Every guide to getting started with AI in a small business gives the same instruction: find your repetitive, rules-based tasks and automate those first.
It is the wrong test. Work that follows consistent rules and needs no judgement is precisely the work a spreadsheet formula, a calendar rule or a twenty-line script already does — more cheaply, and without ever inventing an answer. If a task is genuinely rules-based, a language model is the most expensive and least predictable way to do it.
The test that actually predicts whether AI helps is about what happens when the output is wrong. Here are three questions, in the order worth asking them.
This is a selection framework, not a tool review. Nothing here was benchmarked. The adoption figures come from the Census Bureau; where widely quoted statistics have no traceable source, this article says so instead of repeating them.
Question one: is the work language-shaped?
Language models are good at work where the input and output are words, and where there is more than one acceptable answer. Summarising a long thread. Turning scrappy notes into a clear email. Rewriting the same information for a different reader. Pulling the three commitments out of a meeting transcript. Drafting a first version of something you will edit.
They are bad at work with exactly one right answer that can be computed: totals, tax, stock counts, scheduling conflicts, anything arithmetic. They will produce an answer in the right shape, and it will sometimes be wrong in a way that looks fine, which is worse than an error message.
So the first filter is not “is this repetitive” but “is this a judgement about wording or meaning”. If the task could be written as a formula, write the formula.
Question two: does a human see the output before it matters?
This is the question that decides whether you are allowed to be wrong.
Draft an email that you read before sending: a bad draft costs you ten seconds. Auto-reply to a customer enquiry with no one looking: a bad reply costs you the customer, and you may not find out why. The task is similar. The exposure is not.
Everything on your first list should have a human checkpoint between the model’s output and anyone outside the business. Not because the model is unreliable at everything, but because you do not yet know where it is unreliable for your particular work, and the checkpoint is how you find out cheaply.
Keep the checkpoint until you have seen enough output to know the failure modes. Then decide — with evidence — whether to remove it for that one task.
Question three: what does a plausible wrong answer cost?
The failure mode that matters is not a model that breaks. It is a model that produces something confident, fluent and wrong — which is a normal consequence of how these systems work, not a defect you can configure away. We have written about why that happens at length.
So price it before you start. For each candidate task, answer: if this output were wrong and nobody caught it, what would it cost?
| Cost of a wrong answer nobody catches | Example | Verdict |
|---|---|---|
| Ten seconds of your time | A draft you were going to rewrite anyway | Start here |
| An awkward correction | An internal summary with a wrong date | Fine, with a checkpoint |
| A customer’s confidence | Wrong price, wrong availability, wrong promise in a reply you did not read | Not yet |
| Money or a legal position | Quotes, invoices, contract terms, anything regulated | Not this way |
Rank your candidates by that column, not by how much time the task takes. The biggest time sink is a tempting first target and often the worst one, because the tasks that eat the most time tend to be the ones with the most at stake.
What this rules in
Run the three questions and a fairly consistent first list falls out:
- Inbound triage. Summarise and categorise enquiries so you can see what arrived. You still write the replies.
- Drafting from your own material. Quotes, follow-ups and descriptions generated from notes you wrote, which you then check. The model is rephrasing your facts, not supplying its own.
- Turning a long thing into a short thing. Transcripts into actions, documents into summaries, with the original a click away so anything surprising can be checked.
- Reformatting and extraction. Pull the dates and amounts out of forty emails into a table. Easy to spot-check, and the original is right there.
- First drafts of repetitive writing. Listings, descriptions, standard responses that you edit.
What they share: words in, words out, a human in the loop, and a wrong answer that costs seconds.
The first month, in order
Week one. Pick one task and do it by hand, with the model watching. Use whatever subscription you already have and keep a note of what you asked and what came back. Two useful things happen: you learn where the model is weak on your work, and you end up with a prompt worth keeping rather than a vague intention. Our prompts for work are built to leave a trail you can check in seconds, which is what you want for this.
Week two. Write the check, not just the prompt. Decide what “wrong” looks like for this task and how you would notice in under a minute. If you cannot describe the check, you are not ready to hand the task over. The five-minute verification routine is a reasonable default.
Week three. Run it daily and count the failures. Not a formal measurement — a tally. How often did the output need real correction rather than a tweak. If it is most days, the task is wrong for this or the prompt is, and you have learned that for the price of three weeks of notes.
Week four. Decide, then pick the second task. Keep it, change it, or drop it. Only now is it worth thinking about connecting the model to your actual systems, which is a different and larger job — the protocol most tools use for it is MCP, and it is worth understanding what you would be plugging into what before anyone sells you an integration.
One task a month is slower than the guides suggest and roughly the pace that survives contact with running a business.
What to keep off the list for now
- Anything that sends without you. Autonomous replies to customers are the last step, not the first.
- Numbers. Totals, invoices, stock, tax. Use software built for arithmetic.
- Anything regulated or contractual. Not because AI cannot help draft it, but because the checkpoint has to be someone qualified, which makes it a slower project than a first one should be.
- Anything confidential, until you have checked the setting. On consumer tiers this matters: OpenAI’s help documentation states that for services for individuals such as ChatGPT, “we may use your content to train our models”, with an opt-out under Settings, Data controls, “Improve the model for everyone”. Its enterprise privacy page, updated 8 January 2026, says that for business products and the API, “we do not train our models on your data by default”. If you are putting client material into a free or personal plan, check which of those you are on. Our guide to what free plans actually include covers the rest of the limits.
How common is this, really
Here the honest answer is more useful than a number from a listicle, and the listicle numbers are unusable: pages in this category routinely cite figures like “over 80% of small businesses using AI report productivity gains” with no survey, sample or date attached. We will not repeat those.
What there is: the US Census Bureau’s Business Trends and Outlook Survey, published 26 May 2026, covering collection from 14 December 2025 to 3 May 2026. Overall AI use among US businesses ran between 17% and 20% across that period, with 20% to 23% expecting to use it within six months. Use rises with headcount: as of the collection period ending 3 May 2026, 37% of firms with at least 250 employees reported using AI, 32% of those with 100 to 249, and fewer than 20% of firms with under 20 employees. Among the smallest firms, use did not change significantly over those five months, while it rose among firms with at least 20 employees.
Which is to say: if you run a small business and have not started, you are in the majority, and the gap is widening in favour of larger firms. That is an argument for picking one task deliberately. It is not an argument for hurrying.
What to remember
Do not start with your most repetitive task. Start with the task where a confidently wrong answer costs you ten seconds, keep a human between the output and anyone who matters, and add the second task only once you can say how often the first one fails.