If you asked us which uses of language models have actually earned their place in the businesses we work with, the answer is not exotic. It is drafting and triage. Two unglamorous jobs that the models happen to be very good at, and that quietly eat a lot of everyone's day.
Take drafting first. A huge amount of operational writing is not original. It is the tenth reply this week to a similar enquiry, the standard chaser for an overdue invoice, the summary of a call that follows the same shape every time. The hard part was never the writing. It was the starting from blank, finding the right tone, remembering what to include. A model is excellent at producing a competent first version of exactly this kind of thing. The person then edits, adds the specific judgement only they have, and sends. The work goes from composing to refining, which is faster and frankly more pleasant.
Triage is the mirror image. Instead of producing text, the model reads it and sorts. Which of these messages is urgent, which is a new opportunity, which can wait. Which of these documents is the exception that needs a second look. It does not decide what to do. It puts the right things in front of the right people in the right order, so attention lands where it should. In the operations we have worked in, exception checks are exactly this kind of triage, quietly flagging the cases that warrant a human eye.
Why these two and not others
Both work for the same underlying reason. They sit next to human judgement rather than replacing it. The draft is reviewed before it goes. The triage is confirmed by the person it routes to. The model handles the part that is repetitive and the person keeps the part that needs a head. Neither use asks the model to be the final word, which is precisely why they are safe enough to actually deploy.
They also share an honest cost profile. A good draft is cheap to check. A good triage decision is cheap to confirm or override. The review time, which is the real cost of any AI feature, stays low, so the time saved is real rather than swallowed by supervision.
The trap to avoid
The trap is asking these uses to do more than draft and triage. The moment a draft sends itself, or a triage decision acts without anyone seeing it, you have left the safe ground and taken on risk that needs a much more careful design. Sometimes that is worth it. Often it is not, and the modest version, draft for a human, sort for a human, captures most of the value at a fraction of the risk.
So when people ask where to point a language model in their operation, we usually point at the writing they redo from scratch and the inbox they sort by hand. Start there. Those are the jobs where the model earns its place, week after week, without any drama.
Facing something similar in your business?
Talk it through with our AI guide, or send the team a note. We will tell you straight whether and how we can help.