Learn AI for Business Owners: A Four-Week Plan Built on Real Jobs

Pick one real job each week, practise on it for 30 minutes a day, and finish week four with a written decision on what to keep. In week one use AI to draft replies, in week two to quote, in week three to tidy admin, and in week four to measure and decide.

This is learning by doing the work you already have, not by studying tools. A roofing firm owner who spends 30 minutes a day rewriting her own enquiry replies with an assistant learns more than from a week of videos, because every prompt she writes is one she will reuse. Below is the whole plan, filled in for one owner, so you can swap in your own jobs.

Follow me on Instagram@sagnikteaches

The rules of the plan: 30 minutes, one job, a written note

Four rules keep a busy owner on track.

Connect on LinkedInSagnik Bhattacharya

• One job a week. Choose a job you do at least three times a week, so you get fifteen or more attempts before the week ends.
• Thirty minutes a day, five days. Put it in the diary at the same time. Early morning or just after lunch survives better than "when I have a gap".
• Real work only. Use this week's actual email, quote or invoice, not an exercise. The point is that the output goes to a customer or into your books after your checking.
• A note every Friday. Five lines: what I did, what worked, what failed, the best prompt, minutes saved on one task. It becomes your evidence in week four.

A caution on privacy before you start. Use a business plan, or switch off training in the settings of a personal plan, before you paste customer details, and leave out anything you would not email to a stranger: bank details, medical information, full addresses. The tutorial on checking an AI tool's privacy policy and terms shows what to look for.

Subscribe on YouTube@sagnikteaches

Meet the owner whose weeks we follow

An illustration, not a real person: the owner of a seven-person roofing company. She handles the web enquiries, writes every quote, chases unpaid invoices and does the paperwork each Friday. Her week has about 20 enquiries, 8 quotes, 6 invoices to chase and a pile of job notes. She has never used an AI tool beyond a few experiments. She picks four jobs from her real week and lists them in order of how often they arise and how low the risk is.

Choosing your own four jobs from a messy week

The roofing owner did not guess. Spend one evening before week one writing down everything you did in the last five working days, in a plain list, and tick each item against three questions. Do I do it at least three times a week? Is the output mostly words or numbers I can supply? Can I check it in under two minutes? Anything with three ticks is a candidate.

Here is an illustrative list for an owner of an electrical contracting business, to show how the sorting works.

Three clear candidates emerge: quote replies, quote follow-ups and job summaries. That owner should keep the regulated paperwork out of the plan until the basics are comfortable, and ask an adviser about what an assistant may draft there.

The setup hour before week one

Spend an hour, once, on the things that otherwise interrupt week one. Create accounts on one assistant (use one, not three, so your practice builds up in a single place). Open its settings and look for the control that stops your chats being used to improve the model; vendors name it differently and the label changes, so read the setting rather than trusting a screenshot. Create a folder on your computer called "AI prompts" where the Friday notes and saved prompts will live. Pick a time and block it in the diary for all four weeks now.

Then decide what stays out of the tool. A short rule written on a sticky note works: no passwords, no card or bank details, no medical or other sensitive personal data, no full customer addresses. Anything you would need to ask permission to share belongs outside the chat, or in a business plan after you have read its data terms. If you are unsure how an assistant stores what you type, read its data-use page before day one, not after.

Week one: reply to real enquiries, and learn to give instructions

The goal this week is not speed. It is learning how to tell an assistant what you want, because that single skill drives every later result.

Day by day

1. Day 1: open any assistant, paste one real enquiry with the name removed, and ask for a reply. Read the result and write down two things you would change.
2. Day 2: write a fuller instruction: your role, your tone, the facts it may use and what it must never promise. Reuse it on a second enquiry.
3. Day 3: paste two of your own past replies you liked and ask the assistant to match that voice.
4. Day 4: try three different enquiries, one easy, one angry and one off-topic. Note which the assistant handled worst.
5. Day 5: save your best instruction as a document, time one reply start to finish, and write the Friday note.

The prompt she ends the week with

You write first replies for a roofing company with seven staff. Tone: friendly, plain, short; no jargon.
You may say: we survey within five working days; surveys are free; we confirm a price after the survey.
You must never: give a price, promise a date, or say anything about insurance.
If the message mentions a leak that is happening now, begin with "If water is coming in, call us on [phone number]".
Reply to the message below in under 90 words. End with a question that gets us the missing detail.

Message: [paste]

An illustrative output for a message about a slipped slate: "Hi [first name], thanks for getting in touch, and sorry about the slipped slate. We survey within five working days and the survey is free, after which we confirm a price. Could you tell me roughly how high the slate is and whether anything has come through the ceiling? [your name]". What she fixes: the assistant sometimes adds "we'll see you soon", which implies a date. She adds "never imply a date" to the must-never list. That one correction is the lesson of the week: the rules grow out of what goes wrong.

Week two: quotes, where numbers need a checking habit

Quotes bring figures, and figures are where assistants make mistakes. Week two teaches the habit that makes AI safe to use with numbers: the assistant writes words around numbers you supply, and never supplies the numbers itself.

Words around numbers you supply

Give the assistant the job facts as a short list and ask for a cover note. Then check every figure and date against your own quote document. Spend the first two days on this, the third day on turning the cover note into a reusable template, and the last two on the harder case, quotes that carry options.

Write a short cover note for this roofing quote. Use only these facts.
Customer first name: [first name]
Work: replace 14 m2 of flat roof covering with EPDM, new edge trims
Price given in the quote: [price]
Start window: [window]
Guarantee: [guarantee wording from our standard]
Do not add any figure, date or promise that is not listed. If a fact is missing, write [MISSING] instead.

The [MISSING] instruction is the cheapest safety device in the plan. In an illustrative test the assistant returned the note with "[MISSING]" in the start window, rather than guessing a month. That is the behaviour you want, and you can make a habit of searching every draft for square brackets before sending. A human review rule helps here: how to set up human review for AI work without slowing down describes one.

Measuring the week

Time the old way once: say 11 minutes to write a cover note. With the template the time falls to five minutes including the checking. Eight quotes a week saves 48 minutes, roughly 40 hours over a working year. That is the first number for the week-four decision.

Week three: the dull jobs, invoices and job notes

This is where owners usually find the biggest saving, because admin tasks are repetitive and low in stakes. It is also the week you meet a feature that stores your instructions so you stop pasting them each time.

Two tasks, one tool feature

Task one is invoice chasing. Give the assistant the invoice date, the amount and what has happened so far, and ask for a polite reminder in three strengths: first nudge, second chaser, final notice. Task two is job notes. After a visit she dictates or types rough notes and asks for a clean summary: what was found, what was done, what is next.

The feature is a project, a saved space where your standing instructions and reference documents live. Claude's pricing page lists Projects on every plan, with up to five on the free plan, and OpenAI's help centre says project sharing works on ChatGPT Free, Go, Plus, Pro, Business, Enterprise and Edu (both checked November 2026). Put your tone guide, your standard guarantee wording and your template instructions in a project, so each new chat starts with them.

A before and after for invoice chasing

Before: the owner writes each chaser from scratch, around six minutes each. She does six a week. After: one project holds three templates and a rule for the tone; each reminder takes about two minutes, mostly for filling in the figures and checking. That is 24 minutes a week saved. If the saved time is spent on the quotes that are still unanswered, the gain is worth more than the minutes.

Reminder 1 (day 7 overdue): friendly, assume it was missed.
Reminder 2 (day 21): firm but polite, state the amount and the date it was due.
Reminder 3 (day 35): formal, say what happens next and by when. Do not threaten legal action.
Use only the figures I give you. If anything is missing write [MISSING].

The last line of the rule matters: chasing is an area where a wrong figure or a threat you cannot back up damages the relationship, and sometimes your legal position. If you are unsure what a final notice may say, ask a qualified adviser and put that wording in yourself.

Week four: measure, compare and decide what to implement

The last week has no new job. It turns three weeks of practice into a decision. Spend the five days like this.

1. Day 1: gather your three Friday notes and the times you measured.
2. Day 2: score each job against the rubric below.
3. Day 3: check the privacy and cost of the tool you used, and decide which plan the winning jobs need.
4. Day 4: write the one-page plan for the job you will implement properly.
5. Day 5: choose who else on the team will use it and what they must check.

A five-question rubric for scoring each practice job

Score each job out of ten. In the illustration, the enquiry replies score 9, the quote cover notes 7 (fixing time on the figures) and the invoice chasers 8. Anything scoring under 6 is shelved, not abandoned. The job with the highest score is the one she implements first: it is frequent, low-risk and she has a tested prompt.

The one-page plan for the winning job

JOB: first replies to web enquiries
OWNER: office manager (trained by me in week 5)
TOOL AND PLAN: business chat plan, 2 seats, $25 each monthly (or $20 billed annually)
PROMPT: saved in the project "Enquiry replies"
CHECK: person reads every reply before sending; searches for [MISSING] and any figure
MEASURE: minutes per reply and replies edited, tallied weekly for 4 weeks
REVIEW DATE: [date four weeks on]
STOP RULE: no gain in two review periods, or any wrongly promised date

The tool line reflects a vendor rule worth knowing: ChatGPT Business and Claude Team both require at least two seats. If you work alone, an individual plan may do. To weigh who in the firm needs a paid seat, see who in your team needs a paid AI licence.

What week four looks like with a different owner

To show the decision is not always "implement the top scorer", take an illustrative owner of a small plumbing firm. After three weeks her scores are: customer replies 8, job write-ups 9 and parts lists 5. The job write-ups score highest, but the engineers who would use the tool dislike typing on a phone. Rolling it out would fail on people, not on the tool. She chooses the customer replies first, because the office staff will use it today, and puts the write-ups on a list for a later test with voice dictation. Scores inform the decision; they do not make it.

A second, shorter example shows an owner who finishes the month with no job worth implementing. Her work is mostly on-site, and her admin is a couple of emails a week. Her week-four decision is to keep the free plan for occasional drafting, spend nothing and re-test in six months. That is a good result: she spent about ten hours to avoid spending a monthly fee on something that would not pay.

Questions to answer in the week-four note

1. Which one job would I miss most if I had to stop using AI tomorrow?
2. What is the single worst mistake the tool made, and have I written a rule against it?
3. How many minutes did one task save, and how many tasks a month is that?
4. What would it cost a month to run this properly, including a second seat?
5. Who will check the output, and how will I know they are doing it?

If the answers to questions three and four give a saving of more than twice the cost, you have your implementation. If not, your plan is to keep practising on a better-chosen job.

Four ways a practice week goes wrong, and the fix for each

Some weeks will not work. Do not treat it as a verdict on AI or on you.

The assistant keeps producing generic text

Your instructions are too thin. Add three facts about your business, one example of a reply you liked and one thing it must never say. Generic output is nearly always a signal of generic input.

It takes longer than doing it by hand

Often true in the first three days, because you are learning. If day five is still slower, the job may be too short or too varied to benefit. Swap it for a more repetitive one. Heavily varied writing, such as one-off letters, is a poor first job.

You catch a mistake after sending

Log it, add a rule, and add a check step for that type of error. In one illustrative case, an owner found a reply that offered a Saturday survey the company does not do. The fix was a line in the facts: "We do not work weekends." Every owner has one of these in week one, and the rule from it prevents the next.

Nobody else will use it

Adoption is a people problem. Pick a person who is curious, not the most senior. The tutorial on choosing and supporting an AI champion in a small team covers how. Once the winning job is live, five measures a small firm can check gives you a weekly sheet for keeping score.

Where this plan sits beside other ways of learning

There are other routes. Courses teach vocabulary quickly and are worth a few hours if you like structure. A gentler version, at 20 minutes a day, is a four-week plan that turns one task into a routine. This plan differs in two ways: you pick four jobs from your own week with the three-question sort, and week four ends with a scored, written decision on what to implement rather than a single habit. Either way it ends with working prompts instead of a certificate.

One honest limit: thirty minutes a day will teach you to use assistants well on writing, summarising and checking. It will not teach you to connect tools to each other, automate a process across systems or judge a custom build. Those are separate steps, taken only after you have a job that proves worth. If you decide you want help with that stage, mapping which jobs to automate and handing over something your team can run is what my AI implementation consultation covers, and you can ask about it here.

A calendar you can paste into your diary

Week 1  Mon-Fri 08:30-09:00  Enquiry replies. Friday: note + saved prompt.
Week 2  Mon-Fri 08:30-09:00  Quote cover notes. Friday: note + [MISSING] check.
Week 3  Mon-Fri 08:30-09:00  Invoice chasers + job notes. Friday: set up a project.
Week 4  Mon     08:30-09:00  Gather notes and times.
        Tue                  Score the three jobs.
        Wed                  Check privacy, plan and cost.
        Thu                  Write the one-page plan.
        Fri                  Choose who joins; set review date.

If you follow this for four weeks you will have three tested prompts, three measured savings, a rubric score for each job and a one-page plan for the best one. That is a far firmer basis for spending money on AI than any demo, and it is entirely your own.

Questions busy owners ask before they start

Do I need a paid AI plan to follow this?

Not for the first two weeks. Free plans from the main assistants are enough for drafting and summarising. If you will paste customer or financial information, move to a business plan in week three, since those do not train on your content by default. Check each vendor's terms first.

What if I miss a few days?

Skip the day and keep the order. The plan works on weekly jobs, not on a streak. If you lose more than three days in a week, repeat that week rather than rushing the next. A job half-learned is worse than a week of delay.

Should my staff do this alongside me?

Yes, for one person at least. Give a trusted colleague the same weekly job so you can compare results in week four. Do not roll anything out to the whole team until you have a prompt that works and a rule for what gets checked.

What if none of my jobs seems suited to AI?

Start with writing: replies, summaries and checklists are the safest tasks. If after two weeks nothing saves time, that is a valid finding. Record it, and revisit when your work changes or the tools do.

Is a course better than learning on real jobs?

A course can teach vocabulary quickly, but it will not know your invoices, your customers or your tone. Real jobs give you reusable prompts and honest evidence about what to adopt. Use a course for gaps, not as the plan.

Further reads

• A Simple AI Risk Register for Small Businesses (With Template) — Log what went wrong in your weeks of practice.
• Questions to Ask an AI Vendor Before You Sign Anything — Ask the right things before paying for a tool you tested.
• How to Brief a Developer on a Custom AI Workflow You Need — Brief a developer once you know which job to automate.
• What Does an AI Implementation Consultant Actually Do? — See what outside help covers if you want it.
• Do I Need an AI Consultant? 8 Signs It's Time to Get Help — Eight signs the plan alone will not be enough.
• AI Security Risks for Small Businesses and How to Close Them — Know what an assistant stores before you paste real work.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI help centre article on projects in ChatGPT, OpenAI's ChatGPT business pricing page and Claude's pricing page, checked November 2026; the plan is a worked illustration, and its timings and figures are examples.