AI isn't the strategy. The business is.

The useful question is never whether to use AI. It is which specific cost, delay or ceiling in the business it removes — and what happens if it removes nothing.

Germar4 September 20262 min read

A monitor showing an abstract node and probability visualisation in warm tones

Every company now has an AI position. Very few have an AI answer to a specific problem they already had.

That order matters. A business with a clear bottleneck can evaluate AI honestly, because there is something to measure against. A business without one ends up buying capability and hoping a use case shows up.

Start from the constraint

Ask what stops the business from doing more of what already works.

For a consumer goods operation the answers are unglamorous and consistent. Producing enough good product content to launch a range without it taking a month. Reading demand early enough to buy inventory correctly. Answering the same forty customer questions without adding people. Finding the next category before it is obvious.

Each of those is a real ceiling with a real cost. Each is a candidate. Notice that none of them is “we should have AI”.

The test is unit economics

Applied well, AI changes one of three numbers: what something costs, how long it takes, or how good it is at the same cost.

If a candidate does not move one of those, it is a demo. Demos are useful for learning and expensive to run in production.

The honest version of this test is uncomfortable, because it rules out most of the things that are fun to build. It also protects the business from a category of project that looks like progress for two quarters and then quietly gets switched off.

Judgement does not delegate

There is a boundary worth defending. AI is good at volume, variation and pattern. It is not good at deciding what a brand should stand for, which category to enter, or what quality standard is acceptable.

Those decisions compound over years and are hard to reverse. They should stay expensive and human.

The useful split: let the machine do the work that scales linearly with output, and keep the work that sets direction.

Build it in, do not bolt it on

The version that fails is the one added later — a tool bought after the operation exists, sitting beside the systems rather than inside them, requiring someone to remember to use it.

The version that works is designed in from the start, so the output lands where the work already happens: in the catalogue, in the forecast, in the reply. Nobody has to choose to use it.

That is a less impressive thing to announce. It is a better thing to own.