AI for a small business: the four uses that actually pay

For a business of two to twenty people, four uses of AI reliably return more than they cost: drafting and repurposing content you already produce, triaging incoming customer messages before a human answers them, pulling structured data out of documents like invoices and forms, and summarizing research before a decision. Everything else is either a nice extra or a project with a budget. What these four have in common is that the work is text, it repeats, and somebody can check the result in less time than doing it from scratch. What they also have in common is that none of them replaces the person. Each one moves a human from producing to checking.

Drafting and repurposing what you already say

The highest return version of this is not generating new content. It is multiplying what already exists. One decent case study becomes a short post, three social captions, a paragraph for the newsletter and an answer for the frequently asked questions page. You supplied the facts, so the model is reformatting rather than inventing, and that is the safest job you can give it.

The same applies to the writing nobody enjoys: product descriptions from a spec sheet, a delivery delay email, a polite refusal, a job advert from the notes you scribbled about the role. Give it your own past examples so the tone comes out like you rather than like everyone else, which is what showing the model a few examples is for.

Review requirement: read every word that carries a claim, a price, a date, a commitment or a name. Anything generated from thin air rather than from your material needs checking against a source, for the reason set out in why AI makes things up. Marketing copy that overstates what you deliver is a consumer protection problem in most countries, not just an embarrassment.

Triage, not replies, for customer messages

The tempting move is an automatic reply bot. The move that pays is triage: read the incoming message, classify it, route it, and draft a suggested response that a human sends or edits. You keep the relationship and you lose the ten minutes of sorting.

Concretely, a small team can categorize by type (order status, refund, complaint, new sales enquiry, spam), flag the ones showing anger or a legal threat, extract the order number, and attach a draft built from the answers you already use. The person opens a queue that is sorted and pre filled instead of a pile in date order.

Review requirement: a human sends. Always, until you have a few months of evidence, and even then only for the narrowest categories like acknowledging receipt. Never let a model decide a refund, quote a price, promise a delivery date or interpret a contract term.

Pulling structured data out of documents

This is the quietest of the four and often the most valuable. Supplier invoices into a spreadsheet. Receipts into expense lines. Handwritten order forms into a system. A stack of resumes into a comparable table. Delivery notes reconciled against what was ordered. The output is a table rather than prose, which means you can check it, and asking for an exact output format is what makes it usable rather than merely impressive.

The trick that makes this trustworthy is to demand a source reference for each extracted field, for example the line or page a figure came from, and to have the model leave a field empty rather than guess. A blank you can fill in beats a plausible number you cannot detect.

Review requirement: spot check a fixed percentage every week, permanently, not just at the start. Reconcile totals against something independent, like the bank statement. Extraction accuracy drops the moment a supplier changes their template, and nothing will announce that.

Research and summarizing before a decision

Before you choose a supplier, enter a market, respond to a tender or read a forty page report, a summary gets you oriented in minutes. Ask what a document commits you to, what the unusual terms are, which questions you should be asking. That is a genuine saving on the slow reading phase.

Review requirement: this is the use where a wrong answer is most expensive, because summaries feel complete. Treat every figure, regulation, deadline and quoted clause as unverified until you have seen it in the original. Use the routine in how to fact check an AI answer in five minutes, and never send a summarized legal or tax position to a client as though it were your own advice.

Customer data, consumer accounts and business terms

The moment you paste a customer name, address, order history, health detail or anything financial into a chatbot, you have shared personal data with a processor. That is allowed in most places, but it is your responsibility, not the tool's, and there are conditions.

The difference that matters is consumer versus business terms. A free or personal plan typically permits the provider to retain conversations and may use them to improve its systems unless you change a setting. A business or team plan normally rules training out contractually, offers administrative control and retention settings, and provides the data processing agreement you need if you are subject to the GDPR in the EU or the UK. Several US states now give consumers similar rights over their data. If you serve customers in the EU or UK, or in California, assume you need that paperwork rather than hoping you do not, and check the clauses described in what to look for in an AI vendor contract. This is general information rather than legal advice for your situation.

Three practical rules cover most of the exposure. Use business accounts for business data, because the terms differ and the price gap is small. Minimize what you paste: an order number and a question beat a whole customer record. And know roughly where your prompts are processed, since the country and the retention period are the two facts a customer or a regulator will ask about first.

Write the rules down even if there are three of you. One page listing the approved tool, the categories that must never be pasted in, and who to ask when unsure is enough, and a policy people will actually follow is short by design. Without one, staff use personal accounts on their own phones and you have no idea what left the building.

A two week trial with a real baseline

Do not start a project. Start a measurement.

1. Pick one task from the four above that somebody on your team does at least weekly and hates.
2. Before you change anything, spend one week recording the baseline: how long it takes, how often, and how often it has to be redone. Rough minutes are fine. Without this number you will never know whether anything improved.
3. Run two weeks on one tool with one person responsible. Not three tools, and not everyone at once.
4. Record the same numbers plus a new one: how often the output needed fixing, and how badly.
5. Decide on the comparison. Keep it if the time saved is real after the checking. Drop it if the rework rate is high, and try a different task rather than a different tool, because the task is usually what was wrong.
6. If you keep it, write down the prompt that worked so the next person is not starting over. A shared document of the five prompts your business actually uses is worth more than any subscription upgrade.

Two failure patterns are worth naming. The first is buying tools nobody adopted, which happens when the tool sits outside the software people already have open. The second is measuring enthusiasm instead of minutes. The broader pattern of where the time genuinely goes is covered in where AI actually saves time at work, and the short version applies cleanly to a small team: if checking takes as long as doing, you have added a step rather than removed one.