AI helps a small business most with the writing and reading you already do

AI is a writing and reading assistant, not a growth engine

The honest answer is narrower than the hype. Tools like Claude and ChatGPT are very good at two things: turning messy input into clean text, and pulling the important parts out of long text. For a small business, that means drafting replies to inquiries, writing up quotes and scopes from your notes, summarizing email threads, rewording policies so customers understand them, and sorting a pile of leads by what each person actually asked for.

What AI does not do is just as important. It does not bring in customers on its own. It does not know your prices, your availability, or your refund policy unless you tell it, every time or once in a saved note. It does not do arithmetic you should trust without checking. It does not know that Mrs. Patel always wants the 8am slot.

So the realistic win is not that AI runs your business. It is that it shaves three or four minutes off dozens of small writing jobs a week. For an owner who does their own admin in the evenings, that adds up to real hours. If you are still working out what these tools are in the first place, start with the three ways a small business can use Claude, then come back here.

The rest of this guide is about picking which jobs to hand over. Most people get that part wrong. They start with the flashiest idea, like a chatbot on the website, instead of the dullest and safest one, like reply drafts they read before sending.

The check-time test tells you which jobs to hand over

Every piece of AI output needs a human to look at it before it reaches a customer. That is not a temporary limitation you can wait out. It is how you should run it. So the real cost of an AI job is the time it takes you to check the result, plus what it costs when you miss something.

That gives you a simple test with two questions:

1. Can I check this in a fraction of the time it would take me to write it? If writing takes five minutes and checking takes twenty seconds, hand it over. If checking means redoing the work line by line, you have saved nothing.
2. If a mistake slips through, how bad is it? A slightly stiff sentence in an email is cheap. A wrong total on a quote, a wrong date on a booking, or a wrong dosage instruction is not.

Run some everyday jobs through it.

A customer emails: "Are you free on June 14 for a two-hour shoot, and what do you charge?" Writing a warm, complete reply takes a few minutes. Checking an AI draft takes seconds: you glance at the date, the price you supplied, and the tone. That passes.

A quote for a kitchen remodel with 30 line items is different. If the AI adds them up, you have to recheck every line to trust the total, which is the same work as doing it yourself. And a wrong number either costs you margin or costs you the job. Let software with real line items do the math. Let AI write the plain-English description of the work above the numbers.

Summarizing a 40-message email chain with an insurance adjuster passes too. You know the case, so you can tell in a moment whether the summary missed something. Writing your cancellation policy from scratch only half passes. AI can draft it, but you need to read every word, because customers will hold you to it.

Most owners find that their best AI jobs are boring: replies, follow-ups, rewrites, summaries. That is the point. Boring and frequent is where the minutes are.

What that looks like in four kinds of business

Here is how the test plays out in businesses that do not employ a marketer. The common thread is the same in each: you feed in something you already have, and you get back a draft you can check fast.

A wedding photographer

• Inquiry replies. Feed it the couple's email, your package prices, and your available dates. You get back a reply in your voice, ready for a final read.
• Shot lists. Paste the client's questionnaire answers and ask for a shot list grouped by part of the day. You will spot gaps instantly because you know weddings.
• Gallery delivery and review-request emails. Write the template once with AI's help, then personalize each one in a minute.

Do not have it write the contract itself. Start from a template a lawyer has reviewed. AI can write a plain-language summary to send alongside it, which you check against the real document.

An auto body shop

• Translating technician notes. "LF fender repl, blend door, R&I bumper cover" becomes two clear sentences a customer understands. This is one of the highest-value uses there is, because confused customers call, and calls take longer than emails.
• Delay updates. "Parts are back-ordered until Thursday" gets turned into a polite, specific update instead of silence.
• Insurance threads. A summary of what the adjuster approved, what is disputed, and what you are waiting on.

A small clinic or practice

• Patient instructions in plain language. Rewrite pre-appointment or aftercare instructions at a reading level most people follow easily. A clinician still approves the final version once, and then it gets reused.
• Scheduling replies. Questions about parking, forms, and what to bring are a good fit.

Clinical advice to a specific patient fails the test. So does pasting patient details into a consumer chat app. More on that below.

A general contractor

• Scope of work from site notes. Dictate notes on your phone at the job, paste the transcript in, and get back an organized scope. You were there, so checking it is quick.
• Change-order letters. Clear wording about what changed, why, and what it costs. You supply the cost.
• Comparing supplier quotes. Paste two quotes and ask what each one includes that the other does not. Then verify the differences it flags against the actual documents.

Where AI costs you more than it saves

The jobs that fail the check-time test are usually the ones getting the most attention. These are the ones to skip or handle very carefully.

Unsupervised chatbots that answer policy or price questions. A chatbot that talks to customers with nobody checking can promise things you never offered. This is not hypothetical. In Moffatt v. Air Canada (2024), a Canadian tribunal held the airline responsible after its website chatbot gave a customer wrong information about bereavement fares. The airline argued the chatbot was responsible for its own words, and the tribunal rejected that. If your bot says it, you said it. A small business has less room to absorb that than an airline.

Reviews and testimonials. The US Federal Trade Commission's 2024 rule on fake reviews bans fake consumer reviews and testimonials, explicitly including ones generated by AI, and it allows civil penalties. Using AI to help you write a polite request asking real customers for a review is fine. Using it to write reviews is not.

Mass-produced website content. AI can write fifty blog posts in an afternoon. They will read like everyone else's fifty posts, because they are built from the same average of the internet. A single page written from your own experience, such as what a real job cost and why, does more for you.

Numbers you will act on. Totals, taxes, payroll, deposits. These belong in software that calculates, not software that predicts the next likely word.

Sensitive data. If you run a clinic in the US, patient information is covered by HIPAA. Do not paste it into a personal chat subscription. If you want AI to touch patient data at all, ask the provider whether it will sign a Business Associate Agreement for the specific plan you would use, and get that in writing before you start. For other businesses, a reasonable house rule is to leave out card numbers, ID numbers, and anything you would not put in a normal email.

What a month of AI help actually costs

Most people assume AI help means a $20 or $30 monthly subscription per person. That is one way to pay. The other is to pay the AI company directly through an API key, by usage. For the writing jobs above, usage costs are tiny. Here is the math, so you can redo it with your own numbers.

AI usage is billed in tokens. A token is roughly three-quarters of a word. At the time of writing, Anthropic lists Claude Sonnet at $3 per million input tokens (what you send it) and $15 per million output tokens (what it writes back). Check the pricing page linked below before you rely on these figures, because they change.

• One inquiry reply: about 1,500 tokens in (the customer's email, your price list, your tone notes) and about 300 out. That works out to $0.0045 + $0.0045, or roughly one cent.
• One quote write-up from site notes: about 2,000 in and 800 out, which is about 2 cents.
• One long email thread summarized: about 6,000 in and 400 out, which is about 2.4 cents.

Now picture a busy one-person business. It sends 20 replies a day on 22 working days (440 a month), writes 40 quotes, and summarizes 20 long threads. That comes to about $3.96 + $0.72 + $0.48, or roughly $5 a month. Anthropic's smaller Haiku model is listed at $1 in and $5 out per million tokens, which would bring the same workload to under $2.

The point is not that AI is free. The point is that, for this kind of work, the AI itself is rarely the expensive part. The expensive part is whatever someone stacks on top of it. For the full range, from a few dollars to heavy use, see what Claude actually costs a small business. If the idea of an API key is new to you, what you actually pay for with an API key explains it.

Where the AI lives matters more than which AI you pick

Owners spend a lot of time asking whether Claude or ChatGPT is better. For drafting replies and summaries, both are good, and the difference between them matters less than one other thing: the copy-paste tax.

If your AI lives in a separate chat window, every job starts the same way. You copy the customer's email, paste in your prices, remind it of your tone, copy the result back, and fix the formatting. That might take a minute each time. At 20 replies a day, it eats a big share of the time you were trying to save. It is also where mistakes creep in, such as pasting last week's prices.

There are three realistic setups:

1. A chat subscription. Claude Pro or ChatGPT Plus at around $20 a month. This is fine if you are one person, your volume is low, and you also use it for other things. How many ChatGPT seats a small shop really needs covers this, and most need fewer than they think. Keep in mind that the subscription does not include API usage: Plus and the API are billed separately.
2. AI bundled into a tool you already use. This is convenient because the AI can see the customer record. But it is often priced with a markup on top of the underlying usage, or behind a higher plan tier, and the AI goes away if you switch tools.
3. Your own API key plugged into your tools. You pay Anthropic or OpenAI directly at their published rates, the AI works where your emails and quotes already are, and the key stays yours if you leave.

Fynl takes the third approach. Its CRM, Gmail-connected email, quotes, and automations run on your own Claude or ChatGPT key, with no markup on usage, for $49 a month. Be clear on the gap, though: it does not do SMS or phone calls yet. If most of your customers text you instead of emailing, AI that drafts email replies will miss most of your conversations. A platform with built-in texting, such as GoHighLevel, may fit you better, even after you account for how it bills AI.

What I would do this week

Skip the big rollout. Spend two weeks finding out which jobs actually save you time.

1. Log your writing for three working days. Every email, quote description, follow-up, and policy rewrite, with a rough number of minutes next to each. Most owners are surprised by how much of the day this is.
2. Mark the ones that pass the check-time test. You could verify the draft at a glance, and a slip would be cheap. Cross out anything involving totals, clinical advice, or legal terms.
3. Write a one-page house notes document. Include your prices, availability rules, cancellation policy, how you sign off, and three emails you have sent that sound like you. This page is what makes AI drafts sound like your business instead of a generic one. You will paste it in or save it in whichever tool you use.
4. Run your top two jobs through AI for two weeks. Read every output before it goes out. Note when you had to rewrite more than a sentence.
5. Keep what saved time and drop the rest. If a job needed heavy rewriting more than now and then, it failed the test. Go back to doing it by hand.

Only after that is it worth deciding how to pay: a chat subscription, a bundled tool, or your own key. By then you will know your real volume, and the math in the cost section above becomes your own numbers instead of an example.

Questions people also ask

Sources

• Anthropic: Claude API pricing
• OpenAI: API pricing
• Moffatt v. Air Canada, 2024 BCCRT 149 (CanLII)
• FTC: Final rule banning fake reviews and testimonials (August 2024)