Why AI projects disappoint, and what has to be true instead

Most AI projects in operations disappoint, and it is usually not the model’s fault. That is what the doctoral research at Aalto University behind this company is about — and it runs from how these systems have to be built to what has to change in a company before any of it makes a difference.

What makes an agent trustworthy

An answer nobody can check is an answer nobody acts on. In operations that is the whole problem: the numbers decide what gets ordered, what gets made and what a customer is promised — and somebody has to be able to see how one was reached.

So the first half asks how such a system has to be built: where the calculations belong, where the agent belongs, and where the decision belongs. It was tested where it is hardest — work in which a number has to be defensible weeks later, to somebody who was not in the room.

Why projects disappoint anyway

The other half is organisational, and it is the harder one. It asks what has to change inside a company before any benefit arrives at all — and why so much AI work fails, not because the technology cannot do the thing, but because the organisation is not arranged to receive it.

What to do about it

The practical version is short: pick one process rather than a programme, prove the thing on it before committing to it, and do not buy a level your organisation is not yet arranged to use.

That last one is the most expensive mistake in this field, and it is almost always made with good intentions.

Why we work the way we do

The levels on What’s possible are that argument in plain language. It is also why we start at the operation rather than at the systems, and why we say what has to exist before an agent is worth having.

None of this makes us right. But it is why we will sometimes tell you that the thing you came to buy is not the thing you need yet — and then help you build what is.