Five mistakes small businesses make when they start using AI

An owner is behind on quotes, so they paste a customer's name, address and job details into a chat window to draft the reply faster. Then they stop. Was that okay? The customer never said the details could go into a tool they know nothing about. That pause is worth listening to, because the businesses that struggle early with AI almost always struggle in one of five predictable ways, and each has a simple fix.

This article lays out those five mistakes and the rule that prevents each one. None of them requires technical skill. All of them can be written down before the team touches a tool, which is the cheapest moment to fix a habit.

Mistake 1: Pasting customer data into a chat window

Chat tools such as ChatGPT and Claude are convenient, and that convenience invites people to feed them whatever is on the screen: names, addresses, invoice details, medical notes, payment information. Once that text leaves the business, the owner no longer controls where it sits or how it is used. The larger world takes this seriously. Ohio State, for instance, routes its researchers to language models that were vetted by the university specifically to protect institutional data, according to its Office of Technology and Digital Innovation. A small business deserves the same caution, at its own scale.

The rule that prevents this: never paste anything a customer would be surprised to see there. Strip names, phone numbers, addresses and account details before the text goes in. Use placeholders such as [customer] and [job site]. The draft still works, and the private information stays private.

Mistake 2: Letting AI write to customers unchecked

AI drafts emails quickly and they usually read fine, which makes it tempting to hit send. But the tool can invent facts with full confidence, and the customer will take an invented promise or a wrong price at face value. A small business lives on its reputation, and one confident wrong sentence can cost a job.

The rule: a human reads every AI draft before it goes to a customer, and the same human owns the facts in it. If the draft quotes a price, a date or a scope, the reader checks it against the real job. This is not a trust problem with the team. It is a standing step, like counting the drawer at closing.

Mistake 3: Buying tools before picking a task

Software subscriptions are easy to buy and easy to forget. Teams sign up for a handful of AI tools, each with a monthly fee, and none of them gets used because nobody decided what work the tool was for. The money leaves, the receipts stay, and the owner concludes AI was hype.

The rule: pick the task first, the tool second. Write down two or three tasks that repeat every week, such as follow-up emails, intake forms or quote drafts, and start with one. A general chat tool such as ChatGPT or Claude covers most starting tasks, and only a proven need justifies adding another subscription. The order matters, because the task tells you what to buy and, just as often, what not to buy.

Mistake 4: No rules for the team

When an owner uses AI privately and never tells the team, each person invents their own habits. One person pastes customer data everywhere. Another never touches the tools. A third sends raw drafts to customers. The business gets the worst of all approaches at once, and the owner finds out only when something goes wrong.

The rule: write the rules down before the team starts, and keep them to one page. Cover what data may go in, who reviews what goes out, which tool the team uses and what stays off limits. Ground rules written in plain language do more work than a long policy document nobody reads.

Mistake 5: Giving up after one bad result

The first attempt at a prompt often produces something useless, and the conclusion follows fast: this does not work for our kind of business. That conclusion is usually wrong. A vague prompt gives a vague answer, and a second, more specific prompt often gives something usable. The teams that get value are the ones that kept going past attempt one.

The rule: judge AI on a run of attempts, not on one. Give a task three or four tries with more context each time, then decide. Write down the prompt that finally worked so the team reuses it instead of starting over. What survives that process becomes a routine, and routines are where the actual time savings live.

How to put the five rules in place

The five rules fit on a single page, and the whole setup takes less than an hour. Here is the sequence.

Ohio has the technical ground to stand on here. The Ohio Supercomputer Center has served the state for almost 40 years and Ohio State offers vetted access to major language models for builders. The infrastructure around AI in the state is serious. The difference for a small business is not access. It is habits, and habits are set in the first month.

Questions owners ask before starting