The tool is not the problem

A smaller company trying generative AI rarely gets stuck on the technical side. Models come with a subscription, the interfaces are simple, and the first demo almost always works. The difficulty comes after: the trial never turns into regular use, because nobody defined what it was meant to replace or how anyone would know it was working.

A project that starts with “what could we do with this?” produces demos. A project that starts with “which task costs us two hours a day?” produces use.

Pick a first case you can measure

The best first cases share three traits: the task repeats, its output can be checked by whoever already does it, and a mistake carries no irreversible consequence. Drafting a report from notes, sorting incoming requests, pulling information out of a standard document — these lend themselves to a gradual rollout.

A first case that touches a customer, an amount or a regulated decision sets the bar too high. You will get there, but not while learning.

  • The task repeats often enough for the gain to accumulate.
  • Its output can be checked by whoever already does it.
  • A mistake can be caught before it reaches a customer.
The question is not whether AI can do the work. It is whether your organisation can check what it produces.

What happens when the data is not ready

Many projects fail upstream of the model. An assistant answering questions about your products needs an up-to-date product base; automatic extraction needs consistent documents. When those live across several files and several versions, the tool inherits the mess.

This is usually the point where the question stops being about AI and becomes a question about data — and rejoins the less spectacular work of centralising and structuring it.

Measure before you extend

Before generalising, you need a baseline: how long the task took, how many rounds of correction it needed, what share still goes through a human. Without those markers, a rollout is a shared intuition rather than a decision.

A smaller company does not need an AI strategy. It needs one use that holds, an honest measurement, and a clear reason to move to the next one.