# AI for small businesses: where it pays off and how to start > Which tasks in a 5 to 50 person company are worth handing to AI today, how to pick a first project, what data to keep out and how to measure results. URL: https://quinatec.es/en/blog/artificial-intelligence-business-practical-guide/ Tomas E. Cordoba · 04 May 2026 · Updated on 25 September 2026 The question we get from owners of small companies is no longer "should we be doing something with AI?" but "we've tried it, now what?". This guide tries to answer that without the hype: what actually works in a company of 5 to 50 people, how to choose a first project, what GDPR demands and how to tell whether it was worth it. ## Where AI pays off today in a small company Generative AI is good at one thing: reading, summarising, classifying and drafting text with reasonable judgement. Almost everything that works in a small business today builds on that. ### Documents and email - Pulling data out of invoices, delivery notes and orders that arrive as PDFs or photos, so they land in the ERP without anyone typing them in. - Triaging incoming email: sales enquiry, support issue or advertising, and who should get each one. - Drafting replies to customers from your own templates, so a person reviews and sends rather than writing from scratch. - Meeting notes: transcript, summary and action list, with nobody taking notes while people talk. ### Internal knowledge An assistant that answers questions from your own documents (procedures, price lists, manuals, contracts) cuts interruptions and shortens onboarding for new staff. It works if the documentation exists and is organised; if not, sorting it out comes first, as we describe in our guide to [document digitisation](https://quinatec.es/en/blog/document-digitisation-paperless-office-in-90-days/). ### Agents that run a workflow end to end The next step up is agents: instead of answering a question, they run a process from start to finish. Receive an order by email, check stock, create it in the ERP and reply to the customer with a delivery date. This is where the most hours are saved and where design matters most, with a human check at the steps that have consequences: payments, shipments, commitments to customers. ## How to pick a first project The most common mistake is starting with the most ambitious idea. The first project has to meet three conditions: 1. **Small**: one process, one team, live in weeks rather than months. 2. **Measurable**: you must be able to say how much time or how many errors it cost before. If you do not know, measure it for a couple of weeks before touching anything. 3. **Low risk**: if the model gets it wrong, someone sees it before it reaches a customer or the tax office. A good candidate is usually a repetitive task with clear volume and a person who suffers through it and would be delighted to stop: email triage or supplier invoices almost always fit. ## Privacy and GDPR: the non-negotiables This is where we see the most problems, nearly always out of ignorance rather than intent. - **No personal data in free tools.** Free or personal tiers of AI chat tools generally use what you type to train models. Pasting a payslip, a contract or a customer's email into one is handing data to a third party with no legal basis. - **Business accounts with a data processing agreement.** The business plans of the main providers include that agreement and do not train on your data. Get it in writing before you sign up. - **EU data residency.** More and more providers let you keep processing inside European data centres. For health, financial or children's data that is a requirement, not a preference. - **The EU AI Act.** For most uses in a small company (drafting, summarising, classifying) the obligations come down to transparency. Hiring, credit or safety fall into higher-risk categories and demand more care. ## What it costs Less than you expect in licences and more than you expect in your own people's time. - Licences for general-purpose assistants are billed per user per month, in the same range as an office suite subscription. - API usage is billed by the volume of text processed; for a small company that is usually a modest monthly amount. - The real cost is designing the workflow, integrating it with your systems and the weeks of tuning until the output is reliable. That is where the project succeeds or fails. ## Mistakes we see again and again - **Starting with the tool instead of the process.** "We want to use product X" is a poor starting point. "We spend four hours a day keying in delivery notes" is a good one. - **Nobody owns the output.** If the AI drafts and nobody reviews, the errors reach the customer. Every workflow needs a named person who owns its output. - **Expecting 100% accuracy.** It will not deliver that. Design the process so doubtful cases are set aside for a person to check; it is cheaper and safer than chasing perfection. ## How to measure whether it works Fix the metrics before you start and review them after one month and after three: - Hours per week spent on the task, before and after. - Error rate, or the share of cases sent back for review. - Total monthly cost (licences, API, review time) against the hours freed up. If the numbers do not add up after three months, either fix the design or switch it off. Deciding in advance what "not working" means stops you keeping things running out of inertia. ## Where to start next week Pick a task that repeats daily, measure what it costs, check what data it handles and set up a pilot with one person in charge. If you want a second opinion before investing, our [AI consulting](https://quinatec.es/en/services/ai-consulting/) service does exactly that analysis. And if the problem is more about process than intelligence, start with [automation](https://quinatec.es/en/services/automation/); very often that is enough.