Learning to Use AI for Business: A Four-Week Plan for Owners

A realistic four-week plan for a busy owner is 20 minutes a day, one tool and one real task per week. Week one maps your repeating tasks and tries a single assistant on email. Week two moves to customer writing, week three to your numbers, and week four turns the best result into a routine with rules.

The reason this works better than a course is that courses teach features, and owners need outcomes. You will not remember a feature you never used on your own paperwork, but you will remember the afternoon an assistant turned a messy supplier email into a clean reply. For an independent bookshop owner, week one alone might save around 40 minutes on routine enquiries, and a small, measured win like that is what carries the habit into week two.

Follow me on Instagram@sagnikteaches

Choosing the one assistant without agonising

Owners lose the first week to comparing tools. Skip it. For a month of learning, the differences between the major general assistants are smaller than the differences between a thin prompt and a good one. Pick on three practical grounds. Which account do you already have: eligible Microsoft 365 business subscriptions include Copilot Chat at no extra cost, and every Google Workspace plan includes the Gemini app, with Gemini inside Gmail and Docs on the higher tiers. Where do you do most work: if your life is in Outlook and Word, Copilot is closest to hand; if it is Gmail and Sheets, Gemini is. And does the free plan let you upload a file, which week three needs.

Connect on LinkedInSagnik Bhattacharya

Whatever you choose, read three things on its settings page before typing anything: whether your conversations are used to improve the model, how to switch that off, and how to delete a conversation. Plans change often, so check them on the vendor's page that week and write the answer in your log. If the answer is unclear, use the tool only for tasks with no personal data until you know.

Subscribe on YouTube@sagnikteaches

The ground rules before day one

Four rules keep the four weeks safe and cheap. First, use one general assistant for the whole month, ChatGPT, Claude, Gemini or Microsoft Copilot, whichever you already have an account for. Switching tools weekly teaches you menus, not skills. Second, keep customer personal data, passwords and anything under a confidentiality promise out of every prompt until you have read the tool's privacy terms; what to check in a tool's privacy policy shows how. Third, never send AI-written text to a customer without reading it. Fourth, keep a one-page log with three columns: task, minutes before, minutes after.

Block the time like an appointment. Twenty minutes at the same time each weekday is better than a two-hour Sunday session, because the tool comes to mind when real tasks appear. Aim for 15 minutes of doing and 5 minutes of writing down what happened. Over four weeks that is roughly seven hours, about the cost of one afternoon's course, but spent on your own work.

Learning log (keep this open all month)
Date | Task | Tool | Minutes before (estimate) | Minutes after | Edits needed (0-3) | Worth repeating? Y/N

Week one: find the tasks, then try one on your email

Spend the first two days listing every task you do more than twice a week. Write down the plain verb and the time: answering supplier emails, 25 minutes a day; writing the weekly newsletter, 90 minutes; chasing late invoices, 30 minutes a week. Do not judge yet. For the bookshop owner the list has nine items and totals about eleven hours a week.

From the list, mark the tasks that are mostly words, are low-risk if slightly wrong, and repeat. Those are the right starting points. A bookshop's good first candidate is replying to routine enquiries (opening hours, special orders, event bookings), because the answer is mostly facts the owner already knows.

On days three to five, try one prompt on real emails, with names removed. This is the first example, with an illustrative sample output.

Prompt:
You are helping the owner of a small independent bookshop reply to customers. Tone: warm, brief, no exclamation marks.
Here is the enquiry (names removed): "Hi, do you do special orders and how long do they take?"
Facts to use: yes; orders placed before 2pm arrive in 2 to 4 working days; no deposit for books under $30; collect in store.
Write a reply of under 90 words.

Sample output (illustrative):
Hello [first name], yes, we can order almost any book in print. If you ask before 2pm, it usually arrives within two to four working days, and we'll message you when it's here. There's no deposit for books under $30. You can collect in store whenever suits you. Let me know the title and author and I'll get it started.

Read the sample the way a reviewer would. "Almost any book in print" is a claim the owner did not supply, so strike it. "It usually arrives" is fine. The owner's habit of signing off with a name is missing, and the facts were used correctly. The edit count for this one is 1. Over the week, the bookshop owner handles eight enquiries this way, taking about 4 minutes each instead of 9, and ends week one with a number: about 40 minutes saved across the week, and one invented claim caught.

The lesson of week one is that the assistant only knows the facts you give it. A paragraph of facts at the top of the prompt is the single biggest quality lever you will find in the entire month.

Week one day by day

A plan is easier to follow when each day has a job, so here is week one in detail, at 20 minutes a day.

• Monday: write the task list from memory, then check it against your sent folder and calendar for the last fortnight. You will find three tasks you forgot, usually the small repeating ones.
• Tuesday: put a time next to each task and mark the ones that are words, tables or both.
• Wednesday: pick the easiest wordy task, gather three real examples with names removed, and write the facts block.
• Thursday: run the prompt on the three examples, read each reply aloud, and mark every edit.
• Friday: run it on live messages, send the ones you are happy with after your own edit, and fill in the log.

Reading aloud is a deliberate step. Slightly off phrases that look fine on a screen sound wrong when spoken, and those are the ones a regular customer will notice.

A fourth example: a supplier email

The bookshop owner receives a long email from a distributor about a late delivery with three questions buried in a paragraph. The prompt: "Summarise this email in three bullets, list the questions I need to answer, and draft a two-line reply that confirms the new date and asks for a revised invoice." The sample output lists the three questions correctly but confirms a date the email only offered as a possibility. The owner changes "confirm" to "ask whether the 14th is firm". Summaries are where assistants most often turn maybes into certainties, so always check any date or number against the original.

Week two: customer-facing writing with a reusable prompt

This week the task is writing that customers see: a newsletter, product descriptions, event posts. The new skill is not a better prompt but a saved one. Take the best week-one prompt, strip it to a template, and store it in a document called "prompts I reuse". A template has fixed parts (tone, length, banned phrases) and slots for the facts that change.

Template: weekly newsletter
Role: write for the owner of [business], voice: friendly, plain, short sentences, house spelling (colour, organise).
Audience: regular customers, mostly [description].
This week's facts: [3 to 5 bullets: new stock, event, opening-hours change].
Rules: no invented prices, dates or quotes; under 220 words; end with one clear next step.
Output: subject line plus body.

Then do the second discipline: compare the output with a version written without help for the same facts. The bookshop owner writes the newsletter the old way on Monday (about 90 minutes) and uses the template on Thursday for the next week's issue (about 35 minutes including edits). That is a saving of 55 minutes on one task, though one comparison is anecdote, so repeat it twice.

A second example: the before and after

Before, as a rough draft the owner types: "new stock in, poetry table moved, event friday, come along." After the template, the assistant returns three sentences with the event named and a clear invitation. The fix needed: it called the event "a celebrated evening of readings", which is puffery the owner would never say. Add a rule line "no adjectives like celebrated, delightful or unmissable" and re-run. Banning words you dislike is the fastest way to make the voice yours, and the list becomes part of the template.

A third example: a realistic mistake

One week-two failure to expect: the assistant states an opening time that was true last year. The owner almost sent a newsletter saying "open until 7pm on Thursdays" when the shop closes at 6. The rule is that facts like times, prices and dates must come from you in the prompt, never from the tool's memory. Add a final line to every template: "If a fact is not in the list above, leave a [CHECK] placeholder."

Check-in measure for week two: edits per output. Count the changes you make to each draft; most owners start at 6 to 10 edits and settle at 2 to 3 once the template is tuned. If edits are not falling by the end of the week, the facts section is too thin.

Week three: ask questions of your own numbers

Week three moves from words to figures. Most owners hold the answer to a pricing or stock question in a spreadsheet they stopped opening. The skill here is asking a plain question of a small table and then checking the arithmetic yourself.

Export a simple table: last quarter's sales by category, no customer names. For the bookshop: category, units, takings, stock on hand. Paste it into the assistant, or upload the file where the tool allows it, and ask for specific things.

Prompt:
Here is last quarter's sales by category (no personal data). Tell me: 1) which two categories made the most takings, 2) which category has the most stock on hand relative to units sold, 3) three questions I should ask my accountant. Show the sums for 1 and 2.
Sample output (illustrative): the two top categories with their takings, then a line such as "Stock on hand is 6.1 quarters of sales for gift books against 1.4 for fiction."

The habit to build: always ask for the sums, then do one of them by hand. If the assistant says gift books hold 6.1 quarters of stock and your table says 122 on hand and 20 sold, you can check in ten seconds: 122 divided by 20 is 6.1. Any answer you cannot reproduce on a calculator is a hypothesis, not a finding. For spreadsheet-specific help, the ChatGPT with Excel guide goes further than this week needs.

A before-and-after for this week: the owner used to spend a Sunday evening scanning the stock sheet for slow lines. Now it takes 25 minutes to ask three questions, check two sums and write down four lines to mark down. Week-three measure: errors caught. If the assistant produced three statements and you found one wrong, that is a normal rate at this stage and a good reason to keep checking.

Week four: turn the best result into a routine with rules

By now there are three candidates: email replies, the newsletter and the stock questions. Choose the one with the largest minutes saved and the lowest edit count, and make it routine. For the bookshop it is the newsletter, 55 minutes saved a week, which is about 48 hours across a year. Put it in the diary: Thursday, 35 minutes, template open, facts list ready.

Then write the one-page rules, short enough to pin up. This page is also the first thing to show a staff member.

How we use AI in [business] (illustrative)
1. One approved assistant: [name]. Staff use the business account, not personal ones.
2. Never paste: customer names, emails, card or bank details, passwords, supplier contracts.
3. Facts (times, prices, dates, stock) come from us, never from the tool.
4. A person reads everything before it reaches a customer.
5. If the tool gets something wrong, tell [name] and we add a rule to the template.
Reviewed monthly on the first Monday.

Finally, decide the next step deliberately, not by default. There are three honest options. Stay on the free plan if the routine works. Pay for a plan if you hit limits, using who in your team needs a paid AI licence as the test. Or hand the routine to a tool that runs it for you, which is a bigger project that the AI cost guide prices out.

What to skip while you learn

Some activities feel productive and teach little. Watching hour-long overviews of the latest model is one. Building elaborate multi-step prompts before a simple one works is another. Collecting prompt lists from the internet is a third: other people's prompts carry other people's facts, so you spend your time swapping words. Browsing course catalogues from Microsoft or Coursera can help later, once you know what you want to learn, but a month on your own tasks tells you what that is faster than any syllabus.

Skip, for now, tools that promise to run your whole business, connect to everything or act on your behalf. Learn the manual loop first, ask, read, edit, send, because every automation is that loop with the human step removed, and you should know what you are removing. If you later add automation, how to set up human review for AI work explains where to leave the human in.

Bringing one team member in after week four

Once the rules page exists, choose one person who likes the idea and give them one workflow, not the whole toolbox. The bookshop owner hands over the enquiry replies, with the template, the facts block and the rule that the owner reads the first ten. Ask the colleague for the same four numbers on a Friday for two weeks. If their edits per output are close to yours, widen to a second person. If they are far higher, the template is missing something you do by instinct, such as how you apologise, and writing it down helps every future user.

Keep a shared list of facts that change, such as hours, delivery times and prices, in one document the template points to. When a fact changes, change it once. This single habit prevents the most common team-wide error, the old opening time that keeps coming back.

The measures to check in on each Friday

Keep the Friday check to five minutes and four numbers. Total minutes saved this week, from your log. Average edits per output. Errors caught before they left the building. And a yes or no: did I use the tool on a real task every day this week? Write them on one line. After four weeks you will have a table that settles arguments, including your own about whether this is worth doing.

Notice that week three's saving is the largest but is not repeated weekly, because the stock review happens monthly. The honest annual figure comes from the routine that repeats, not the biggest single win. That is why week four picks by weekly repetition.

If your business is not a shop

The shape holds for any owner. A driving school owner's week one might be replying to lesson enquiries, week two writing reminders and a newsletter for learners, week three looking at cancellations by day of the week, and week four routinising the reminders. A florist's week three might be waste by flower type. A photography studio's might be enquiry-to-booking rates by source. The point is that every week uses your own data and your own words, which is what a generic course cannot offer.

Check each week's task against three filters before starting: it repeats, it is mostly words or small tables, and a wrong answer is annoying rather than costly. If a task fails the third filter, such as anything involving legal wording, tax figures or health claims, keep it for later and take advice from a qualified adviser.

What to do if a week goes badly

Some weeks fail. The outputs are bland, or you do not find time. Look at the log: if the minutes after are not lower than the minutes before, the task was wrong, not you. Pick a smaller one. If the outputs are bland, the facts section is thin; add three more specifics and a sample of your own writing for the tool to imitate. If you cannot find the time, shrink the slot to ten minutes, not zero. A short plan that continues is better than a good plan that stops.

After four weeks, if you would like someone to look at your own task list and help choose what to automate next, that is the kind of thing my AI implementation consultation covers; you can reach me through the contact page. Otherwise the plan above is complete: repeat it for the next workflow, one tool and one task at a time.

Questions owners ask before starting

Do I need to pay for anything in week one?

No. Start on a free plan of one general assistant such as ChatGPT, Claude, Gemini or Copilot, and decide on a paid plan only after week two, when you know which task you repeat. Free plans have limits and change often, so check the current terms before you rely on one.

What if I miss a few days?

Carry the week's task over rather than rushing. The plan is built around one real task a week, not a calendar. Two weeks of slow progress still beats four weeks of skimming videos, because each week ends with something you actually use in the business.

Should staff follow the same plan?

Not yet. Finish week three yourself, write the one-page rules in week four, then let one trusted team member repeat the single workflow that worked. Rolling out to everyone before you know the failure modes is how customer-facing mistakes happen.

How do I know I am learning and not just playing?

Each week ends with a number: minutes saved, edits needed per output, or errors caught. If the number is flat for two weeks on the same task, change the task or the prompt, not the tool. Learning shows up as less editing, not more experiments.

Further reads

• Do I Need an AI Consultant? 8 Signs It's Time to Get Help — Signs that self-teaching has reached its limit.
• What Does an AI Implementation Consultant Actually Do? — What outside help covers once the plan is done.
• Common AI Mistakes Small Businesses Make and How to Avoid Them — Mistakes to watch for in the first month.
• Questions to Ask an AI Vendor Before You Sign Anything — Questions for any tool you consider adding after week four.
• How to Prepare Your Business Data for AI, Step by Step — Tidying the information your next workflow will use.
• A Simple AI Risk Register for Small Businesses (With Template) — A simple record of what could go wrong and who watches it.
• AI Tools for Learning and Development: A Small-Company Shortlist — A job-by-job shortlist for whoever runs training in a small firm: skills matrix, training plan, onboarding and checks, using tools you may already have.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: Microsoft Learn overview of Copilot licences (Copilot Chat included with eligible Microsoft 365 licences), Google Workspace pricing and feature comparison (Gemini by plan). Plans change often; check the current terms.