The conversation about artificial intelligence in small businesses tends to skip the only question that matters: which specific task a person is going to stop doing. Without that answer, any project is just an expensive demo.
Use cases that genuinely work today
- Searching your own documentation. Asking in plain language and getting an answer that cites the procedure or the contract it drew it from.
- Classifying and summarising what comes in: tickets, emails, incidents. It saves you the triage, which is pure wasted time.
- Repetitive writing: quotes, minutes, standard replies. With a human check, always.
- First-line support on your website, answering with your own information and handing over to a person when it does not know.
What you need to tidy up before switching anything on
This is the warning almost nobody gives you: an assistant connected to your documents shows every employee everything that employee could already see, including what they should never have been able to see. Before rolling out Copilot or an in-house assistant, you need to review folder permissions and label anything confidential. Otherwise the AI does not create the problem: it drags it into the open all at once.
Where your data ends up
- Is it used to train somebody else’s models? On business accounts usually not, but you need to see that in writing.
- Which country is it processed in? For sensitive data that matters, and there are options inside the EU.
- Can it be deployed on your own server? With today’s open models, for many use cases yes.
How to measure that it works
You pick one task, measure how long it takes today, run the trial with real data and measure again. If it does not save measurable time, you stop. That discipline is the only thing that separates an AI project from yet another subscription.
And then there is automation, which is not AI
A good share of what people ask for as “AI” is better solved with plain old automation: copying data between programs, generating reports or triggering alerts. With tools like n8n, hosted on your own server and with no per-user licence, it comes out cheaper, it is predictable and it never makes anything up.
