How AI actually works for a small business

The mental model: text in, text out

ChatGPT, Claude, and Gemini are text-in, text-out predictors. You give them some words. They predict what words should come next. That is it.

That sounds reductive. The implications run deep.

• It writes well because predicting good writing is exactly what it does.
• It rewrites well because changing tone is predicting the same meaning in different words.
• It summarizes well because a summary is just shorter predicted text.
• It does math badly.The right answer to 47 × 83 doesn't come from predicting what looks like a math answer.
• It will make things upwhen it doesn't know. It will always predict something, even if there is no factual basis.

If you remember one thing: AI is great at tasks where the work is moving words around. It is bad at tasks where being correct about specific facts matters.

Which tool to start with

Three tools matter for a small business owner: ChatGPT (OpenAI), Claude (Anthropic), and Gemini (Google). They look the same from the outside. They are not the same.

• Claude is the strongest at writing in your voice. It refuses to make things up where ChatGPT will guess confidently. For a business owner whose week is customer emails and quotes, this is the right starting tool.
• ChatGPThas the strongest ecosystem — custom GPTs, better web search, the best phone app for voice mode. Right if your week is documents and research.
• Gemini is built into Google Workspace. If you already pay Google for email, you get it for free, and the Gmail and Docs integrations are real. Weaker on craft.

All three cost roughly $20/month for the paid tier. The free tiers are usable but hit limits fast. The most common mistake is paying for two; start with one, use it for 30 days, add a second only if you have a specific task the first is bad at.

For most Muskegon-area service businesses, the answer is Claude Pro. Open it. Paste a customer email you have been avoiding. Edit two sentences. Send. That is usually the lightbulb moment.

What you should never paste in

When you paste text into any AI tool, it travels over an encrypted connection to the provider's servers, gets a response, and the conversation is retained for a period (sometimes used for training on free tiers, almost never on paid). Three categories should never go in:

• Personal identifiers and payment data.Social security numbers, full credit card numbers, bank routing numbers, driver's license numbers. Redact first — "SSN: ***-**-1234" is enough context for the AI and meaningless to a hypothetical leak.
• Privileged or contractually-protected communications. Attorney-client privilege. NDAs. HIPAA-covered information. Client tax returns. These have a sharper edge than general privacy — pasting them can be a professional breach even if nothing bad happens to the data.
• Internal financials and trade secrets. Pricing strategy memos, customer lists with revenue per customer, the recipe for your process. Anything that gives you a competitive edge.

The habit that prevents most problems is one question, asked before you paste anything:

If this exact text were on a public blog post tomorrow, would it hurt me, my customer, or my business?

If the answer is no, paste it. If the answer is yes, redact first, or take that part out of the workflow entirely.

One more thing: hallucinations

When AI doesn't know something, it predicts a plausible answer, not nothing. That's a hallucination. Treat any specific fact in AI output — a name, a date, a phone number, a citation — as unverified until you check. The model is great at the words around the facts. It is unreliable for the facts themselves.

Now go try a tool

That is the brief. The right next move is to open one of the tools and use it with your own data. The Drafter is live; the others land over the next few weeks.