how we should govern a capability investment rather than a solution investment.

It’s a different type of discipline, maybe, but it is still a discipline. This is a more honest answer to the original concern. Most organisations don’t yet have a governance model that distinguishes between them.

Measuring AI projects by capability rather than ROI

The value from AI initiatives is different. The early months of delivery can look like a loss.

This isn’t because there won’t eventually be a return; it’s because what is being built at the start are the foundations. The outcomes at that time aren't ROI targets. Instead, the initial outcomes should focus on things like data-handling discipline, cost transparency, and, of course, learning and behavioural adoption.

You put in a stake you can afford to lose, and you set rules that stop if the value doesn’t materialise later. The real problem in most organisations is that governance doesn’t treat soft outcomes the same way it treats ROI.