Ten real uses of generative AI in a small business

No transformation promises: ten concrete tasks where AI genuinely saves hours today, three where it is still not worth using, and the data rule that applies to all of them.

2026-08-15 · Automation

The conversation about AI in small businesses swings between two equally useless extremes: that it changes everything, or that it is good for nothing. Reality is duller and more usable: there is a handful of concrete tasks where it saves real hours from day one, and others where it still creates more work than it removes.

Here is the list, with the criterion of what works today and what does not.

The ten that work

1. Drafting the first version of any repetitive text. Quotes, proposals, product descriptions, follow-up emails. AI does not write the final version: it writes the version you start from, and that is half the time.
2. Summarising long documents. A forty-page contract, a tender document, a report. Asking it «what obligations am I taking on and by when» saves an afternoon of reading. Afterwards you have to read what it flags, not trust the summary.
3. Extracting data from documents. From a pile of invoices or delivery notes in PDF, pulling supplier, date, net amount and tax into a table. It is one of the things that works best and gives back the most time.
4. Classifying and prioritising incoming email. Separating a sales enquiry from an incident from advertising, and routing it.
5. First-line customer service replies. For the questions that repeat. With an explicit limit: when the question goes off script, it passes to a person without pretending to know.
6. Translating and adapting content into another language or another register. Especially useful if you sell in several countries.
7. Turning notes into documents. From meeting notes to minutes with agreements, owners and dates.
8. Helping with spreadsheets. Writing the formula you cannot remember, explaining what an inherited one does, working out why a reconciliation does not reconcile.
9. Preparing internal training material. Procedure manuals, how-to guides, frequently asked questions for new joiners.
10. Exploring data. Asking a sales table questions in plain language and getting a first analysis to start from.

The three where it is still not worth it

Deciding anything with legal or financial effect. Which expense is deductible, how a dismissal is classified, what a rule requires. AI produces plausible answers with a confidence that does not match their reliability, and in tax or employment matters a plausible false answer costs money. It is useful for preparing the question, not for giving the answer.

Figures going into an official document. Any number destined for a return, a payslip or a contract is calculated with the appropriate tool and checked. No exceptions.

Anything published without a person reading it. Texts, replies to customers, posts. The day it goes wrong, the problem belongs to the business, not to the model.

The data rule that always applies

This is the part almost nobody raises before starting, and it is the one that can turn an hours saving into a breach.

Personal data of your customers and your team does not get pasted into a chat without checking three things first: what contract you have with that provider, whether that data is used to train models and where it is stored. In Spain the GDPR applies and in Chile Ley 21.719 with its own calendar; in both cases, putting third parties' data into a tool with no legal basis and no processor agreement is a problem.

The practical solution is not banning AI: it is pseudonymising first. Replace names, identifiers and account numbers with codes on your own machine, work with the pseudonymised text and put the real data back at the end. It is a one-minute step that removes most of the risk.

We develop it in using AI with customer data within the law, and what the European rule already in application requires is in the AI Regulation in Spain.

How to start without spending

The order that works:

1. Pick one task from the list above that you do every week and that wears you down.
2. Measure how long it takes you today. A number, even a rough one.
3. Try it for two weeks with the tools you already pay for. Almost every office suite now includes something.
4. Compare. If it does not save measurable time, drop it and move to the next. Do not persist out of enthusiasm.
5. Write down the instruction that worked and share it with the team. The value is not in the tool: it is in the well-tuned instruction, which is reusable.

And a spending warning: before buying a specific AI tool, check what is already included in what you pay for. Many businesses are paying twice for the same thing.

If this sounds like you

What pays best is not the tool, it is knowing which of your tasks gets automated and in what order. That is the process and AI audit: it comes out with the prioritised list, the estimated saving and what is not worth automating. Training the team to use it well is AI training, and the approach is in what to teach and in what order.

If the specific case is that invoices are typed in by hand, the direct short cut is automatic invoice capture.

We are Mindset & Code: automation, data and development for small businesses. You can see what we do and what it costs.

General guidance. Each tool's data processing terms change frequently and are checked in its contractual documentation before using it with customer data.