AI is the Means Not The End

Many AI projects begin with the technology and go looking for a problem. A board asks for an AI strategy. A vendor offers a pilot. A team builds a demonstration that impresses everyone in the room and changes nothing in the business. Six months later, the pilot is quietly shelved, and the organization concludes that AI does not work for it.

AI did not fail. The project started at the wrong end. We start from the other end: the decision that needs to get better.

Start with the decision

Every organization has a handful of decisions that shape its year: what to produce, what to buy, where to invest, whom to hire, what to price, which risk to accept. Each depends on information that arrives on time, or does not; that is accurate, or is not; that people trust, or argue about.

Name the decision first. Then ask what it needs: which information, how fresh, how accurate and in whose hands. Only then ask how to provide it. Sometimes the answer is a model. Often it is not.

Four questions before any model

We ask every client the same four questions, in order.

• What are you working on? Not the AI ambition. The work: the product, the service, the operation, and the decision inside it that matters most this year.
• What do you use today to collect information, make calculations and make decisions? Spreadsheets, an ERP, a weekly meeting, a phone call to the plant. This tells us what can be built on and what has to change first.
• What are you missing? The number that arrives too late, the view nobody has, the question nobody can answer without a week of work. This is where value lives, and it is often data, not models.
• What is the biggest pain point? The one problem that, solved, changes how the team works. We start there, with one decision, one system and a fixed scope.

The assistant on our consulting page asks the same four questions, so the first conversation with a principal starts where it should.

Often the answer is simpler

A late report is frequently a process problem: an input that arrives after the deadline, a handoff nobody owns. Numbers nobody trusts are usually a data problem: two systems that define the same thing differently. Two teams arguing over different figures is a governance problem. A forecast nobody uses is a trust problem.

None of these needs a model. Fixing them first makes any later AI work faster, cheaper and more reliable, and sometimes it makes the AI unnecessary. We tell clients when that is the case. An engagement that ends with a cleaner process and no AI is a success if the decision got better.

The simplest model that works

When a model is the answer, the best one is the simplest that meets the need. When Bartosz led treasury cash-flow forecasting for the Central Bank of Somalia, working with the International Monetary Fund, the evaluation criteria were fixed before any model was trained, and four families of models were compared on the bank’s own treasury data. The winner was a classical statistical model, SARIMA, which reached 96 percent accuracy on held-out data at a six-month horizon. It was not the most fashionable choice. It was the one that worked, that the bank’s own engineers could run and maintain, and that its leadership could understand.

That is the pattern we look for. A method the people who own the decision can understand, check and keep running is worth more than a more impressive one they cannot.

When AI earns its place

AI earns its place when four conditions hold.

• The decision repeats often enough that improving it pays back the effort.
• The data exists, or can be collected, at the quality the decision needs.
• A person can check the result against reality, and does.
• The people who own the decision trust the method enough to act on it.

Forecasting demand or cash, reading documents at volume, flagging the transaction that looks wrong, matching records across systems: these are frequent, data-rich and checkable. They are good candidates. Strategy, negotiation and judgment about people are not.

Measure the decision, not the model

Model accuracy is a means too. The measure that matters is whether the decision got better: fewer stockouts, a faster close, cash planned with more confidence, fraud caught earlier, less time spent arguing about numbers. Agree on that measure before work starts, and report against it after. If the decision improved, the AI did its job. If not, it was a means to nothing, however accurate it looked.

What we promise

Clear numbers, sound systems and, where it earns its place, AI. One decision, one system and a fixed scope to start. The goal is a better decision your team can defend, and a system they can run without us.