Useful AI in a business does four jobs: reading, sorting, drafting and judging.
Each one has a shape, a cost, and a case it reliably gets wrong. Knowing the fourth part is what separates something you can leave running from something that does damage nobody notices.
Reading
Taking information that exists in a form software couldn't use, and turning it into fields. Handwriting, photographs, PDFs, scans, forms people filled in badly.
This is the strongest of the four and it's the one businesses underuse. Anything that arrives on paper and gets typed into a computer is this job.
Where it fails: handwriting unlike anything it's seen, and documents where the layout carries meaning it can't see. A number in the wrong column is still a number. Check totals, not fields.
Sorting
Taking what arrives in no order and putting it where it belongs. Enquiries, applications, tickets, returns, job applications, photographs.
The value here isn't accuracy, it's that the pile is already in order when a person opens it. Even a sort that's right eight times in ten saves the entire first pass.
Where it fails: categories nobody defined properly. If two of your buckets overlap, it'll pick one, and it'll pick a different one next week. Fix the categories before blaming the tool.
Drafting
A first version of something a person finishes. A reply, a summary, a description, the boring half of a quote, notes from a recording.
The point is the blank page, not the finished text. Editing something mediocre is faster than starting, for almost everybody.
Where it fails: it will invent a fact rather than leave a gap. If you didn't give it the price, it'll write a plausible price. Anything factual has to come from you or from a system, never from the draft itself.
Judging
Looking at something and saying which bucket it belongs in, and why. Is this finished to standard, is this claim complete, is this the right part, is this urgent.
This is the one to be careful with and it's also where the largest savings sit, because it's the step that currently needs your most experienced person.
Where it fails: close calls. It's good at the obvious yes and the obvious no, and it's no better than a coin toss in the middle, where a person would ask a question. Send the middle to a human on purpose.
What it costs to run
AI charges per use rather than per month, so the cost follows your volume.
Reading one document costs a fraction of a cent. That's easy to dismiss, and it's the right way to think about it at ten a day. At two thousand a day it's a real line, and it grows exactly as fast as your business does.
Model the running cost at today's volume and at three times today's volume before you build. A build that's worth it now and unaffordable at scale is a worse outcome than not building it.
What it isn't good at
Worth being blunt, because this is where money gets wasted.
- Arithmetic you care about. It'll produce a number that looks right. Use software for math and AI for words.
- Anything needing today's facts unless you hand them to it. Your prices, your stock, your availability.
- Being the only check on something that matters. Money, safety, legal, medical. It can do the first pass and a person signs it off.
- Doing the same thing identically every time. Ask twice and you may get two phrasings. When you need exactly one answer, that's a rule's job.