There are two conversations about AI in small business. One is about transformation, ten-times productivity and replacing whole functions. The other is what actually happens on a Tuesday: someone uses it to draft a reply, summarise a long document, or turn messy job notes into a quote.
The second one is real. This is about that.
What businesses actually use it for
The QuickBooks AI Impact Report for Australia found small businesses using AI mainly for data processing (38%), administration (36%), marketing (34%), customer service (32%) and bookkeeping (32%).
Notice what those have in common. They are all tasks where the work is producing a first version of something or finding something in a pile. None of them are the business's actual craft. Nobody is reporting that AI does the plumbing, wins the tender or decides the pricing.
That is the useful signal, and it is more honest than most of what is written about this.
Be careful with the numbers you will read
You will find claims that AI cuts support costs by 40 to 60 per cent, or removes ten to twenty hours of manual work per person per week. Those figures generally come from companies selling AI implementation services, and they are describing best cases from projects they were paid to deliver.
The more grounded figure in the same research is that ordinary use saves one to two hours a week. That is not nothing — it is most of a working day each month — but it is a different decision from the one a "60 per cent cost reduction" implies. Plan for the smaller number and treat anything more as upside.
The four jobs it does well today
Drafting the first version. Emails, quotes, job descriptions, social posts, a policy you have been putting off. The value is not that the output is finished — it usually is not. It is that editing something mediocre is much faster than starting at a blank page, and blank pages are where small business admin goes to die.
Summarising. A long email chain, a supplier contract, meeting notes, a report you have been avoiding. Ask for the specific thing you need — the obligations, the dates, the decisions — rather than a general summary.
Finding things. Not searching the internet, but searching your own material. What did we quote this client last time? What was the agreed scope? This is the one people underuse.
Turning mess into structure. Voice notes into a written quote. A photo of a whiteboard into a task list. Scribbled measurements into a materials order. This is where trades and field businesses get the most out of it, and it is the least talked about.
Where it does not help, and will waste your time
Anything where being wrong is expensive and you cannot check the answer. If you do not know enough to spot a wrong answer, you cannot use it for that task safely. This is the whole game.
Facts about your own industry that change. Rates, thresholds, standards, prices. Models are confidently out of date. Look those up at the source.
Work where the value was the thinking. If you delegate the reasoning rather than the typing, you get something that reads well and means little. Your clients can tell.
Anything you would not have done anyway. A lot of AI "productivity" is producing more content nobody asked for, faster. That is not a saving.
How to actually get the hours back
Most people try AI on something big and impressive, get an underwhelming result, and stop. The businesses that get value do the opposite.
- Pick one task you do every week that you dislike. Not the most important one — the most repetitive one.
- Do it with AI five times. The first attempt is always the worst; you are learning how to ask.
- Save the prompt that worked. In a note, a doc, anywhere. Most of the value is in reusing the instruction, not in the tool.
- Only then add a second task.
Businesses that adopt this way keep it. Businesses that run a big "AI initiative" usually do not.
What to do about accuracy
Have a rule about which tasks get checked and by whom, and write it down before anyone else in the business starts using it. A workable default: anything that leaves the business — goes to a customer, a supplier, the public — is read by a person first. Anything internal and reversible does not need to be.
Also decide what must never be pasted into a general AI tool: customer personal information, anything covered by a confidentiality agreement, payment details, and staff records. That decision is easier to make once than to litigate later.
What to do next
- Name the one weekly task you would most like to stop doing manually.
- Try it five times before judging it.
- Write down your check-before-it-leaves rule, and what is never pasted in.
- Revisit in a month and count the hours honestly, not optimistically.
If a single hour a week comes back, that is fifty hours a year for effectively no capital outlay. That is a good result, and it is a much easier target to hit than the ones in the advertising.
Sources
- QuickBooks AI Impact Report — Australia, 2026, for reported AI usage by task.