ChatGPT for small business: where it saves hours, and where it costs you

The test: repetitive, well-defined, reviewable

A job is a good candidate when you do it often, you can describe what 'done' looks like in one sentence, and you can spot a wrong answer in under a minute. Miss any one of those three and you will spend more time correcting the output than you would have spent doing the work.

That test is why 'write my monthly newsletter' works and 'decide my pricing' does not. One has a shape and a reviewer. The other is a judgement call that only you can carry.

Start here: the five jobs that pay off first

In rough order of hours returned per hour invested:

• First-draft customer replies, generated against your real policy text.
• Turning one finished thing into its own promotion — a product into a post, a post into an email.
• Cleaning and reformatting data you already have: exports, lists, notes into a table.
• Documenting a process you already run, by describing it out loud and having it written up as steps.
• Pre-flight review: reading your own page or email back as a sceptical stranger before you publish it.

Set it up once so you are not re-explaining your business

The single biggest quality jump is a saved context block. Write it once, paste it at the top of any serious session, and update it when the business changes.

Business: {name, what it sells, to whom}. Price points: {list}. Customer: {who they are, what they already tried, what they fear}. Voice: {three adjectives and one sentence of a real example}. Never claim: {results, guarantees or credentials you do not have}. When you lack a fact, ask.

The three jobs to keep away from it

Not out of principle — because the failure is expensive and hard to see:

• Anything with a legal or financial consequence: tax treatment, contract terms, refund law.
• Claims about results. If the model writes it and you cannot personally back it, it is a liability, not copy.
• Anything customer-facing sent unread. One hallucinated policy detail costs more than the hour it saved.

How to tell whether it actually helped

Pick one job. Time it by hand for a week, then time it with AI for a week, and count the corrections. If corrections take more than about a third of the original time, the job failed the reviewability test and you should hand it back to yourself.

I keep this deliberately crude. The point is not a dashboard, it is catching the tasks that feel faster and are not.

A realistic first week

Day one, write the context block. Day two, do support replies with it. Day three, document one process you keep in your head. Day four, turn one finished thing into a post and an email. Day five, run the objection pass on your own best page. That is a week and it costs nothing.