AI for Small Business Owners: A Plain-English Beginner's Guide

Know five things before you start. AI chat tools predict plausible text rather than look facts up, so everything they produce needs checking. Business plans protect client data better than free ones. The quickest wins are everyday writing, reading and summarising. Paid seats cost about $20 to $30 per person a month. And you start with one task, not a strategy.

Everything below expands those five points in plain English, with an illustrative six-person accountancy practice as the running example. Jargon is defined the first time it appears. When you want a step-by-step plan afterwards, getting started with AI in seven steps picks up where this guide stops.

Follow me on Instagram@sagnikteaches

What "AI" means in the tools you'll actually use

When people say AI in 2026 they usually mean a large language model: software trained on enormous amounts of text to predict which words should come next. ChatGPT, Claude, Gemini and Copilot are all built on models like this. That one fact explains most of their behaviour. They write fluently because predicting fluent text is exactly what they were trained to do. They get facts wrong because fluent and true aren't the same thing.

Connect on LinkedInSagnik Bhattacharya

When you type a question, the model isn't looking the answer up in a database, unless the product has been given a search tool or access to your files, in which case it reads those and then writes a response. Without that, it produces the most plausible-sounding answer from patterns in its training. This is called generative AI, because it generates new text (or images, or audio) rather than picking from a list.

Subscribe on YouTube@codingliquids

A two-minute test makes this concrete. Ask a chat assistant with no access to your files, "What does our practice charge for a sole-trader tax return?" A careful tool will say it doesn't know. A less careful answer, and you will see them, reads something like "Typically between $300 and $600, depending on complexity", phrased as if it were your fee list. It has no idea what you charge; it has produced a plausible range from general patterns. Now paste your actual fee schedule in and ask again, and it reads the answer off the page. Same tool, completely different reliability, and the only change is whether it was given the facts.

You already use an older kind of AI without calling it that: the spam filter in your email, the rules that match bank-feed lines to suppliers, fraud alerts on the business card. Those tools sort and predict; they don't write. The difference matters because they fail differently. A sorting tool puts something in the wrong pile. A generative tool can invent something that was never there.

The four kinds of AI product a small business meets

A token is a fragment of a word; AI usage through an API is priced per million tokens, and a million tokens is roughly 750,000 English words. You'll only meet tokens if you go down the custom route, because chat subscriptions are priced per seat and don't include API use.

Most small firms need only the first two rows in their first year. Before buying anything, check what your current software can do: whether AI works with the tools you already use shows how.

One job done four ways: chasing missing year-end records

The four kinds are easiest to tell apart on a single job. Take the practice's most tedious one: reminding clients who still haven't sent their year-end paperwork.

• Chat assistant. A bookkeeper pastes in what's missing for one client, without the name ("bank statements for the last two months, two supplier invoices"), and asks for a friendly reminder. She copies the draft into her email, adds the name and sends it. About three minutes a client, nothing to set up.
• AI inside existing software. Copilot in Outlook drafts the reply from inside the email thread, so nothing gets copied between windows. Check which Copilot features your Microsoft 365 plan includes before relying on this, because some sit behind the paid Copilot licence rather than the free Copilot Chat.
• Automation platform. Every Monday, an automation reads the practice's "records outstanding" spreadsheet, an AI step writes a reminder for each row still marked missing, and the drafts are saved for the bookkeeper to approve. Setting it up takes an afternoon or two, and the platform charges for each step it runs.
• Custom build. A developer builds a client page that shows each client exactly what's outstanding and sends AI-written nudges on a schedule. That's weeks of work and a build fee, sensible only for a firm with hundreds of clients and the same problem every year.

For a six-person practice the first two routes are nearly always right. The third starts to earn its setup time once the reminders run into dozens a month and follow the same pattern each time.

Where it's reliable and where it trips

AI is reliably useful for: first drafts of routine text such as emails, letters and engagement letters; summarising long documents; rewriting for a different reader or tone; pulling structured details out of messy text; turning rough notes into a tidy document; and explaining an unfamiliar topic so you can ask better questions.

It trips on: facts and figures it wasn't given; adding up long columns of numbers; references, citations and clause numbers; anything that happened after its training unless it can search; the specifics of your firm and clients; giving the same answer twice; and judgement calls about people. The limits of AI in a small business goes through each of these with the checks that catch them.

Pulling details out of messy text is the one beginners underrate, and it shows both sides at once. A client emails: "Sold the van in March for about 6k, got a new laptop, think it was 1,200ish, and the mileage is on the spreadsheet I sent you last year?" Asked to turn that into a list for the file, an assistant might return (illustratively):

Asset sold:      van, March, $6,000
Asset bought:    laptop, $1,200
Mileage:         see spreadsheet supplied last year

It's tidy and saves retyping, but "about 6k" and "1,200ish" have quietly become exact figures, and "last year's spreadsheet" is a reference to the wrong year's mileage. Adding one line to the prompt, "mark anything approximate or unclear with [CHECK]", turns both into questions for the client instead of numbers in the accounts.

For the accountancy practice, this means AI is a good first drafter of a client email explaining what records to send for year-end, and a poor source of the deadline itself. It may state a date or threshold that is out of date, or that applies somewhere else, and it will sound just as sure as when it's right. The fix is simple: paste the facts in, and let the AI do the wording.

In a real prompt, the split is visible: the prompt supplies every fact, and the AI supplies only the wording.

You're drafting for a small accountancy practice. Write a short, warm
email to a sole-trader client asking for their year-end records.
Facts to use (don't add any others):
- Date we need the records by: [typed in by staff]
- Still needed: bank statements for the last two months, receipts for
  equipment bought this year, mileage log
- They can reply to this email with attachments
Under 150 words. Plain English. Sign off as [staff name].

An illustrative draft of the kind that comes back:

Hi [client], I hope business is going well! It's that time of year again. To get your accounts finished on time, could you send us the following by [date]: your last two months of bank statements, receipts for any equipment you've bought this year, and your mileage log. You can upload everything securely through our client portal, or simply reply to this email. If anything's missing, don't worry, we can always estimate. Thanks, [staff name]

Two lines need fixing before it goes out. The practice has no client portal; the AI added one because emails like this often mention them, despite the instruction not to add facts. And "we can always estimate" is a promise nobody authorised, which also undercuts the request. Both slips are fluent and easy to miss on a skim, which is why every draft gets read in full.

Your data: where free ends and business begins

• Consumer plans (the free and individual paid tiers of ChatGPT and Claude) may use your chats to improve the models unless you switch off the model-training setting in privacy settings.
• Business plans (ChatGPT Business and Enterprise, Claude Team and Enterprise, Microsoft 365 Copilot, Gemini in Workspace) don't train on business content by default.
• Connected tools see what the user sees. If you link an AI assistant to a shared drive, it can read everything that person has access to, including folders nobody remembered were open.

That third point catches people who've done everything else right. Picture the practice owner connecting an assistant to the shared drive and asking for "a summary of what's changed in our client folders this month". The summary is accurate, and one bullet mentions last year's staff pay review, because an HR folder had been shared with the whole office years ago and never locked down. Nothing left the business plan, but a junior colleague running the same request would have seen colleagues' salaries. Check who can open each shared folder before connecting anything, and fix the sharing rather than hoping the AI won't look.

For an accountancy practice, client financial records and personal details belong only in a business plan, and even then you should check your engagement letters, any professional body guidance you follow, and data-protection law such as the GDPR. If you're unsure, ask your data-protection adviser before staff start pasting client files. What to check in an AI tool's privacy terms lists the clauses to read.

Whatever the plan, one habit is worth building from day one: give the AI the problem, not the person. Compare two prompts for the same job.

BEFORE
Make this polite: "[client's full name], [home address]. Your account
has been overdue since June, $4,210, and your tax ID ending 4471 is
on the file we sent..."

AFTER
Make this polite: "[Client], your account has been overdue since
[month]: [amount]. We need payment, or a call to agree a plan,
by [date]."

The second prompt gets exactly the same quality of wording, and nothing that identifies the client leaves the practice. The staff member puts the name and figures back in when pasting the draft into the email. It takes ten seconds and takes most of the risk out of everyday drafting.

What it costs, in money and in hours

List prices in USD as of September 2026. Individual plans: ChatGPT Plus and Claude Pro are $20 a month each (Claude Pro is $17 a month billed annually), and Google AI Pro is $19.99. Team plans: ChatGPT Business Standard and Claude Team Standard are $25 per user a month, or $20 billed annually, with a two-seat minimum. Microsoft 365 Copilot Business is $21 per user a month on annual billing, with a promotional $18 through 31 December 2026. Copilot Chat is included with Microsoft 365 business plans, and Gemini is built into Google Workspace business plans.

For six people on annual billing, a business chat plan is about $120 a month. Because the practice runs on Microsoft 365, starting with Copilot Chat costs nothing extra.

The sums change for a one-person business. ChatGPT Business and Claude Team both need at least two seats, so a sole trader on a business plan pays for two, about $40 to $50 a month. A freelance interior designer working alone might reasonably take ChatGPT Plus or Claude Pro at $20 a month instead, switch off the model-training setting in privacy settings on the first day, and keep client names out of prompts as shown above. Look at that setting again whenever the app's settings screen changes, since a switch you turned off once is easy to assume is still off.

The hours matter as much as the money. Budget time for each person to learn one or two tasks properly in the first month, and budget checking time on every output. Checking takes longer than people expect at the start and shrinks as prompts improve and staff learn where the tool tends to slip.

A rough first-month budget for the practice, all illustrative: two hours per person to learn the first task, so 12 hours across the team, plus about three hours of the owner's time writing prompts and the usage rule. That's around 15 hours before anything is saved. If routine client replies then drop from ten minutes to four, as in the fortnight below, and the team sends about 30 of them a week, that's three hours a week back. The learning time is recovered in about five weeks, and every week after that is gain, provided the checking habit stays in place.

Words you'll hear in the first month

• Prompt: the instruction you type. Clear, specific prompts with context get better results than clever wording: "write a reminder" gets a generic email, while "write a 100-word reminder to a client who has missed two deadlines, firm but friendly" gets something close to usable.
• Model: the AI engine underneath a product. Many apps let you choose between a faster model and a more careful one. The quick one is fine for tidying an email; switch to the slower one for a 20-page tender response or anything with calculations in it.
• Context window: how much text the model can consider at once, including your documents and the conversation so far. In practice, if you paste a 200-page lease into a long-running chat and then ask about one clause, the earliest material may be dropped or given less weight, so a fresh chat with just the relevant section often works better.
• Hallucination: a confident answer that's made up. Ask where a rule comes from and you may get a plausible document title and page number for a guide that doesn't exist.
• Connector: a link that lets a chat assistant read from another app, such as your email or file storage. ChatGPT now calls these apps, and on business plans an admin decides which ones are switched on.
• Project: a shared space in ChatGPT or Claude that holds the instructions and reference files for a recurring job, so nobody pastes the same background every time (Gemini's nearest equivalent is a Gem, becoming a skill from November 2026). OpenAI is retiring its older custom GPTs, which stop running on 11 December 2026, so set up shared work as a Project rather than a GPT.
• Agent: AI that takes actions, such as sending an email or updating a record, rather than only writing text for you to use. ChatGPT Work, which replaced the earlier ChatGPT agent in July 2026, is one example: it works through a multi-step job across files and apps and can hand back a spreadsheet or slide deck. Start agents on jobs where they prepare something for a person to approve. A sensible first job for the practice: upload three quotes for new practice-management software and ask for a comparison spreadsheet of monthly cost, contract length and what each includes. The owner then checks every figure against the PDFs before deciding anything, because an agent that misreads one price produces a neat, wrong spreadsheet.

A six-person accountancy practice's first fortnight (illustrative)

Week one: the owner goes first. Using Copilot Chat, already part of the practice's Microsoft 365 plan, the owner tries it on their own work: summarising a long professional update, drafting a fee-increase letter, and turning partners' meeting notes into a list of actions with owners. They keep a note of what worked and what needed fixing. Using AI on your own admin first explains why this order works.

Week two: two staff try one task. The task is first drafts of replies to routine client queries, such as "what do you need from me for year-end?". The rules are simple: no client names or figures pasted in, every reply read before it's sent, and facts such as dates always typed in by the person, never supplied by the AI. Both staff log how long each reply takes.

The result, in this illustration: replies that took about ten minutes now take about four, including checking. One draft contained a wrong deadline because a staff member forgot to paste the date in; review caught it, and the practice added "the AI never supplies dates" to its prompt template. At the end of the fortnight the owner decides to extend the task to the whole team and write a one-page usage rule. No new software was bought.

Four beliefs that cost beginners their first month

"I need lots of data first." Not for chat assistants or built-in features; you need clear instructions and the relevant document. Whether you need a lot of data to use AI covers when data volume does start to matter.

"The free version is the same as the paid one." Usage limits, available features and data terms all differ, and the data terms are the reason a business should pay. The way this usually goes wrong: a new starter, keen to help, pastes a client's bank statement into a free personal account to total up the expenses. The answer is fine, but the chat now sits outside everything the practice's business plan protects, and nobody knows it happened.

"Good results need clever prompts." Brief it as you'd brief a capable new junior: who the reader is, what the document is for, the facts to use, the length and the tone. That beats any trick phrase. "You are a world-class copywriter; write an amazing fee letter" produces gushing filler. "Write a one-page letter telling existing clients our monthly fee rises from [old] to [new] on [date] because we've added payroll support; keep it plain, apologise no more than once, and end with who to call" produces something a partner could sign after one read-through.

"If it's wrong once, it's useless." It will be wrong sometimes. The skill is choosing tasks where a mistake is cheap to catch, like a draft email you'll read anyway, and avoiding tasks where it isn't, like figures nobody rechecks. Suppose the tool miscategorises one expense in twenty when suggesting codes that a bookkeeper reviews line by line: that's still useful, because she spots the odd one in seconds. The same error rate on totals that go straight into a client's accounts unreviewed would be a serious problem. Same tool, same accuracy; the difference is whether a person sits between the output and the consequence.

Beginner questions this guide gets asked

Which AI tool should a beginner start with?

Start with the one included in your office software. On a Microsoft 365 business plan that's Copilot Chat; on Google Workspace it's Gemini. Use it for two weeks on your own admin before paying for anything else. If it falls short, trial a business plan of ChatGPT or Claude for a month and compare them on the same three tasks.

Do I need to learn to code?

No. Chat assistants and the AI built into office and accounting software are used by typing instructions in plain English. Coding, or paying someone who can, only comes in when you want AI connected to your other systems through an API or a custom workflow, and many of those connections can now be set up with automation platforms that don't need code.

Can I ask AI tax, legal or HR questions?

You can use it to understand a topic, draft questions for your adviser or summarise guidance you've been given, but don't act on its answers alone. Chat tools can state rules from the wrong jurisdiction, out-of-date thresholds or invented references with total confidence. For anything with legal or financial consequences, check with a qualified adviser before relying on it.

Further reads

• AI Glossary for Business Owners: 50 Terms in Plain English — Fifty more terms explained in plain English.
• How to Learn AI as a Business Owner: A 30-Day Self-Study Plan — A 30-day self-study plan once the basics make sense.
• AI Myths That Stop Small Businesses Getting Started — More beliefs that stop owners starting, and the reality.
• How Many AI Tools Does a Small Business Actually Need? — Why one or two tools are usually enough at the start.
• ChatGPT vs Claude vs Gemini vs Copilot for Small Business (2026) — Compare the main chat assistants for business use.
• Do I Need AI in My Business? A Decision Guide for Owners — Score whether your business needs AI now or later.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: ChatGPT, ChatGPT Business, Claude, Claude Team, Google AI and Microsoft 365 Copilot pricing pages; OpenAI and Anthropic privacy and data-use documentation; Xero JAX and QuickBooks Online AI agent product pages.