When AI genuinely adds value (and when it is hype)

A practical lens for deciding where AI belongs in your process, and where a plain rule does the job better.

Published 8 August 2026

A great deal of AI is currently being built into places where a plain rule would have been faster, cheaper, and more reliable. At the same time, organisations leave work on the table where AI genuinely does make the difference. The distinction is reasonably sharp, and it has little to do with technology. We work it out per situation across our other articles.

The rule of thumb: can you write it down as a rule?

If a task can be captured in rules that always hold, use the rules. Invoices above a threshold to a second approver, VAT rate by product group, blocking an order on insufficient stock: none of that is AI work. A rule can be checked, tested, and gives the same answer tomorrow as today. AI gets interesting once the input is messy and the rules exist only as examples.

Where it works in practice

Making unstructured text usable: sorting, summarising, and labelling incoming mail, quote requests, and complaints.
Reading documents whose layout differs per sender, such as invoices and packing slips from dozens of suppliers.
Producing first drafts a person finishes: a reply, a product description, a report.
Searching your own documentation, where people would otherwise ask the colleague who happens to know.

Where it usually goes wrong

Mistake one is putting AI on work that must be exactly right with nobody checking. A language model gives a plausible answer, not a guaranteed correct one. Financial entries, stock movements, and anything going out to a customer deserve a check. Mistake two is adding AI to a process that is not in order. If your data lives in four places and contradicts itself, AI mostly accelerates the production of wrong outcomes.

Three questions before you start

What happens when the answer is wrong? Can someone see and repair it before it does damage?
Do we have enough examples of good outcomes to judge whether it works?
What is the alternative, and how much better does AI have to be to justify the effort?

Start where mistakes are cheap

The sensible first application is one where a miss costs little and the time saved is measurable. Sorting incoming mail, for instance: sorted wrongly means someone moves it, nothing more. If it works, you have both evidence and a feel for where the model is reliable. Only then does moving it closer to the money make sense.

Frequently asked

Our data is scattered. Can we still do something with AI?

You can, but it is rarely the best first step. Organising scattered data usually pays off immediately and improves everything that follows. Start there, and use AI on the part where messiness is inherent, such as incoming text.

Is our company information safe?

That is a design choice, not a property of AI. It comes down to which data you send, to which party, where it is stored, and whether it is used for training. Settle that before anything goes live, and keep sensitive fields out of the call when they are not needed.

Does this replace people?

In practice it mostly shifts which part of the work people do. The preparatory, repetitive part shrinks; judging and deciding remains. Which is also why a review step is rarely redundant.

How do we know it is actually better than what we do now?

By running them side by side. Let the existing process continue for a period and compare outcomes. Without that comparison, any verdict is a feeling.