What Can ChatGPT Do for My Small Business? Jobs to Try First

ChatGPT can draft and tidy text, summarise long material, turn rough notes into structured documents, brainstorm options and explain spreadsheets, and those are the jobs to try first. They take little setup and carry low risk because you read the result before it goes anywhere. Leave final facts, legal wording and unchecked sums for a person.

Rank every candidate job on two questions: how much effort to set up, and what happens if the answer is wrong and nobody notices. A reworded email is cheap and harmless. A pricing change based on an unchecked calculation is neither. Jobs where you supply every fact and read every word come first. Anything that commits you to a customer, a price or a legal position comes last, and some of it should never be handed over at all.

Follow me on Instagram@sagnikteaches

How to rank a job: effort, risk and checkability

Three scores decide whether a job belongs in your first fortnight. Score each from 1 (low) to 3 (high).

Connect on LinkedInSagnik Bhattacharya

• Setup effort. Do you need files, connections or a template? A pasted paragraph is a 1. A Project with reference documents is a 2. A connected app with permissions is a 3.
• Damage if wrong. A typo in an internal note is a 1. A wrong figure in a quote to a customer is a 3.
• Checkability. Can you tell in under a minute whether the output is right? A summary of a page you can see is quick to check. A statistic you cannot trace is not.

Do jobs with low effort, low damage and high checkability first. That rule produces the order below. It also explains why so many first attempts disappoint: owners start with the most exciting job, such as an automatic answer to every customer, which scores high on damage and low on checkability.

Subscribe on YouTube@sagnikteaches

Ten jobs ranked from safest to riskiest

Jobs 1 to 3: tidy, summarise, extract

These three can start this afternoon on the free plan. They share a feature: the source is in front of you, so a wrong output is easy to spot.

Rewriting your own message

Say a nursery manager needs to tell parents that the garden will be closed for drainage work for two weeks. Her rough note says "garden shut 3rd to 17th, kids use the hall, bring wellies Fri, sorry". The prompt:

Rewrite this note as a short, warm message to parents (under 120 words).
Keep every date and instruction exactly as written. Do not add facts.
Note: "garden shut 3rd to 17th, kids use the hall, bring wellies Fri, sorry"

A good output keeps the dates and the wellies instruction, adds an apology, and runs to about 70 words. What to fix: models sometimes add details such as "indoor activities will include dancing". Delete anything you did not say. Time saved is small per message, perhaps four minutes, but a nursery sends dozens of notices a term.

Summarising a thread

Paste a long supplier email chain and ask for the decisions made, the open questions and the dates mentioned, as three short lists. Then scan the original for any date you do not see in the summary. The check takes a minute, and it catches the most common error, a date that has moved between messages.

Meeting notes into actions

A care agency's weekly rota meeting produces messy notes. Paste them and ask for a table with columns Action, Owner, Due date, and mark "Not stated" where the notes are silent. The "Not stated" rule matters: without it the model fills the gaps with plausible guesses. Keep names out of the notes if they identify clients; use initials or a role.

Jobs 4 to 6: draft, structure and brainstorm

These jobs make something new, so you must edit more. They are still low risk because nothing leaves the building unread.

For posts and product text, give ChatGPT your facts and one example of your own writing, and ask it to match the voice. A music school might paste three lines from its last newsletter and its term dates, then ask for a 90-word enrolment reminder. Check every date, fee and claim. A line like "limited places" is the model's habit, not a fact.

For checklists and procedures, ask for a first draft, then give it to the person who does the job. A veterinary practice reception might ask for an opening-up checklist of 12 items. The receptionist crosses out four that are wrong for the building and adds two. The draft saved the blank-page effort, and the correction made it true. Drafting for the person who does the job, then letting them correct it, is the habit that makes every later job safer.

For brainstorming, ask for ten options and a one-line reason each, then choose. Questions for a supplier call, ideas for a quiet-month offer or ways to explain a price rise are all fair uses. Keep the decision with you.

Jobs 7 and 8: spreadsheets and small data files

ChatGPT is good at explaining a formula you did not write and at suggesting one you cannot remember. It is less safe at doing arithmetic across a whole file by itself. If you upload a sheet and ask for totals, the answer may be right or may quietly skip rows. The rule: have it show the steps or the formula, then test on three rows you can add up by hand.

An illustrative case: a veterinary practice has 400 rows of appointment data and wants average appointments per weekday. Ask for a pivot-table recipe in your spreadsheet instead of a number, and check one weekday by counting. If the method is right on one day, it is likely right on the rest. The older tutorial on ChatGPT for Excel goes further on formulas, and our tutorial on what ChatGPT can read and what breaks lists file problems to expect.

Here is a prompt that keeps the arithmetic checkable, written for a care agency's visit log of about 250 rows:

The attached CSV is one month of home-care visits. Columns: Date, Carer code, Client code, Start, Finish.
Do not give me totals. Instead:
1. Give step-by-step instructions to build a pivot table showing the average visit length per carer code.
2. List any column problems you can see (blank cells, times in different formats, duplicate rows) with the row numbers.
3. Tell me one figure I can check by hand to confirm the pivot table is right.

An illustrative reply gives a five-step pivot-table recipe and warns that 12 rows have a blank finish time and will drop out of the average. That warning is worth more than the recipe. Before trusting any average, count the blanks yourself with a COUNTBLANK formula on the Finish column. If your count says 14 rather than 12, the model has misread the file, and you know not to rely on anything else it says about it.

Job 9: customer replies, approved before sending

This job pays back quickly and carries the most everyday risk, so keep a person in the loop. A saved prompt keeps the rules in one place.

You draft replies for a small music school. Be warm and brief (under 90 words).
Never promise refunds, discounts, places or dates. If the facts below do not answer the question, say we will check and reply today.
Facts: [term dates, fees, trial lesson policy]
Parent message: [paste, with names removed]

Imagine a parent asks, "Can my son swap to Thursday?" and the facts do not mention Thursday. A correct draft says the school will check availability and reply today. A bad draft says "Yes, Thursday works". The prompt rule above is what prevents the second answer, and your read of the draft is the backstop. Whether to let AI answer messages at all is a bigger question, covered in the tutorial on whether AI should answer customer messages.

Job 10: advice that needs a qualified person

ChatGPT can explain what a term means and help you write questions for your accountant or solicitor. It cannot take responsibility for the answer, and it can be confidently wrong about rules that vary by place and date. Use it to prepare, not to decide. For example, ask: "List ten questions I should ask an accountant before hiring my first employee." That is a good use. "Tell me whether my contract is valid" is not.

What ChatGPT cannot do reliably

• Know current facts without checking. Prices, opening hours, rules and product details change. If it has web search on, ask for sources and open them. Otherwise assume its facts may be old.
• Calculate long sums on its own. Ask for the method, test it, and do the final arithmetic in a spreadsheet.
• Say "I don't know". It tends to produce a fluent answer anyway. Prompts that allow "Not stated" cut this.
• Read every scanned or messy file correctly. OpenAI's file help says ChatGPT can read text and images in PDFs, but also that images embedded in Project files may not be picked up automatically. A blurred scan or a crumpled receipt is still hard to read, so spot-check any number it pulls from one.
• Remember your business by itself. Memory exists on every plan, but the free plan's memory is limited and paid plans get more, so keep real context in a Project you control rather than hoping it was remembered.
• Act safely on its own across your tools. ChatGPT Work runs longer, multi-step tasks across your apps and files. It is available on every plan in the desktop app, and on web and mobile from Plus upwards, and OpenAI describes it as checking in when approval matters. Treat it as a later step, once simple drafting works.
• Keep secrets. Anything you paste leaves your machine. Read the data terms for your plan first.

Setting up a Project so every chat starts informed

Once two or three jobs work, the repeated part is context: who you are, how you write, what you never promise. A ChatGPT Project holds that once. You give the Project instructions and a few reference files, and each new chat inside it starts with them. Keep the files small and factual: a price list, a policy page, a style sample. OpenAI's help page lists 5 files per Project on the free plan, 25 on Go and Plus, and 40 on Pro and Business.

An illustrative set of Project instructions for a veterinary practice's front desk:

You help the front desk of a small veterinary practice write clear, kind messages.
Tone: calm, practical, plain words. No jargon unless the owner used it.
Never give medical advice, diagnoses or dosage. For anything clinical, say a vet will phone back.
Never promise appointment times, prices or availability. Use only the facts I paste in each chat.
If a fact is missing, write "Not stated".
Sign-off: "[your name], front desk".

The clinical rule is the important one. Without it, a friendly model may reassure an owner about a symptom. With it, the draft hands the question to a vet. Review the instructions every few weeks, because they go stale as your policies change. Our tutorial on ChatGPT memory settings for business explains why a Project is steadier than memory.

A mistake that looked fine: the summary that lost a deadline

An illustrative case from a small care agency. The manager pasted a long email chain with a funding body and asked for a summary. The summary was tidy and read well. It left out that the reporting deadline had been moved from the 14th to the 10th in the fourth message. The manager worked to the 14th and sent the report late.

The failure was not exotic. A summary compresses, and compression drops the detail that changed. What would have caught it: asking explicitly for "every date and every change of date, in order", and spending one minute checking those against the thread. The prompt that works is narrower than "summarise this":

From the email chain below, list: (1) every date or deadline mentioned, with who said it and which message it came from, (2) any date that changed, showing old and new, (3) decisions made, (4) open questions. Quote the sentence for each date.

Asking for quotes turns checking into matching, which is faster than re-reading. The same habit helps with contracts, quotes and supplier terms.

A before and after: a quote follow-up

Quote follow-ups are a classic job 9 candidate. Here is a real-looking, illustrative case from a music school that sells term packages.

Before (written in a rush): "Hi, just checking if you got my quote for the term lessons. Let me know."

After ChatGPT, using the school's facts and your edit: "Hi [first name], I wanted to check that the quote for [child]'s ten Tuesday lessons reached you. The term starts on the 14th, and we hold the slot until Friday. Happy to answer any questions by reply or a quick call. [your name]."

The edit step matters: the model initially wrote "limited slots remain", a pressure line the school could not support, so the manager replaced it with the real hold date. The after version gives a reason to reply and a deadline that is true.

Free, Plus or Business: when to pay

Start on the free plan. Paid plans mainly buy higher limits, more files and extra features. As of November 2026, OpenAI's pricing page lists Go at $8 a month, Plus at $20 and Pro from $100; prices change, so confirm there before you buy. ChatGPT Business starts at two users, and OpenAI's Business page says business data is never used to train its models. On the free plan, OpenAI's file help caps uploads at three files a day, and scheduled tasks are not included.

A decision rule that works: move up when a limit stops you twice in a week, or when two people need the same instructions. For the plan comparison, see the tutorial on ChatGPT subscription options for a business team. Custom GPTs are scheduled to stop running on 11 December 2026, with OpenAI moving them to plugins, so put shared instructions and reference files in a ChatGPT Project instead. Project sharing works on every plan from Free upwards; a free account can share a Project with up to five people.

A tidy-up example: turning a messy price list into a clean one

One more low-risk job that surprises owners is cleaning up lists. A childcare nursery keeps its session fees in a note someone typed years ago, with inconsistent capitals, mixed formats and a couple of old rates. Paste the list and ask for a two-column table, Session and Fee, sorted by session length, with any line that looks inconsistent flagged rather than corrected.

Original lines (illustrative):
half day am - 28
HALF DAY PM $28.00
full day  $52
Full day (with lunch) 56.5
holiday club, per day: 45

A sensible output is a table with the five rows in one format, and a flag on "full day" appearing twice with different fees ($52 and $56.50). That flag is the useful part: the model noticed an inconsistency you may have forgotten. The decision about which fee is current stays with you. Never accept a silent correction, because a tidied number that looks right but came from a guess is worse than an untidy one.

Check the result by counting rows in and rows out. If you gave it five lines and it returned four, one has been merged or dropped. This count takes ten seconds and catches the commonest tidy-up error.

A two-week trial you can run

Pick two jobs from ranks 1 to 5. Time each by hand for three repeats to get a baseline, then use ChatGPT for ten repeats and record the time and the number of corrections.

JOB TRIAL LOG
Job: [e.g. parent notices]    Baseline by hand: [minutes]
Run  Time with ChatGPT  Corrections needed  Anything wrong that I almost missed?
1    5 min              2                   a date moved
...
Keep if: time saved per run is [x] minutes AND corrections are under [n] per run.

Illustrative numbers: baseline 12 minutes per notice, ChatGPT version 5 minutes including the read-through, and 1.5 corrections per notice. Weekly, with six notices, that saves 42 minutes. Over a school year of about 38 weeks it is roughly 26 hours. Small, but real, and it costs nothing on the free plan.

If a job fails the trial, look at the prompt before the tool. Missing facts, no word limit and no rule for unknowns explain most disappointing results. A quick repair: add the missing facts, a word limit and the line "If something is not in the facts, write Not stated", then rerun the same three examples and compare the correction counts with the first run.

Handling private information while you learn

Learning with real material is where most owners slip. Keep a simple rule: until you have read the data terms for your plan and switched off training use where it is on, paste only material you would be happy to see in a shared document. Replace customer names with [Customer A]. Use ranges rather than exact health or payment details. The tutorial on putting customer data into ChatGPT explains what happens to pasted data. On the Free, Go, Plus and Pro plans, OpenAI's pricing page marks training on your content as opt-out, so find that setting before the first real paste.

Picking the first one or two jobs for your own week is a good use of an AI implementation consultation. Otherwise run the trial above, keep the jobs that survive it, and add one new job a fortnight.

Quick questions before you start with ChatGPT

Do I need a paid ChatGPT plan to start?

No. The free plan is enough for the first jobs on this list, such as drafting emails and tidying notes, though it has lower limits and fewer features. Move to a paid plan when you hit limits, want larger file uploads, or need shared workspace controls. Check current plan prices on OpenAI's pricing page, because they change.

Can ChatGPT see my customers' details?

Only what you type or upload, plus anything you connect through apps. Treat the chat box as a place to put data deliberately. Leave out names, payment details and health information unless you are on a plan whose data terms you have read and your business is allowed to share that data.

Will ChatGPT remember my business between chats?

Memory is on every plan, but it is limited on Free and expanded on paid plans, so check your settings rather than assume. The dependable way to keep context is a Project: you store instructions and reference files in it, and every chat inside starts from them. Re-check what a Project contains before you rely on it.

Is it worth buying ChatGPT Business for two people?

ChatGPT Business needs at least two seats, so two people can use it, and a solo owner either pays for two seats or uses an individual plan. Business plans do not train on business content by default, per the vendor. The question is whether you need shared Projects and admin controls.

How long should I test before I decide it works?

Two weeks is enough for a first job. Time the task by hand for a baseline, do the AI version for ten repeats, count the corrections you had to make, and keep it only if time saved holds up after checking. If the correction rate stays high, change the prompt or drop the job.

Further reads

• How to Use ChatGPT Scheduled Tasks for Business Reminders — Turn a repeated prompt into a scheduled reminder.
• ChatGPT vs Claude vs Gemini vs Copilot for Small Business (2026) — See how the other assistants compare before committing.
• How to Use ChatGPT as a Business Adviser Without Being Misled — How to use ChatGPT for advice without being misled.
• How to Get Started With AI in Your Small Business: First 7 Steps — The first seven steps across all AI tools, not only ChatGPT.
• Is Free ChatGPT Good Enough to Run a Small Business On? — Whether the free plan can run a small business.
• Customer-Facing or Back-Office: Where Should AI Go First? — Decide whether the first job should face customers or not.
• ChatGPT Prompts for Small Business Owners: 50 Tested Examples — Tested prompt wording to borrow for your first jobs.
• How to Stop AI Tools Training on Your Business Data — Switch off training use before you paste real material.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI ChatGPT pricing page, ChatGPT Business and ChatGPT Work pages, and Help Centre articles on Projects, file uploads and custom GPT retirement (plan limits change; confirm on the live pages).