AI for Small Business, Part 1: Where AI Helps, and Where It Does Not

Imagine you have hired a new intern. This intern has read an astonishing amount, never gets tired, writes a clean email in six seconds and will happily draft your newsletter at three in the morning. The same intern, when asked a question it does not know the answer to, will give you a confident answer anyway, in perfect grammar, and will not blush when it turns out to be invented.

Would you hire this intern? Of course. Would you let them sign your contracts, quote your prices or talk to your most sensitive customer without anyone reading the message first? Of course not.

That image is the whole article in a nutshell, and it is the best mental model I know for a small business owner who wants results from AI, not experiments. This is the first part of a weekly series, AI for small business, step by step. Each Monday I take one decision, such as where AI helps, how to use it for sales, for operations and for customers, how to measure the result and how to write a one-page policy, and I try to make it practical and honest. Today's decision is the one that comes before all the others: where does AI help a small business, and where does it not?

The intern, in more detail

Modern AI tools that write and answer in natural language are, at heart, extremely good pattern machines. They have learned from enormous amounts of text how language usually flows, and they generate what is likely to come next. That makes them superb at some jobs and risky at others.

What the intern is good at:

• Writing a first draft. Emails, descriptions, social posts, job ads, meeting summaries. The draft is not final, but the blank page is gone.
• Summarizing. Turning a long email thread, a call transcript or a ten-page document into five bullet points.
• Sorting and labeling. Reading incoming messages and tagging them: sales inquiry, support, invoice question, spam.
• Rewriting. Changing a text to be shorter, friendlier, more formal, or translated.
• Following a pattern many times. Filling the same template with different details, reminding people at the right moment, answering the same five questions in a consistent tone.
• Brainstorming. Giving you twenty options so you can choose one and throw away nineteen.

What the intern is bad at:

• Knowing what it does not know. It can produce a plausible, detailed and wrong answer. This is often called a hallucination, and it is not rare enough to ignore.
• Facts that change or are specific to you. Your prices, your policies, today's rules, your customer's history. Unless you give it that information, it will guess.
• Judgment. Deciding whether a complaint is a misunderstanding or the start of a lawsuit. Choosing which of two good employees to promote.
• Math and exact numbers. It can slip on arithmetic and on details, so numbers need a second look or a proper calculator.
• Taking responsibility. When a customer is angry, they want a person who owns the answer.
• Keeping a secret. Depending on the tool and its settings, what you type may be stored or used in ways you did not expect. You have to read the terms.

None of this makes AI less useful. It makes it a tool with a shape, and a good tool is one whose shape you know.

The four tests: Repeat, Rules, Risk, Review

When a task crosses your mind ("could AI do this?"), run it through four short questions. I call them the four R's, because small business owners remember lists that rhyme.

1. Repeat. Does it happen often? A task that happens once a year is not worth the setup. A task that happens ten times a week is where time comes back.

2. Rules. Can I write the steps on one page? AI helps most when the task has a clear pattern: "when a new inquiry arrives, reply within an hour with these three things." If you cannot explain the task to a new employee in a page, you cannot explain it to a machine either. If your process is a mess, you will be automating confusion at higher speed.

3. Risk. What does a mistake cost? A wrong summary of an internal meeting costs five minutes. A wrong quote to a customer costs money and trust. A wrong message to a patient may cost much more. The higher the cost of an error, the more human review you need, and sometimes the right answer is "not this task."

4. Review. Can a person check the result in under a minute? If checking the output takes as long as doing the task, the tool has saved nothing. If a person can glance at a draft and say "yes, send" or "no, change this," you have found a good fit.

Four yes answers: a strong candidate. Three: a pilot, with review. Two or fewer: leave it human for now, or fix the process first.

This is the same idea as the filter in AI for small business: what to automate first, seen from the opposite side. There, I help you choose the first task. Here, I help you see where the fence is.

A traffic-light map of small business tasks

To make the tests concrete, here is a rough map. It is general and every business is different, so use it to start a conversation, not to end one.

Green: let the intern work, then glance

These tasks are repetitive, low-risk and easy to check.

• Drafting replies to common questions, using your own approved wording.
• Summarizing meetings, calls and long email threads, with the action items listed.
• Sorting and tagging incoming messages or forms.
• Writing the first version of social posts, newsletters, product descriptions and job ads.
• Reminders: appointments, invoices, follow-ups, review requests.
• Turning your rough notes into a clean checklist or a procedure.
• Rewriting something in plain language or translating a simple message, with a native reader checking anything that matters.

Yellow: use it, but a person must check every output

The value is real, so is the risk.

• Anything with numbers. Quotes, financial summaries, forecasts, estimates. Use a spreadsheet or a calculator, and check the machine's math.
• Customer-facing messages in tone-sensitive situations. Apologies, price changes, complaints. Use the draft as a start and make it yours.
• Analysis and recommendations. AI can suggest patterns in your sales or your feedback, but the conclusion is your job.
• Research. It can point you to topics and questions. Verify claims and sources before you repeat them.
• Screening and sorting people. Be very careful: tools can reflect bias, and in hiring, rules about fairness and privacy apply.

Red: keep it human, or use a specialist

• Legal, tax, medical and safety advice. A machine can explain general concepts, but a qualified professional must advise on your situation.
• Decisions about people. Hiring, firing, discipline and pay.
• Firm promises and contracts. A guarantee, a deadline or a price commitment should come from a person who can be held to it.
• Private or regulated information in public tools. Customer health data, financial account details, employee records and trade secrets deserve a careful look at the tool's data terms, or should stay out entirely.
• Crisis communication. If something has gone seriously wrong, people need a person.

Three small stories (illustrative)

These are invented examples to show how the tests work. They are not real clients.

The clinic front desk. A small clinic gets about thirty calls and messages a day asking the same things: hours, insurance accepted, how to prepare for a visit. A reply drafted from approved wording is green: it repeats, it follows rules, mistakes are cheap and a person reads before sending. But answering "should I come in for this symptom?" is red: it is medical advice, and it must go to a clinician. The line between them is clear when you draw it before you start.

The contractor's office. A small construction company wastes hours every week rewriting the same kind of weekly update for owners. A draft built from the foreman's notes, with the dates and numbers filled in by a person, is yellow-green: it saves time, but the numbers and any claim about delays must be checked. Writing the notice of a delay claim or a contract change is red: wording matters legally.

The consultant. A consultant uses AI to turn call notes into a summary and a list of actions (green), to propose three options for a workshop agenda (green) and to research a new industry (yellow: read the sources). Deciding what to recommend to the client, and putting a name on that recommendation, stays with the consultant.

In each story the pattern is the same. The machine takes the repetitive, language-heavy part. The person keeps the judgment and the signature.

The quiet danger: automating a mess

The most common failure I see in small businesses is not a dramatic robot disaster. It is slower and more boring. A team does not have a clear process for handling new leads. Someone adds an AI assistant to reply to them. The assistant faithfully replies to everyone with the confusion built into the process, faster than before. Now there are more replies, some contradictory, and nobody knows who owns the follow-up.

AI amplifies what is there. If your process is clear, it makes it faster. If it is vague, it makes the vagueness louder. This is why the series begins with deciding where AI belongs and moves on to process, ownership and measurement, as in the sales agent guide, where the process had to be written down before an agent could help.

The numbers support the caution. In October 2026, Gartner reported that only 13% of the CIOs it surveyed said their organizations had realized significant or transformational value from AI so far, and relatively few had rules for stopping unsuccessful pilots. I explain what I make of that in this news post. The lesson for a small business is not "AI does not work." It is that enthusiasm without a number, an owner and a stopping rule is how money disappears.

Five habits for a safe start

1. Pick one task, not ten. Choose one that passes the four R's. Treat it as a small project with a start date, an owner and a number to measure, as in the 90-day sprint.
2. Keep a human in the loop. Every output is read before it leaves the building, at least for the first month. Then decide, with data, which outputs can skip review.
3. Write down the rules. One page: what the tool may do, what it may not, what tone to use, when to hand over to a person. The one-page governance idea applies directly.
4. Measure before and after. How long did the task take, how many mistakes, how many hours did it free? Without a baseline you only have a feeling. Part 5 of the series will go deeper on this.
5. Protect private information. Read the tool's terms about data. When in doubt, remove names and sensitive details before you paste anything, and ask a qualified professional about your legal duties.

Run a task through the test

The tool below turns the four R's, and two more checks, into a quick score for one specific task. Think of a task you do often. Tick everything that is true for it today. You will get a plain verdict: keep it human, pilot it with review, or treat it as a good candidate.

Interactive tool

Should this task go to the AI intern?

Loading the tool…

What to do this week

List five tasks from your last week that felt repetitive. Run each one through the four R's and sort them into green, yellow and red. Pick one green task, write down how long it takes now and ask yourself who would check the output. That is the start of a real pilot, and you did not need to buy anything to get here.

If you would like help choosing and running the first pilot in your business, with a clear owner, a number and a stopping rule, that is part of my work on automation and AI through Project Focus. You can tell me about your situation on the contact page, or take the Project Focus diagnostic to see where your business leaks time.

Next Monday, Part 2: AI for sales, follow-up that never forgets. The series page lists every part, and the publishing calendar shows the dates.

What is the one task in your week you would happily hand to an intern, and what would you want to check before it went out? Tell me in the comments.

General information, not legal advice. Check privacy and employment rules with a qualified professional before putting customer or employee data into any AI tool.

Frequently asked questions

Where does AI help a small business the most?

AI helps most with repetitive, language-heavy tasks that follow clear rules and where a person can check the result quickly: drafting replies, summarizing meetings, sorting messages, reminders and first versions of marketing text. Those tasks save time without putting important decisions in a machine's hands.

What should a small business not use AI for?

Avoid relying on AI alone for legal, tax, medical and safety advice, decisions about hiring and firing, firm promises and contracts, and anything that needs private or regulated information to be shared with a tool whose data terms you have not checked. In these areas a qualified person must decide or review.

What is an AI hallucination?

It is an answer that sounds confident and fluent but is wrong or invented, such as a made-up fact, source or number. It happens because the tool generates likely-sounding text rather than looking up verified facts, which is why every important output needs human review.

How do I know whether a task is a good fit for AI?

Ask four questions: does the task repeat often, can you write the steps and rules on one page, is a mistake cheap, and can a person check the result in under a minute? If all four answers are yes, it is a strong candidate for a small pilot with human review.

Join the conversation

Did this help? A quick reaction tells me what to write next.

Useful Made me smile Go deeper on this

Be the first to react.

Comments

Loading comments…

Leave a comment

NameEmail (required, never shown)

What is the one task in your week you would happily hand to an intern, and what would you want to check before it went out? Post comment

Your name and email create a small profile that shows you are a real person. The email is never shown. The first time, I send you a confirmation link, and every comment is reviewed before it appears. Please keep it respectful and leave links out.