Where should a small business start with AI?

The most expensive mistake with AI is starting with "let's do an AI project". The approach that works is the opposite: pick one tedious job and shorten it.

Pick the right job first

A good candidate has three properties:

1. It repeats often. A job done once a month isn't worth automating.
2. It works on text or documents. This is where language models are strongest.
3. Being wrong isn't catastrophic. The output can be checked by a person.

Where these three don't hold, AI usually adds cost and risk without adding value.

Five uses that pay off in smaller companies

Classifying incoming enquiries. Dozens of emails a day — which are quote requests, which are support, which are invoice questions? Automatic labelling and routing is one of the fastest-returning uses there is.

Extracting data from documents. Invoices, delivery notes, order forms. Pulling line items out of a PDF by hand is slow and error-prone.

Summarising long records. Meeting recordings, long email threads, client notes. Producing summaries is one of the biggest time sinks for people.

Answering from your own documentation. A team repeatedly asking the same questions. An assistant that answers from your own product docs helps both staff and customers.

Drafting content. Product descriptions, listings, email drafts. A person writes the final version, but not from a blank page.

How to start: a small trial

Before committing, run a two-week trial:

1. Collect 50 real examples of the job (not invented test data).
2. Run them through the model and check the output by hand.
3. Count how many were right and how many were wrong.
4. Work out the cost per use.

After those four steps you can decide. If accuracy isn't good enough, the project stops there and you've lost nothing.

Three things to design in from the start

What happens when it's wrong? There has to be a path that hands the work to a person when the model isn't confident. Critical decisions are never left to the model alone.

Where does your data go? Enterprise APIs don't feed data into training, but that setting has to be configured explicitly. If sensitive data must never leave your environment, that part can run on your own infrastructure.

How is cost controlled? Work out cost per use up front and set a monthly ceiling. Measuring real usage during the trial means no surprises in production.

A pattern worth avoiding

"Let's put a chatbot on the website" is the most commonly started and most commonly abandoned project. The reason is simple: most visitors don't ask a bot anything, and those who do are usually looking for contact details.

Pick an internal job your team repeats every day instead. The benefit is easier to measure and it doesn't put the customer experience at risk.

---

We can work out together which job fits. See AI integration or write to us.