Where AI is actually useful — and where it is not

A model that can summarise a document is not, by itself, a reason to reorganise a company. The useful question is narrower: what would change if this organisation could do a particular job more accurately, more cheaply, or with fewer people in the loop — and what would go wrong if it did.

AI is often useful where there is already a corpus, a repeated judgement, and a human who remains accountable. Retrieval over an internal knowledge base. Assistance at a well-defined interface. A draft that a specialist will still sign. It is less useful where the organisation has not decided what “good” looks like, where the data is not actually available, or where an agent is asked to act without a boundary.

That is an engineering and architecture conversation, not a product conversation. Model choice, retrieval design, evaluation, and the places the system should refuse are the work. Applying AI indiscriminately is a way of avoiding it.

Cognitecta’s consulting starts there: where AI is genuinely useful in this organisation, and how to build it so that the people who will own it can still explain what it does.