How Not to Burn
Millions on AI

Companies of every size are making the same mistake right now. They are burning billions automating processes with artificial intelligence. An MIT study found that 95 percent of these projects show no measurable effect. Only one in twenty makes the leap from pilot to real, value-creating operation.

So why do nineteen out of twenty fail?

Many try brute force. Bigger, supposedly better models, more expensive subscriptions. But the problem remains.

Because it isn't the model. It's what the model gets to see.

Put the smartest person in the world in front of a problem and give them no information about it. They won't solve it — not because they aren't clever enough, but because intelligence without information grasps at nothing. It is exactly the same with artificial intelligence. A model is only ever as good as what it knows about the reality it is meant to decide in. A small, unremarkable, inexpensive, sovereign model with an excellent data foundation easily beats the most expensive frontier model in the world that lacks one.

What actually happens in the 95 percent is not automation but manual labour with a very expensive tool. An employee laboriously gathers the relevant documents, types them out, uploads PDFs, and explains to the model, over and over, what it should already know.

The companies in the five percent do something else. They don't feed their model with random finds. They give it a data foundation that exists permanently — already there before the question is asked. The model doesn't have to search. It already knows.

They build an integrated data foundation. Not a database in the conventional sense, but a structured map of what actually exists in the operation and how it all connects. The AI can navigate this map — picking out one piece of information and moving on from there the moment it needs another that relates to it.

Any service company with enough expertise and the courage to build such a unified, integrated data foundation before putting AI on top of it can finally use AI effectively.

And yet AI can still make mistakes. In the office, a badly phrased word in an email can be quietly corrected before it does any harm.

But whoever wants to govern the physical world does not merely visualise data or draft emails. Here you steer a live operational system, in real time, while it runs. It is like permanent open-heart surgery. There is no going back, no draft to review once more before it goes out. In the physical world everything has to mesh seamlessly. Machine and human have to act together. The decision is already reality the moment it is made.

The person who confirms that decision in the end must be able to make it well-informed. But a substantial part of the information needed for that decision often isn't with them at all. It sits with a partner — an operator whose network borders their own, or a service provider handling a stretch of the route they only see their own piece of. And these are exactly the data partners can rarely simply hand over, whether because of trade secrets no one outside their walls should see, or data-protection rules that forbid sharing from the outset.

So is it impossible to build such an integrated data platform for operations that work in the physical world — that move real people, not just rows in a table; companies that push goods through ports and drive electricity through lines?

Or is it possible to skyrocket the abilities of every single employee with technology? To turn a clerk in front of a dashboard into a leader who sees the disruption coming before it happens and resolves it with a snap of the fingers — and who already knows every consequence of their decision the moment they make it?

Galith makes it possible.

But how does Galith make it possible?

The foundation is a single, continuously updating model of what is happening in a network right now. Part of the data is integrated live; another part is extracted from its sources. Timetable data, sensor data, contract states flow unchanged into the data foundation. Other things have to be derived first — if a camera recognises that a particular vehicle has reached a particular place, that information too is written into the same foundation, with its source, timestamp, and confidence value.

Above the foundation sits the decisive layer. Here Galith does not rely on a model that guesses. The comprehensive data foundation makes it possible to compute decisions mathematically instead of letting an artificial intelligence estimate them. Because a mistake in the physical world is unforgiving. If a system holds a train ten minutes too long because a recommendation was wrong, it misses the connections it was built for, and the effect spreads further through the network than the original delay ever did. That is why at Galith a model never decides on its own.

Its job is a different one. Until now an operator had to press buttons and flip switches to speak to their system. Galith replaces that unnatural form with natural language, making the control of highly complex systems effortless. A language model translates the operator's intent into a structured request to the algorithms above the data foundation, and returns concrete courses of action with their consequences — which the operator confirms with a single click, or refines with further instructions. The model organises. It does not decide. It takes the strenuous, data-gathering groundwork off the human's hands, so they can focus entirely on the one thing that matters most: the decision itself.

Galith calls this principle controlled visibility. And it applies to the data itself. Every record on the platform carries its origin and its access rules down to the level of individual properties. This makes it permanently traceable where a piece of information came from, and access rights can be distributed seamlessly. If a trip is handled through a booking partner, that partner needs only the passenger's name to assign the ride. They don't need the address or the payment details. They get what the task requires, and not a crumb more. This mechanism is what makes cooperation between companies possible at all, without any of them giving up control of their own data.

And what the employee finally sees is never a generic dashboard. The interface is built individually, together with the teams who work with it every day — tailored to their actual working reality. Because even the best information is useless if it arrives in a language the operation doesn't speak.

How can Galith be put to work in your own operation? The first step is a pilot with a clearly defined zone and acceptance criteria agreed in advance — a concrete, verifiable result both sides settle on, as partners, before it begins. If those criteria aren't met, you aren't left holding an unused licence. You don't pay for a promise. You pay for a result you define yourself and can verify afterwards. If it meets the bar and moves into live operation, a fixed annual licence is agreed — so that what would otherwise be variable compute costs stays predictable.