The unglamorous half of automation
· Most of the automation work that pays for itself is not clever. It is removing the retyping, the waiting, and the missing trail from a process that already works.
Most of the automation conversations I have start in the same place. Someone has read that AI can do a thing, and wants to know whether it can do that thing for them. It is a fair question. The honest answer, most of the time, is that the AI is the least interesting part of the job.
What turns up instead, over and over, is a process that already works. People know what to do and they do it. The cost sits in everything around the work: the same information typed in twice, a status that can only be found by asking somebody, a decision everyone remembers slightly differently. None of that needs a model to fix it. It needs somebody to look at the process honestly and rebuild the plumbing underneath it.
Three shapes keep recurring.
Capture once, at the source. If information is recorded where the work actually happens, most of the downstream retyping disappears and so do the errors it introduces. This is the least impressive item on the list and often the most valuable.
A trail instead of a memory. When the system records the decision, who made it, and when, as a side effect of the normal flow, a disagreement about what happened becomes a lookup rather than an argument.
Searchable, not merely stored. Plenty of organisations already keep everything. Keeping is not finding. Turning a pile of files into something a person can answer a question from is mostly a matter of structure, and it is worth more than another place to put things.
Where AI earns its place. Inside systems like these there are messy edges: free text that needs reading, a document that needs summarising, a first draft that needs writing, a category that needs guessing. That is genuinely good work for a model, and it is worth doing. It works when it sits inside a system that is otherwise boring and deterministic, and when a wrong answer is visible and correctable rather than quietly authoritative. AI makes a good component and a poor foundation.
How you know it worked. A system that landed well stops coming up. The adoption push stops being necessary, because the friction it removed has stopped being a topic, and people spend their time talking about the work rather than the tool. That is a harder outcome to put in a headline than a percentage, and it is the one worth aiming at.
The practical version of all this: before asking what AI could do here, find out where the same thing is being written down twice. That question is free to ask, and it tends to answer the first one.