Automating processes with AI in a small business: where to start

Two years ago, automating with artificial intelligence was something for large companies with data departments. Today a ten-person business can have an agent that reads order emails, pulls out the data and enters it into the ERP for less than the cost of a laptop. The problem is no longer the technology: it is knowing where to start.

The most common mistake we see in Madrid and the surrounding area is starting with the tool ("we have bought AI licenses for everyone") instead of the process. This guide goes the other way round: first you choose a process that is worth it, then you measure it, and only then do you decide which technology to use.

What we mean by "automating with AI"

It helps to separate three things that often get mixed up:

• Classic automation. Fixed rules: when an order comes in, create a task. It has been around for years and is still the cheapest option when the process has no ambiguity.
• Generative AI. Language models that understand text, documents and images and produce text. They are used to read, classify, extract, summarize and draft.
• AI agents. A language model with access to your tools (email, ERP, CRM, calendar) that carries out a multi-step task and asks for help when it is unsure.

What really changes a small business's bottom line is combining all three: rules for what is clear, a language model for what is ambiguous, and a person reviewing what matters.

How to choose the first process: five criteria

Make a list of ten repetitive tasks and score each one from 1 to 5 on these criteria. The first process is the one with the highest score, not the one that sounds most exciting.

1. It is frequent. It happens several times a day or, at the very least, several times a week. Automating something that happens once a month is rarely worth it.
2. The input is text, email or documents. Invoices, delivery notes, requests, forms, transcribed calls. That is where generative AI makes the difference over a script.
3. It has clear rules, but with variation. If it were exactly the same every time, classic automation would do. If there are no rules at all, AI will not fix it either.
4. A mistake is not serious and someone reviews it. At first, the agent proposes and a person approves. Leave the processes where a failure costs money or a customer for later.
5. It can be measured. Hours per week, days of delay, number of errors. If you cannot measure the before, you will not know whether it worked.

Start with what is boring, frequent and measurable. The glamorous process can wait.

Examples of AI agents already working in small businesses

These are use cases being deployed today with mature technology, with no research or experiments:

• Inbox agent. Reads the orders or customer-service mailbox, classifies each message, extracts the data (customer, references, quantities, dates) and creates the order or the ticket in your system. Anything doubtful is flagged for review.
• Reading supplier invoices and delivery notes. Receives the PDF or the photo, pulls out the data, matches it against the purchase order and leaves it ready to post. An accounting firm or a purchasing department pays it off in weeks.
• Assistant over your documentation. Answers customer or employee questions from your manuals, price lists and terms, citing the source and escalating to a person when it does not know.
• Quote preparation. From a free-text request, it proposes a quote using your price list and leaves it as a draft so the sales rep can review it in two minutes instead of twenty.
• Sales follow-up. Transcribes and summarizes calls and meetings, updates the CRM and drafts the follow-up email.
• Data quality. Detects duplicate customers, incomplete addresses or misclassified products and proposes corrections.

Notice the pattern: in all of them the AI does the heavy lifting and a person decides. That is the formula that works in 2026.

A concrete example: the accounting firm that received 300 emails a day

Take an accounting and tax advisory firm in Madrid with nine people and around 180 clients. Every day some 300 emails come in: invoices to post, payslips to review, questions about deadlines, loose documents. Two people spend half their working day opening, classifying, renaming and filing.

The typical solution has three pieces. An agent that classifies each email by client and type. An extractor that reads the attached invoices, pulls out the date, total, net amount, VAT and supplier, and saves them in the client's folder with the correct name and a row in the control spreadsheet. And a drafter that, for frequently asked questions, prepares a draft reply that a person reads and sends with one click.

What does not show in the table, and what clients value most: documents stop getting lost and replies go out the same day.

Common mistakes (and how to avoid them)

• Starting with the website chatbot. It is the most visible thing and usually the one that adds the least value in a small business. Start with what happens inside.
• Automating a process that is not defined. If everyone does it differently, you first need to agree on how it is done. AI does not fix disorganization, it speeds it up.
• Taking the person out of the loop too soon. In the first weeks the agent proposes and someone approves. Once it has gone a month without significant errors, extend its autonomy.
• Forgetting GDPR. If the agent reads customer data, you need a data processing agreement with the model provider, services that do not train on your data and, ideally, processing within the European Union. This can be solved, but it has to be done.
• Not measuring. Without a "before", the project turns into a debate of opinions.
• Buying licenses for everyone and waiting. Giving the whole staff an AI subscription without a specific process generates curiosity for a week and abandonment within a month.

How much it costs and how long it takes

Indicative ranges for Spain, excluding VAT, with a small, senior team:

The monthly cost depends mostly on volume: a language model processes an invoice or an email for a fraction of a cent, so what weighs is the server, supervision and improvements. Also check the public digitalization grants currently open in your region, as several have included specific lines for artificial intelligence.

Steps to get started this week

1. Inventory. Ask each person on the team to write down the three tasks they repeat most and how much time they take per week.
2. Score each task against the five criteria above and choose the one with the highest total.
3. Measure the before. Hours, delays, errors. Two weeks of data is enough.
4. Define the review loop. Who approves what the agent proposes, and what happens when it is unsure.
5. A two- to four-week pilot with real data and a single person reviewing everything.
6. Measure the after and decide: extend the agent's autonomy, adjust, or stop.
7. Document and train. Make sure the team knows what the agent does, what it does not do and how to correct it.
8. Second process. With what you have learned, the next one costs half as much.

Want help choosing your first process?

At AltavenirMS we offer a free automation assessment: we go through your task list with you, score the processes and tell you which one to start with, what it would cost and what return to expect. Get in touch through our contact page or see what we do in automation and applied AI.