The most useful AI might be the least dramatic.

A sample opinion piece. Scenarios are hypothetical and imagery is illustrative.

Imagine a small distributor whose operations team spends every Friday checking delivery notes against invoices. The company does not need a digital personality. It needs fewer missing references, faster exception handling and a clear record of who approved each change.

This hypothetical brief is less spectacular than “an AI employee”. It is also easier to evaluate honestly.

Find the friction before the model

Walk through the task with the person who does it. Which steps are repetitive? Which require judgement? Where does information get lost? Sometimes a better form or a deterministic rule will solve the problem more reliably than a model.

If language or unstructured documents create the bottleneck, an AI component might help. Keep its responsibility narrow: extract candidate fields, suggest a category, or highlight a mismatch. Let the rest of the workflow enforce the rules.

Measure the whole job

A demonstration can make extraction look effortless. The actual cost includes checking uncertain results, handling unusual documents and fixing mistakes. Evaluate the time needed to complete the whole task, including review, rather than the speed of the first response.

A useful pilot ends with a decision, not just a more impressive demo.

Start with a reversible pilot

Run suggestions alongside the existing process. Keep original documents accessible. Ask reviewers to record why an output was wrong. Agree in advance what would justify expansion, redesign or stopping.

For a team in Nairobi or New York, the same question matters: does the system make an important piece of work easier without hiding its mistakes? That is a quieter ambition than replacing a department. It is often a better place to start.