AI diagnosis

Diagnosis, not advice

The word is borrowed from medicine on purpose. A diagnosis names what is wrong. Advice suggests what might be nice. The distinction matters because most AI reports in mid-sized companies are the second kind: twenty recommendations, no cause, and therefore nothing anyone is obliged to act on.

A useful diagnosis ends in a sentence of the form: this is not working because X. If it cannot, the examination was not deep enough.

What gets examined

• The process, as it actually runs, not as the flowchart says it runs
• The data, specifically whether the same term means the same thing in two departments
• The decision path, meaning who is allowed to say no and whether they use it
• The skills, specifically who can judge an output rather than only produce one

The causes that keep coming back

In practice the answer is rarely technical. Most often it is one of these: nobody owns the process, so no change can be enforced. Or the data means different things in different systems, so any model built on it is wrong in a way nobody notices. Or a supplier built something that works and left, so nobody can change it.

All three are organisational. That is why a diagnosis that only inspects the technology tends to conclude that everything is fine, while nothing works.

Diagnosis and scan

The terms overlap. In our own vocabulary the AI Readiness Scan is the named product that performs this examination and then goes further: it ranks opportunities and produces a roadmap. A diagnosis is the part that establishes the cause.