What we actually automate with AI, and what we leave alone

The useful question is never which model. It is which twenty hours a month are going on work software does better, and who will own it afterwards.

What we automate

Work that is repetitive, rule shaped and high volume. Extracting data from documents and putting it where it belongs. Classifying and routing inbound enquiries. Drafting the first version of something a human then edits. Reconciling two systems that should have been integrated years ago.

What we do not

Anything where being wrong is expensive and nobody would notice it was wrong. Anything requiring judgement about a specific customer relationship. And anything where the process itself is broken, because automating a bad process just produces bad outcomes faster.

The part that decides whether it lasts

Handover. If nobody inside the business can explain what it does, change it, or turn it off, it will be quietly abandoned within a year. We write the runbook and train the person who will own it, and we consider that part of the build rather than an extra.

I check these against the original source before publishing, and I will send you the source if you want to see it. I do not quote a number I have not read myself.