The attraction of artificial intelligence is understandable.
A system that can answer customers, assist salespeople, interpret technical information or support management decisions appears to offer a relatively direct route to greater productivity.
But AI introduces a problem that conventional information systems can sometimes conceal.
It exposes weaknesses in the underlying knowledge.
If product names are inconsistent, relationships are unclear, terminology varies between departments, rules are implicit or important exceptions exist only in people's heads, an AI system has no reliable foundation from which to reason.
The first AI project is therefore often an information project.
Before asking what the model should say, it is useful to establish what information is authoritative; which terms mean the same thing; which products, customers or processes are related; what rules govern the decision; where exceptions exist; what the system may infer; what it must ask; when it should stop; and when it should escalate to a person.
These are governance questions, but they are also operating questions.
Consider a recommendation system.
Knowing that two products belong to the same category is rarely enough. The system may need to understand use, environment, compatibility, customer preference, exclusions and the circumstances under which additional information is necessary before recommending anything at all.
A conventional database can leave these relationships implicit.
A useful AI system cannot.
This is one reason AI projects can produce unexpected organisational value even before deployment.
Preparing the knowledge forces the business to make its own logic explicit.
Terminology has to be reconciled. Contradictions become visible. Missing information is exposed. Informal rules have to be examined. Exceptions have to be decided rather than remembered.
In that sense, AI can become a diagnostic tool for information architecture.
The important management question is therefore not simply: Where can we use AI?
A better starting point is: Where do we possess sufficiently reliable knowledge for AI to make a useful decision, and what would need to be clarified before we trusted it?
That produces a less dramatic implementation path.
It is also more likely to produce something useful.
Artificial intelligence can extend organisational knowledge remarkably effectively.
It cannot compensate indefinitely for knowledge the organisation itself has never properly structured.