Most writing on this subject is either breathless or apocalyptic, and neither helps you decide anything on a Tuesday. So here is the plain version, for a business turning over a few million with a team who are already busy.
There is a simple test that gets you most of the way. AI is good at jobs where the answer is already somewhere, and someone just has to go and get it, shape it, or check it. It is bad at jobs where someone has to decide something that matters, and be answerable for it.
Anything that arrives in a pile and has to be looked at and put somewhere. Enquiries that need routing to the right person. Invoices that need matching to a purchase order. Certificates and forms from suppliers. CVs into a shortlist for a human to actually read.
This is the least glamorous category and usually the most valuable, because it is the work that quietly eats mornings and nobody has ever counted.
Quotes and proposals from a call and your own rate card. The monthly report from your own numbers. Standard letters, engagement paperwork, job sheets, follow-ups. The kind of writing where a good version already exists and the new one is a variation.
The point is not that it writes better than your team. It is that your team goes from writing to checking, and checking takes a fraction of the time.
Most businesses this size have four or five systems that each hold part of the picture, and a person who spends a day a month copying between them. That job is almost always automatable and almost always underestimated, because the person doing it has stopped noticing how long it takes.
Does this order have everything it needs. Has this supplier sent the right certificate. Is anything in this batch outside the usual range. Machines are patient and consistent in a way people cannot be at four in the afternoon, and they do not skip the boring one.
What did we quote them last time. What does our policy say about this. What did we agree in that meeting. The answer exists, in a folder, and finding it takes twenty minutes and interrupts two people.
Who to hire, who to let go, who gets credit, who gets a price rise. Beyond the fact that we refuse to build these, they are the wrong shape for the technology: the sample is small, the cost of being wrong is high, and someone has to be able to stand behind the decision and explain it.
These systems are excellent at plausible and only mostly reliable on exact. That works for a first draft someone checks. Wherever being roughly right would cost you, keep a person in the loop. A better model does not solve it; deciding where the human sits does.
Not because it cannot do them, but because the maths does not work. Something that happens four times a year, even if it takes a day each time, is rarely worth building and maintaining. Spend the effort on the thing that happens forty times a week.
Three questions, and you can do this yourself this week without buying anything.
In a business your size, the wins are boring. They are the intake, the paperwork, the drafting and the chasing. They are assistants with a narrow remit. Anyone promising the second thing is either selling something or has not worked in a business like yours.
The good news is that boring compounds. Four or five of those jobs, done properly and actually used, is usually a person and a half of capacity back, without anyone leaving.