Most advice about artificial intelligence (AI) in small business is written either by people selling it or by people frightened of it. Neither group is much help when you have twelve staff and a real problem in front of you. So here is the practical version: what it does reliably, what it does badly, and the single test that separates the two.
The test
AI is useful wherever being wrong is cheap and checkable. It is dangerous wherever being confidently wrong is expensive and nobody would catch it. Every good use below passes that test. Every bad use fails it. If you remember one line from this post, make it that one.
Reading and filing documents
This is the strongest case and the least exciting one. Invoices, purchase orders, delivery notes, and application forms arrive as scans and attachments, and somebody types the numbers into your system. AI reads them, pulls out the fields, and files them where they belong.
It passes the test because a wrong invoice total shows up immediately, since the totals stop reconciling. Set it up so a person reviews only the ones the system is unsure about, rather than all of them, and a full day a week turns into about an hour.
Drafting the first version
Quotes, follow up emails, job descriptions, policy documents, meeting summaries. The blank page is the expensive part, not the editing. A first draft you disagree with is still faster than no draft, because correcting a wrong sentence is much easier than writing a right one from nothing.
It passes the test because you read every word before it leaves your business. The failure mode here is not producing an odd draft. It is not reading the draft.
Answering questions from your own material
Your manuals, price lists, warranty terms, and installation guides hold the answers your team currently gets by interrupting your most experienced person. Pointed at those documents, an assistant answers in seconds and names the document the answer came from.
Two conditions make this work. Ground it in your documents rather than general knowledge, and make it show its source so any answer can be checked in one click. Without both, you have built a confident guessing machine.
Checking work
This is the use most people miss. AI is very good as a second reader: comparing a quote against the price list, flagging the one line in a hundred that looks unusual, spotting the missing field before the form goes out. It is a fast junior employee with no ego and no fatigue, and it does not mind checking the same thing five hundred times.
Where it goes wrong
Two categories, and they are worth being blunt about.
- Anything where being confidently wrong is expensive. Final answers on tax, safety, legal wording, medical questions, and anything a regulator will read. The system produces a fluent wrong answer with exactly the same confidence as a right one, and there is no change in tone to warn you.
- Anything you cannot check. If you would not be able to tell whether the output was correct, do not use the output. The problem is not that the technology is bad. It is that you have no way to notice the day it stops working.
There is a third, smaller failure worth knowing. It knows nothing about your business that you did not give it. Ask about your stock levels, your margins, or what you agreed with a customer last March, and it will answer as though it knows. Connect it to your real data, or do not ask it those questions.
How to start
Pick the single task that eats the most hours and where a mistake would be obvious. For most businesses that is document handling. Run it alongside the current process for a few weeks, count the hours before and after, and only then decide whether to do more. Do not start with the most impressive idea. Start with the most checkable one.