A small business owner asked me a fair question last month: everyone says to use AI, but nobody says what to do on Monday morning. This is the short version of what I would check before spending anything - the same ground covered in more depth in my full guide on AI for small business.
Question one: which job are you already paying for?
Write down the tasks you currently hand to a person, a freelancer or an evening of your own time. Then ask which of those has a clear input and a checkable output. Drafting first replies to customer questions qualifies. "Improving the business" does not. If you cannot name the job, the tool will not find it for you.
Question two: what happens when the output is wrong?
Every assistant produces confident mistakes at some rate, and the honest answer is that the rate is not zero. Decide in advance who reads the output before a customer does, and what the fallback is. A business that sells physical goods can absorb a clumsy product description; one that issues tax paperwork cannot absorb an invented figure. Match the risk of the job to the amount of human review you can genuinely sustain.
Question three: is the cost a subscription or a habit?
Published rates as of September 2026 put a single paid assistant at US$20 a month, or US$17 a month when billed yearly, and a seat inside an existing office suite from US$18 per user each month. A realistic first year for a five-person business that does this properly lands between US$250 and US$1,500 in software. The surprise is never the subscription line - it is the first month of writing instructions and re-reading output, which nobody bills you for and which you pay in evenings.
Question four: where does the data actually go?
If customer records, contracts or identity documents are involved, the tool you pick decides where that data lives. Vietnam's rules here are not optional, and they are covered in detail in the data and legal section of the original article. For a first experiment, keep real customer data out of the pilot entirely and work with samples. Once the workflow proves itself, revisit the question with the compliance obligations in front of you rather than behind you.
Question five: what does success look like in ninety days?
A rollout that does not blow the budget usually follows the same shape: pick one job and measure it in the first month, connect it to the systems you actually use in the second, and formalise and document it - with one named owner - in the third. If nothing measurable has changed by day ninety, the correct decision is to stop, not to buy a bigger plan.
The version I would actually run
Start free, prove one job, then pay for exactly one assistant and one specialist tool for the single most expensive repetitive task. Businesses processing a hundred or more documents, or three hundred or more support messages a month, are the ones where a paid automation earns its place. Everyone else should stay on the free tiers and spend the money on training instead.
Three things worth repeating
- A tool that saves time on a job you do not have is a cost, not a saving.
- The named person who checks the output matters more than the model you choose.
- Ninety days of measurement beats a year of enthusiasm.
FAQ
Do I need AI at all if the business is small?
Not necessarily. If no single repetitive job consumes meaningful time each week, a better process will usually beat a new tool. The value appears when one job is both frequent and checkable.
What is the cheapest honest way to start?
Use the free tiers of an assistant you already have access to, on one job, for a month. Spend nothing until you can describe what improved.
Is it safe to put customer data into these tools?
It depends on the tool and the data. Keep real customer records out of a pilot, and treat the legal obligations as a design constraint rather than a later problem.
How long before I know it is working?
Give it one full cycle of the job you picked - often a month - and compare the same numbers you would have tracked without it.
Top comments (0)