There is a lot of noise about AI and very little of it is useful if you run a business with eleven staff and a stock problem. So here is the plain version: which tasks are worth handing to software, which are not, and how to tell the difference before you spend money.
Start with the boring work, not the impressive demo
The businesses getting real value from this are not doing anything glamorous. They are using it to stop somebody retyping the same information into three places, to answer the same twelve questions at eleven at night, and to produce the weekly report that used to eat a full morning.
That is unexciting, and it is exactly where the return is. It is also the easiest to verify, because you already know how long those tasks take.
The four-part test
A task is a good candidate when all four of these are true.
It repeats. Dozens of times a week, not twice a month. Automating something that happens rarely costs more to build and maintain than it saves.
It follows rules. You could write the steps down and hand them to a new employee. If the answer is "it depends, you develop a feel for it", that feel is judgement and it is not ready to be automated.
Being wrong occasionally is survivable. There has to be a human check, or a cheap way to correct it. Anything where a single error is expensive or unsafe stays with a person.
It does not need a relationship. Nobody feels insulted to have a delivery status answered by software. People do feel insulted when a complaint, a bereavement, or a negotiation is handled by it.
What passes the test
First response to enquiries. Someone messages at 10pm asking whether you have an item, what it costs, and when you open. Answering instantly and handing over to a person when the question gets specific keeps the lead warm through the night.
Reading documents into records. Supplier invoices, waybills, delivery notes, application forms. Somebody currently types these into a spreadsheet. That is slow, dull, and where errors get in.
Drafting the repetitive writing. Quotes from a specification, listing descriptions from property details, the weekly summary of sales and stock. Drafted in minutes, then reviewed by a person before it goes anywhere.
Sorting and routing. Deciding which enquiries are serious, which are urgent, and who should see them, so a real buyer is not stuck behind casual browsers.
What fails it
Anything requiring judgement about a person -- hiring, discipline, whether to extend credit. Anything where the relationship is the product, which in most professional services is most of it. Final decisions of any kind. Low-volume work, where the build cost never pays back. And anything you cannot check, because you will not notice when it goes wrong.
Be especially wary of a customer-facing assistant with no handover to a human. Sooner or later somebody arrives angry with a problem the software cannot solve, and being trapped in a loop turns an annoyed customer into a former one.
Questions to ask before you sign anything
What exactly leaves my systems, where does it go, and what is kept? You should get a straight answer in writing. If it is vague, that is the answer.
What does it do when it does not know? The correct behaviour is to say so and fetch a person. Confident wrong answers are worse than none.
Can we run it alongside the manual process first? Anyone confident in their work will agree. This is the single most useful thing you can insist on, because it lets you judge the result on real work rather than a demo.
What happens when it breaks, and who fixes it? Automation that nobody maintains quietly stops working, and often nobody notices for weeks.
What does it cost to run, not just to build? These systems have ongoing costs. Get the monthly figure before you commit.
Is my business too small for this?
Usually the opposite. If you have three staff, there is nobody to delegate the repetitive work to, so it lands on whoever is least busy and gets done inconsistently or not at all. The case for automating it is stronger, not weaker.
But the honest answer for some businesses is: not yet. If the underlying process is broken, automating it makes the mess arrive faster. Fix the process, then automate it. That is not a sales pitch, and we would rather say it before you spend than after.